pmid stringlengths 8 8 | pmcid stringlengths 8 11 ⌀ | source stringclasses 2
values | rank int64 1 9.78k | sections unknown | tokens int64 3 46.7k |
|---|---|---|---|---|---|
38667221 | PMC11048303 | pmc | 1 | {
"abstract": "Friction, wear, and the consequent energy dissipation pose significant challenges in systems with moving components, spanning various domains, including nanoelectromechanical systems (NEMS/MEMS) and bio-MEMS (microrobots), hip prostheses (biomaterials), offshore wind and hydro turbines, space vehicles, solar mirrors for photovoltaics, triboelectric generators, etc. Nature-inspired bionic surfaces offer valuable examples of effective texturing strategies, encompassing various geometric and topological approaches tailored to mitigate frictional effects and related functionalities in various scenarios. By employing biomimetic surface modifications, for example, roughness tailoring, multifunctionality of the system can be generated to efficiently reduce friction and wear, enhance load-bearing capacity, improve self-adaptiveness in different environments, improve chemical interactions, facilitate biological interactions, etc. However, the full potential of bioinspired texturing remains untapped due to the limited mechanistic understanding of functional aspects in tribological/biotribological settings. The current review extends to surface engineering and provides a comprehensive and critical assessment of bioinspired texturing that exhibits sustainable synergy between tribology and biology. The successful evolving examples from nature for surface/tribological solutions that can efficiently solve complex tribological problems in both dry and lubricated contact situations are comprehensively discussed. The review encompasses four major wear conditions: sliding, solid-particle erosion, machining or cutting, and impact (energy absorbing). Furthermore, it explores how topographies and their design parameters can provide tailored responses (multifunctionality) under specified tribological conditions. Additionally, an interdisciplinary perspective on the future potential of bioinspired materials and structures with enhanced wear resistance is presented.",
"introduction": "1. Introduction Sustainable development is closely linked to improvement in tribology (low friction, low wear, and enhanced lubrication), which has a significant and beneficial impact on both the environment and society [ 1 ]. Global energy consumption is increasing year on year, putting pressure on both resources and the environment. Notably, the repercussions of friction and wear extend beyond substantial energy and economic loss (approximately 23% of global energy loss and 1–2% of a nation’s GDP) and serve as a major source of CO 2 emissions [ 2 , 3 ]. Crucially, tribological losses (material degradation), comprising high friction and wear inefficiencies, expenses related to part replacement and remanufacturing, and maintenance costs, affect virtually all moving elements, ranging from industrial machinery to the intricate mechanics of natural bone joints and artificial implants integrated into the human body [ 4 ]. Any effort aimed at curbing these losses yields a direct and positive impact, not only conserving energy but also fostering economic stability, and enhancing the well-being of both individuals and the environment. Consequently, it is entirely valid to assert that tribology, with its inherent capability to diminish friction and wear, characteristically contributes to the noble pursuit of a sustainable and human-centric world, and directly or indirectly contributes to Sustainable Development Goals 3, 6, 7, 8, 9, 11, 13, 14, and 15 [ 5 ]. Figure 1 shows the triangle of sustainability, revealing nature, bioinspired materials, and bioinspired tribology as its three corners. In this context, various methods can mitigate friction and its negative effects, such as wear. Those can be classified into (i) the mechanical approach (e.g., well-lubricated systems, novel composites, microstructural enhancements, and heat treatments, etc.) and (ii) the surface topography engineering approach (e.g., adjusting surface roughness or texture to reduce contact area). According to a model proposed by McFarlane and Tabor [ 6 ], friction force is the sum of two components: a mechanical contribution resulting from surface deformation and an adhesive contribution mainly influenced by surface energy. This model suggests that a minimum coefficient of friction (CoF) can be achieved at a critical value of real contact area [ 7 ]. Consequently, controlling surface topography can lead to CoF control. However, traditionally, the focus has been on mechanical approaches, overlooking the potential of engineered interfaces in friction control due to incomplete understanding of surface texturing (optimal choice and design of textures). Additionally, an innovative approach utilizes a combination of both approaches to enhance tribological properties and reduce friction and wear. The synergy may enable simultaneous tribological benefits, such as favorable effects of wear debris, continuous lubricant replenishment, formation of protective tribolayers or tribofilms, etc. [ 8 , 9 ]. Multifunctionality within tribological materials/surfaces represents a paradigm shift in engineering [ 10 ]. These materials not only mitigate wear and friction but also offer measurable benefits for efficient operation in specific applications. For instance, the wear rate of modern hip implants has been reduced by several orders of magnitude [ 11 ]. However, to enhance their longevity and minimize the necessity for revision surgeries, it is equally crucial for biomaterials to positively carry dynamic fluctuating high loads (during human movement), facilitate cell transfer and growth (biological interaction), and ensure seamless tribological integration within the human body (chemical interaction), even leveraging the positive effects of wear debris (e.g., hydroxyapatite-based biomaterials can generate wear debris that stimulates bone growth and integration around the implant, while optimized surface roughness/texture can lead to improved osteoconductivity or lubrication) [ 11 , 12 , 13 , 14 ]. Similarly, in the domain of photovoltaic (PV) solar cells, materials/surfaces are engineered to reduce wear (erosion) while simultaneously improving light absorption and self-cleaning properties [ 15 ]. This innovation can result in remarkable, up to 15%, enhancement in the overall efficiency and lifespan of solar panels [ 16 ]. In hydro-bearings and hydrofoils, meticulously designed surface textures that reduce drag lead to elevated energy conversion rates [ 17 , 18 ]. In certain instances, these materials have demonstrated substantial efficiency gains exceeding 20%, thereby translating into increased renewable energy production. In other cases, such as surgical instruments or soft robotics, it is advisable that the surface demonstrate higher friction or higher adhesion in order to improve the grip strength while accommodating high load transfer (pick up/drop) [ 19 , 20 ]. In space shuttles and satellites, the surfaces are designed to resist impact (wear) from space debris (meteors) while being lightweight, anti-weathering, etc. [ 21 , 22 , 23 ]. Space rover wheels are designed to navigate easily through harsh terrain, requiring enough flexibility while delivering tribological robustness [ 24 ]. Highly durable surfaces that demonstrate water-repellent properties, efficient water harvesting and spreading, or exceptional resistance to oils play a pivotal role in various applications [ 25 , 26 , 27 ]. These surfaces are essential for ensuring rapid lubricant dispersion, enabling rainwater harvesting systems, and powering nanogenerators, among other uses [ 28 ]. For instance, in the context of lubrication, surfaces with superoleophobic characteristics facilitate swift and uniform distribution of lubricants in machinery, reducing friction and wear. Moreover, surfaces engineered for efficient water spreading enable collection and storage of rainwater for various purposes while being resistant to enduring impact, erosion, etc. In summary, the creation of wear-resistant surfaces with specific functionalities, such as water harvesting and superoleophobic properties, has far-reaching implications across a range of applications, from industrial machinery to sustainable water management and energy harvesting technologies [ 29 ]. Frictional anisotropy is a critical tribological feature that spans from the molecular scale to the macroscale, characterized by directional asymmetry in friction response during sliding [ 30 ]. This feature is particularly valuable for microbots designed for friction-based locomotion, offering high maneuverability and precise targeting in confined three-dimensional (3D) spaces. Additionally, it ensures stress compatibility with soft living matter by limiting interface stresses [ 31 ]. As the demand grows for microbots in various applications like robotic assembly, drug delivery, lab-on-chip technology, sensing, microsurgery, and cancer treatment, there is a practical need to construct artificial prototypes [ 32 ]. These prototypes, exploiting geometric features to induce controllable anisotropic motion responses under external oscillating loads, serve as models for understanding and predicting the dynamics of future industrial microactuators [ 33 ]. For instance, in medical microbots navigating complex biological environments, anisotropic friction-based locomotion coupled with physical propulsion holds promise as an enabling technology for their development [ 34 ]. Table 1 shows the essential functional features required in key tribologically challenged fields. Recently, significant attention has been paid to the exploration of materials, surfaces, and architectures inspired by nature [ 25 , 26 , 62 ]. The fact that natural objects have survived the harshest conditions through multifaceted evolution inspires engineers to create or imitate the construction of similar structures. Nature-inspired materials are man-made materials that mimic the structure, properties, or functions of natural materials or living organisms and offer the potential for sustainable synergy between tribology and multifunctionality [ 63 ]. The properties of these materials and surfaces result from complex interplay between the surface structure and the morphology and physical and chemical properties. Moreover, many materials, surfaces, and devices with such designs provide multifunctionality. Various terms such as bioinspiration, biomimicry, biomimetics, nature inspiration, and nature mimicry are commonly used interchangeably by researchers [ 63 ]. Biomimetic materials imitate the evolutionarily developed structural features, leading to adapted architecture, especially in living species [ 64 , 65 , 66 ]. Often, the architectures exhibit a graded structure spanning multiple scales, including macro-, micro-, and nanoscales. For example, bone’s complex porous structure, characterized by complex ligament shapes and variations in density, allows it to achieve superior mechanical (stiffness, energy absorption, stress distribution, etc.) and biological characteristics (e.g., porosity leads to facilitating nutrient exchange and cell proliferation) compared to most man-made materials (biomaterials) [ 65 ]. Friction and adhesion are common in nature. Notable examples involve leveraging nature-inspired surface textures found in animal scales and skins, such as those of sharks (for drag reduction and hydrophobicity), snakes (for erosion resistance), pangolins and turtles (for flexibility combined with erosion/abrasion resistance), and gecko feet (for improved adhesion) [ 24 , 63 , 67 ]. Some species feature adaptive systems that enable changes in color, pattern, or texture for defense, signaling, temperature regulation, or reproduction (also called the ‘chameleon mimetic system’) [ 68 , 69 ]. Nacre, an organic material found in mollusk shells, exhibits remarkable strength and resilience [ 70 ]. Spider silk, renowned for its exceptional mechanical properties and supercontraction abilities, holds great potential for various structural applications [ 71 ]. Examining beetle adhesion systems at the nanoscale has uncovered a diverse array of intricate multiscale architectures serving crucial functions like wing fixation, crawling, mating, and external protection, mainly utilized regarding locomotion for microrobots (anisotropic friction and earthworm-inspired) [ 30 , 34 , 72 ]. These adhesion systems utilize different mechanisms; some rely on liquid secretion (capillary force and lubrication), while others operate through direct interlocking of high-density microfibers or contact of mushroom-shaped hairy structures (van der Waals force) [ 70 ]. A comprehensive analysis of materials (structures) inspired by nature reveals their unique functionality, encompassing various domains of sustainable science [ 63 ]. Bioinspired structures combined with surface topography (or texturing) of these materials demonstrate promising outcomes in customizing friction, wear, and other captivating properties such as antifouling [ 73 , 74 , 75 ], self-lubrication [ 8 , 61 , 76 ], self-adaptation [ 77 , 78 ], hydrophobicity [ 27 ], self-cleaning [ 16 , 25 , 73 ], cell growth [ 12 , 79 ], drug delivery [ 80 , 81 , 82 , 83 , 84 , 85 , 86 ], antibacterial [ 79 , 87 , 88 ], color manipulation [ 68 ], anti-reflection [ 89 ], anisotropic friction in MEMS/NEMS/microrobots/microactuators [ 30 , 70 , 72 ], and enhanced adhesion [ 24 , 25 , 70 , 90 ]. Figure 2 shows various biological organisms (animals/plants) for inspiration regarding creating an efficient multifunctional tribological material. Achieving such complex nature-inspired geometries in materials or on their surface through conventional materials and technologies often proves challenging and expensive [ 92 , 93 , 94 , 95 ]. Historically, progress regarding materials (tribological) has primarily depended on modifying their chemical composition to alter their tribological and mechanical properties. Although this approach has yielded positive results, the journey from discovering new tribomaterial to its commercial availability has typically been time-consuming [ 92 ]. To this end, additive manufacturing (3D printing or laser surface texturing) offers promising prospects, especially in terms of generating complex bioinspired surfaces/materials and also allowing for faster fabrication and up-scaling production [ 94 , 95 , 96 , 97 , 98 , 99 ]. Any bottom-up approach to developing complex, multifunctional, and multiscale mimetics that can provide multiple robust functionalities requires a multidimensional strategy. The research in this multifaceted tribology domain is still at an early stage. The purpose of this review is to provide comprehensive information on the latest advances and future prospects in the field of “nature inspired” or “biologically inspired” tribological materials, with particular emphasis on their design, manufacturing processes, sources of inspiration, as well as friction and wear, performance, and other additional functions. The key focus is to establish a correlation between various tribological scenarios, bioinspired material characteristics, and their resilience to specific environmental conditions, aiming to formulate guidelines for the creation of bioinspired wear-resistant materials and systems. We have attempted to sum up relevant tribological information to a large extent. Nevertheless, due to some missing parameter specifications from the relevant reference publications that do not clearly and unambiguously characterize the physical conditions or tribological stress or the fact that tribological behavior is a system response and mere representation of some value such as force, rotational frequencies, etc., it is not in itself (alone) suitable for characterizing the stress level of the tribological system in question. Also, mixed or incorrect use of data/units for mass and forces should be avoided or carefully analyzed. For this, we recommend that readers defer to the relevant cited source(s). Figure 3 illustrates the research methodology framework for the current work."
} | 4,020 |
36071037 | PMC9452534 | pmc | 2 | {
"abstract": "Light-induced microbial electron transfer has potential for efficient production of value-added chemicals, biofuels and biodegradable materials owing to diversified metabolic pathways. However, most microbes lack photoactive proteins and require synthetic photosensitizers that suffer from photocorrosion, photodegradation, cytotoxicity, and generation of photoexcited radicals that are harmful to cells, thus severely limiting the catalytic performance. Therefore, there is a pressing need for biocompatible photoconductive materials for efficient electronic interface between microbes and electrodes. Here we show that living biofilms of Geobacter sulfurreducens use nanowires of cytochrome OmcS as intrinsic photoconductors. Photoconductive atomic force microscopy shows up to 100-fold increase in photocurrent in purified individual nanowires. Photocurrents respond rapidly (<100 ms) to the excitation and persist reversibly for hours. Femtosecond transient absorption spectroscopy and quantum dynamics simulations reveal ultrafast (~200 fs) electron transfer between nanowire hemes upon photoexcitation, enhancing carrier density and mobility. Our work reveals a new class of natural photoconductors for whole-cell catalysis.",
"introduction": "Introduction Living cells have been incorporated with quantum dots and nanostructures for fluorescent labelling and drug delivery for over two decades 1 . However, light-absorbing nanostructures have not been used to drive catalytic reactions inside of cells due to lack of biocompatibility and high cytotoxicity of foreign materials, such as photosensitizers, inside the cell which often limits the operational efficiency 1 . Furthermore, inherent defects in synthetic photosensitizers cause several problems such as photocorrosion, photodegradation and the generation of photoexcited radicals, which results in low stability, irreproducibility and lack of sustainability of biohybrid materials 2 . Some bacteria produce light-absorbing centers but suffer from low electron transfer efficiency and a lack of durability 1 . Natural electron transfer proteins such as azurins, myoglobin and c-type cytochromes do not show photoconductivity 3 , 4 due to picosecond carrier lifetimes of the heme iron which typically inhibits any charge separation 5 . Covalently linking artificial photosensitizers to these proteins yields low electron transfer rate on the 10 ns timescale or slower, greatly limiting their applications 6 . Furthermore, it is not feasible to use longer excited state lifetimes, such as electron injection from the triplet states due to rapid degradation caused by reactive oxygen species produced in these processes. Therefore, there is an urgent need to develop novel biomaterials capable of ultrafast primary electron transfer to achieve efficient charge separation, followed by sequential secondary electron transfer for long-lived charge separation and charge accumulation 6 . To evaluate the use of engineered living materials as living photoconductors, we chose the electroactive soil organism Geobacter sulfurreducens because it has evolved the ability to export electrons, derived from metabolism, to extracellular acceptors such as metal oxides and electrodes in a process called extracellular electron transfer (EET) 7 , 8 . Bacteria establish direct electrical contact to electron acceptors via micrometer-long, polymerized cytochrome nanowires, called OmcS, which eliminates the need for diffusive redox mediators 7 , 8 (Fig. 1b ). Hemes in the OmcS nanowire form a parallel, slipped-stacked pair with each pair perpendicular (T-stacked) to the next pair, forming a continuous chain over the entire micrometer length of the nanowire 7 (Fig. 1d ). The minimum edge-to-edge distances is 3.4–4.1 Å between the parallel-stacked hemes and 5.4–6.1 Å between the T-stacked pairs. Fig. 1 Living photoconductors. a Measurement schematic. Biofilms are grown on transparent fluorine-doped tin oxide (FTO) electrodes. b Transmission electron microscopy of CL-1 cells producing OmcS nanowires. Scale bar, 200 nm. c AFM height image of a single OmcS nanowire on mica (left) and respective height profile (right) shown where the red line is indicated. Scale bar 50 nm. d Hemes in OmcS stack seamlessly over the entire micrometre-length of nanowires. Edge-to-edge distances are in Å. e UV-Visible spectroscopy of biofilm on FTO electrode with the excitation wavelength of 408 nm marked as a purple triangle. f Current voltage response of biofilm with the laser on and off. Percentage increase in conductance value represents mean ± standard deviation (S.D). of two biological replicates. Source data are provided as a Source Data file. Owing to this evolutionarily optimized OmcS nanowire structure with seamless stacking of hemes, G. sulfurreducens can transfer electrons over distances of one hundred times their size by forming more than 100 µm-thick highly-conductive nanowire networks in biofilms 9 , 10 , which enables G. sulfurreducens cells to generate the highest current density in bioelectrochemical systems 11 . Owing to large electron storage capacity, cytochromes also confer high supercapacitance to biofilms with low self-discharge and reversible charge/discharge 12 . Moreover, a network of purified nanowires can transfer electrons over distances of 10,000-times the size of a cell 9 . Therefore, G. sulfurreducens serves as an ideal model system for electrocatalysis, metal corrosion and production of fuels 13 , 14 . It was previously thought that conductive filaments on the surface of G. sulfurreducens are pili 15 and a network of pili confers conductivity to G. sulfurreducens biofilms 10 , 13 , 16 . However, structural, functional and subcellular localization studies revealed that nanowires on bacterial surface are composed of cytochromes 7 , 8 whereas pili remain inside the cell during EET and are required for the secretion of cytochrome nanowires to the bacterial surface 17 , 18 . Nanowires could be widespread and their photophysical properties could be physiologically important because many Geobacter -like metal-reducing bacteria form highly conductive biofilms 19 , 20 and are widely distributed at the surface of earth in shallow sediments that contain abundant sunlight and metal oxides 21 – 23 . The sediments are capable of transporting electrons over centimeters 24 and can convert incident light into electricity 25 . Illuminating visible light on G. sulfurreducens cells has been shown to improve their catalytic performance such as increases in metabolic electron transfer to metal oxides 26 or other semiconducting materials by over 8-fold compared to that observed under dark conditions 21 . Furthermore, light-induced bacterial electron transfer correlated well with the rates of microbial respiration and substrate consumption 26 . However, the underlying molecular and physical mechanism for this increased photocatalytic performance has remained unclear. In addition to light-induced whole-cell catalysis 21 , 26 , artificially expressing cytochrome OmcS in photosynthetic cyanobacteria, increased catalytic performance in diverse processes such as an increase in photocurrent by 9-fold 27 , increase in nitrogen fixation by 13-fold 28 , and improved photosynthesis due to 60% increase in biomass 29 compared to the wild-type cyanobacteria. These studies highlight the vital role of OmcS in light-driven biocatalysis. However, intrinsic photophysical properties of OmcS, which could account for these catalytic improvements, have not been investigated. Out of 111 cytochromes in G. sulfurreducens , OmcS is the only nanowire-forming cytochrome essential for EET to Fe(III) oxides abundant in subsurface 14 . Indeed, cytochromes abundant in subsurface during uranium bioremediation function similar to OmcS 30 . OmcS is also important for EET to electrodes during initial stages of biofilm growth 14 . OmcS is also required for interspecies electron transfer in Geobacter cocultures to conduct “electric syntrophy” 13 , 31 , 32 . This interspecies electron transfer via naturally conductive microbial consortia is important in diverse methanogenic and methane-consuming environments that affect global climate 33 – 35 . Photosynthetic bacterial species have also been shown to perform electric syntrophy with light-driven conversion of CO 2 to value-added chemical commodities 2 . However, the components and pathways responsible for such light-driven biocatalysis have not been identified and potential for photoactivity beyond photosynthetic microorganisms remains largely unknown. We hypothesized that cytochrome nanowires in the biofilms could be photoactive, enabling efficient electronic interface between microbes and electrodes. Here we show that living biofilms of Geobacter sulfurreducens use nanowires of cytochrome OmcS as intrinsic photoconductors. Surprisingly, nanowires show photoconductivity with ultrafast, sub-picosecond heme-to-heme electron transfer which could explain their influence on photocatalytic performance mentioned above. These rates are among the highest for excited-state electron transfer in biology 36 .",
"discussion": "Results and Discussion Photoconductivity in living biofilms made up of OmcS nanowire network To determine the role of OmcS nanowires in light-induced electron transfer, we used the genetically engineered G . sulfurreducens strain CL-1 because it overexpresses OmcS nanowires (Fig. 1b–d ) and forms highly conductive and cohesive biofilms that can be easily transferred to multiple surfaces 37 (Fig. 1a ). Upon laser photoexcitation (λ = 408 nm) which is specific to the Soret band of c-type hemes 4 , biofilm conductance remained ohmic and increased by 72 ± 21% (Fig. 1e, f ). These studies show that living G . sulfurreducens biofilms can serve as intrinsic photoconductors. As biofilm conductivity determines the bacterial rate of EET 11 , our results could explain the increased photocatalytic performance by G. sulfurreducens 21 , 26 . Rapid photoconductivity in purified nanowires reversible for hours To determine the origin of photoconductivity in biofilms, we purified nanowires from the CL-1 strain (Fig. 2a ). The ultraviolet-visible (UV-Vis) absorbance spectrum of nanowires showed a strong Soret band at 410 nm for air-oxidized nanowires (Fig. 2b ). Nanowires were fully oxidized under these conditions because addition of oxidant (ferricyanide) did not change the spectrum (Supplementary Fig. 1a ). We placed the nanowires on interdigitated gold electrodes and illuminated from the top (Laser Power = 100 mW/cm 2 ). Photoconductance of nanowire network initially increased more than 6-fold (Fig. 2c ), but the extent of conductance increase decreased over time, likely due to laser damage. Nanowires responded faster than 100 ms (Fig. 2c inset). The photoresponse persisted for hours but decreased over time (Fig. 2c ). Both the dark current and photocurrent were proportional to an applied voltage ranging from –0.2 to +0.2 V (Fig. 2d ), indicating an ohmic conduction behavior of nanowires similar to biofilms. Remarkably, nanowire networks, with and without laser excitation, showed a linear current-voltage response with an average conductance increase of 230 ± 28 % ( n = 7), which is higher than common perovskites 38 , 39 and porphyrin nanowires 40 (Fig. 2d, e ). Fig. 2 High photoconductivity in purified protein nanowires. a Heme staining gel of nanowires showing a single band of OmcS. b UV-Vis spectrum of oxidized (green) and reduced (red) nanowires. c Photocurrent response of nanowire network at 200 mV with the current decay of the off-state subtracted. Inset : Fast (<100 ms) photoresponse of nanowires. Axes are same as in Fig. 2c. d Current-voltage response of nanowire network and cytochrome c for comparison e Comparison of conductance of nanowire network with laser on or off. Values represent mean ± standard error of the mean (S.E.M) with individual data points shown as grey dots (n = 7 independent experiments). ** indicates p value = 0.003 using a paired two tail t -test. f Schematic of pc-AFM of individual nanowires. g Current-voltage response of an individual nanowire with a linear fit shown by a purple dashed line. h Comparison of conductance increase upon photoexcitation in individual nanowires. Values represent mean of all current-voltage curves measured on individual nanowires (number of curves ranges from 10 to 120 Supplemental Table 2 ). i Comparison of average conductance of individual nanowires with laser on or off. Values represent mean ± S.E.M. with individual data points shown as grey dots ( n = 15 independent experiments). ** indicates p value = 0.007 using a paired two tail t -test. Source data are provided as a Source Data file. Multiple control experiments confirmed that the observed photoconductivity is an intrinsic property of nanowires, owing to their polymerized cytochrome architecture. For example, the monomeric horse-heart cytochrome-c showed very low dark current and photocurrent as expected 3 , 4 when measured under identical conditions (Fig. 2d ). Upon addition of a chemical reductant sodium dithionite, the Soret band for reduced nanowires red-shifted to 420 nm as expected 4 (Fig. 2b ). These chemically reduced nanowires (λ Soret = 420 nm) did not show significant photoconductance upon excitation at λ = 405 nm, confirming that photoreduction of oxidized hemes are necessary for photoconductivity in nanowires at this excitation (Supplementary Fig. 1b ). Switching the electrode material from gold to tungsten also retained photoconductivity, confirming that the measured response is not an artifact of the electrode material (Supplementary Fig. 2 ). The ratio of laser-on/ laser-off (on/off) current of nanowires increased with increasing laser power, further demonstrating that the measured photoconductivity is solely due to laser excitation (Supplementary Fig. 3 ). All these experiments together confirm that the nanowires show intrinsic photoconductivity which can account for observed photoconductivity in living biofilms. The difference in photoconductivity between biofilms and purified nanowires is likely due to non-conductive materials such as cells and polysaccharides present in the biofilms. Individual nanowires show up to 100-fold photoconductivity increase To quantify the photoresponse of individual nanowires, we used photoconductive atomic force microscopy (pc-AFM) 41 (λ = 405 nm, Initial Laser Power = 3.20 kW/cm 2 , Fig. 2f ). Individual nanowires showed up to 100-fold increase in conductance upon photoexcitation (Fig. 2h–i , Supplementary Table 2 ). The differences in photoconductance are likely due to variation in the laser power caused due to experimental setup (see methods and Supplementary Fig. 11 for details). The difference in the photoconductivity between individual nanowires and nanowire network is likely due to inter-nanowire as well as nanowire-electrode contact resistance. Notably, the observed 10 to 100-fold increase in conductance for protein nanowires at relatively low bias (< 0.5 V) is substantially greater than that of synthetic porphyrins 42 that show only up to a 5-fold increase at very high bias of 12 V. These experiments on individual nanowires confirm that the observed photoconductivity response in networks of nanowires is due to nanowires alone and not because of an artifact of the measurement setup. Furthermore, the observed photoconductivity is not due to heating effects because all pc-AFM experiments were performed in a temperature-controlled environment thus inhibiting any substantial increase in temperature. Furthermore, the linearity and stability of our IV curves indicate that measured conductivity increase is not due to heating (Fig. 2g ). In addition, the conductivity of OmcS nanowires decreases upon heating 43 whereas we observed up to 100-fold increase in conductivity upon photoexcitation. fs-TA revels sub-picosecond charge separation in nanowire To understand the mechanism of photoconductivity in protein nanowires, we performed femtosecond transient absorption (fs-TA) spectroscopy by determining the electron dynamics upon photoexcitation on an ultrafast (~100 fs) time scale 5 , 44 (Fig. 3a ). The fs-TA tracks the UV-Vis spectral changes by changing the time delay Δτ between the femtosecond laser pump and the probe pulses and recording a differential absorbance spectrum (ΔA) at each time delay 44 (Fig. 3a ). This difference spectrum contains information on the dynamic processes occurring in the system such as excited state energy migration, electron or proton transfer processes and isomerization 44 . In contrast to the above studies of photoexcitation in the Soret band (Figs. 1 , 2 ), we performed fs-TA using excitation in the Q-band (λ = 545 nm) to avoid thermal damage, and to monitor changes in the region of the strongest absorption bands 44 . It is important to note that Soret and Q-band transitions arise from the same ground state making Q-band excitation a suitable proxy to monitor these processes 44 . Photoexcitation with λ = 530, 545 and 400 nm yielded similar dynamics, demonstrating a wide spectral range for photoconductivity (Supplementary Figs. 4 , 5 ). Neither buffer alone nor the blank substrate showed any response, measurements in solid and liquid state are similar, and the ET dynamics were independent of the laser intensity and power (Supplementary Figs. 6 , 7 , 8 ) indicating that observed dynamics are due to nanowires and not an artifact of the environment or the substrate. Fig. 3 Ultrafast (<100 fs) charge transfer between hemes in nanowires revealed by femtosecond transient absorption spectroscopy (fs-TA). a Schematic of fs-TA. A pump beam (λ = 545 nm) excites a nanowire sample and is followed by a probe beam after a time delay. The differential absorption between the initial and time-delayed spectra is detected and reported as optical density. b Averaged transient absorption data of nanowires ( n = 6 independent experiments) where colours represent the milli optical density (mOD). c Normalized change in differential absorption with wavelength at different delay times. Key wavelengths are marked as λ = 410 nm (green), λ = 424 nm (red), and λ = 367 nm (blue). d The experimental (solid) and simulated (dashed) spectra of oxidized, reduced, and singlet doubly-oxidized nanowires. Wavelength markers are same as in Fig. 3c. e Normalized change in differential absorption over delay time at key wavelengths. Time-markers are shown in the same colour as time traces in Fig. 3c. Traces in c and e represent mean of n = 6 independent experiments. Source data are provided as a Source Data file. Upon photoexcitation of protein nanowires, electrons are promoted from the ground state to the excited state, decreasing the ground state population. This decrease caused a negative signal in ΔA at 410 nm known as a ground state bleach 44 (Fig. 3b, c ). In addition, we observed a positive ΔA which is indicative of excited-state absorption at λ = 367 nm and λ = 424 nm after Δτ = 0.1 ps and 2 ps, respectively (Fig. 3c, d ). These absorptions are absent in the native, air-oxidized, unexcited nanowires (Fig. 3d ), indicating that the photoexcitation is causing these absorptions. In particular, the absorption at λ = 424 nm agrees well with the absorption of chemically reduced nanowires (Fig. 3d ), suggesting that upon photoexcitation, excited-state electron transfer is reducing the hemes in the nanowires and thus photoreduction contributes to nanowire photoconductivity. To understand the origin of different transient oxidation states, we determined the kinetics at the key wavelengths mentioned above using a sequential model 44 that yielded the first excitation timescale of 19 ± 23 fs (see methods). This timescale is faster than the instrument response function (100 ± 50 fs) and can, therefore, be treated as an instantaneous excitation on the timescale of the measurement (Fig. 4a ). Following this excitation, the charges were transferred between hemes with a decay time of 212 ± 27 fs. The corresponding spectra are a superposition of a ground state bleach and the appearance of a new feature around 367 nm, which can be attributed to the doubly oxidized hemes as per the spectral simulations (Fig. 3d ). Based on simulations (see below, Fig. 4d ), we conclude that the ultrafast charge transfer also results in the formation of a reduced heme in its excited state, which is spectroscopically dark. We also found a second decay with a time constant of 1.0 ± 0.1 ps that can be attributed to relaxation of the excited reduced heme that increases in absorption at the reduced Soret at λ = 424 nm. In addition, we found a third decay time of 7.9 ± 0.3 ps that can be attributed to the recombination to their initial state, including charge transfers back to the singly oxidized heme ground states. Fig. 4 Model for origin of photoconductivity in protein nanowires. a Simplified energy level diagram for hemes depicting the changes that occur upon photoexcitation in transient absorption and their respective decay times. b The dark current in the ground state arises due to propagation of a reduced state created by electron injection from the electrode. c The photocurrent is due to the laser excitation initiating an ultrafast charge transfer between hemes, creating newly reduced (red) and double oxidized hemes (blue). The photoreduction provides additional charge carriers and larger driving force for charge transfer, which therefore increases the current under bias. d Quantum dynamics simulations of ultrafast charge transfer between hemes in protein nanowires, forming a doubly oxidized heme and an excited state of a reduced heme. Computations suggest excited state electron transfer between hemes We further compared our experimental UV-Vis spectra of nanowires with time-dependent density functional theory calculations of hemes in the nanowire (Fig. 3d ). The maximum of the computed Soret band in the reduced heme (λ = 420 nm) is shifted by 9.5 nm to the red of the band maximum for the oxidized heme, in good agreement with the 10.5 nm shift observed experimentally (Fig. 3d ). These computational analyses further suggest that photoexcitation causes reduction of hemes in the nanowires. Our finding is consistent with prior studies of photoreduction of monomeric cytochromes mediated by the light-induced excited state of hemes even in the absence of external electron donors 45 , 46 . To evaluate the transient kinetics data obtained using the sequential model fitted to experimental data, we performed quantum dynamics simulations at the Extended Hückel level of theory 47 , 48 . We simulated the propagation of an electron wavepacket in the excited state from hemes in the nanowires, in a slip-stacked as well as in T-stacked orientation 7 . Our simulations suggest a ~100 fs timescale for photo-induced charge transfer between the slip-stacked pair of hemes (Fig. 4d ). This timescale agrees with the experimentally determined timescale for the excited state charge transfer (212 ± 27 fs). The survival probability for electron transfer in a slipped-stack heme pair remains low (<60%) for most energy levels, indicating a high probability for electron transfer to a nearby heme within 100 fs (Supplementary Fig. 9 ). Thus, the timescale for electron transfer between slipped-stack heme pairs remains similar for most energy levels. Hemes are likely electron source for observed photoreduction As no external electron donor was added, our results suggest that the additional electrons that reduce the heme are intrinsic to the nanowire itself. We further analyzed the possibility that surrounding protein causes the observed photoreduction of OmcS hemes. Several aromatic amino acids, including tryptophan and tyrosine, are within 5 Å of the hemes in the OmcS. Although excitation of either tryptophan or tyrosine is not possible at the wavelengths used in this study 45 , 46 , we considered a possibility that electron transfer can quench a photoexcited heme in a manner similar to flavins in a cryptochrome 49 . This quenching would reduce a heme and leave behind an amino acid radical. The most likely amino acid candidate for radical formation is tryptophan because its radicals have absorbance which would explain the 367 nm species 50 . While the formation of such radicals is possible, the signal strength in fs-TA measurements is determined by the molar extinction coefficients (ε) of the (transient) species. The molar extinction coefficient of the Soret band for OmcS is approximately 100 times larger than those of tryptophan radicals 50 , 51 . The ground state bleach represents all the photoexcited hemes in the OmcS nanowires and the species corresponding to λ = 367 nm and 424 nm have differential absorptions of ~20 and 10% of the total magnitude, respectively (Fig. 3e ). Therefore, the number of tryptophan radicals created from electron transfer needs to be larger than the number of excited hemes in OmcS if the radical species at λ = 367 nm arises from tryptophan. Such a possibility seems unlikely because only one radical can be created for every quenched excited heme. Thus, the observed spectra cannot be accounted for by amino acid radicals. We also evaluated the possibilities of other electron sources causing photoreduction. We found that multi-photon processes are absent in our experiments because the ET dynamics were independent of the laser intensity and power (Supplementary Fig. 8 ). The magnitude of photocurrent is also linear with increased power (Supplementary Fig. 3 ). Redox impurities also did not contribute to the measured spectra because of identical dynamics in solution and in solid-state (Supplementary Fig. 7 ). Photodegradation also did not change the electron transfer dynamics, only the magnitude of spectra by <10% over two hours. We therefore considered an alternative possibility that parallel-stacked hemes can serve as an electron donor and acceptor pair (Fig. 4 ). We hypothesized that the excited state charge transfer is occurring between two neighboring hemes with only one of the hemes being in the excited state. Such charge transfer would result in the appearance of a reduced heme and leave behind a doubly oxidized heme (Fig. 4 ). The computed UV-Vis spectrum of a doubly oxidized heme indeed showed an absorption maximum at λ = 365 nm which agrees with the experimentally observed species at λ = 367 nm. Our computed spectrum of a doubly oxidized heme thus recaptures the blue shift observed in the transient absorption experiment (Supplementary Fig. 10 ). The qualitative agreement between the computed and experimental spectra is independent of the spin state of doubly-oxidized species such as the singlet and triplet state. To identify the nature of doubly oxidized species, we performed an analysis of atomic spin populations. We found that the change in the spin populations occurs only on the ligands and not in the iron center. Therefore, our analysis suggests that doubly oxidized species are Fe 3+ + porphyrin radical which agrees with the observed spectra at 367 nm. These analyses further suggest that the doubly oxidized species are not Fe 4+ due to lack of change of spin density on the iron center upon additional oxidation of the heme in the Fe 3+ state (Supplementary Fig. 10 and Supplementary Table 1 ). To further evaluate the thermodynamic feasibility of radical heme species, we used the Rehm-Weller cycle. This analysis requires four energetic terms: (1) energy required to form radical heme species (based on iron-porphyrin systems 52 ) (1.7 V), (2) the ground state redox potential of OmcS (−212 mV) 51 , (3) the photon energy used to excite OmcS nanowires (λ = 545 nm = 2.3 eV), and (4) the vibrational energy difference between the ground and excited states, called the Coulomb stabilization energy associated with the intermediate radical ion pair 53 ( ω p ) ~60 meV. Therefore, the energetics of this process would be ΔG et = [1.7 eV–(−0.212 eV) + 0.06 eV]−2.3 eV = −0.4 eV. Thus, ΔG et < 0 for the formation of the radical heme species, making them energetically feasible. Our analysis is a lower estimate for the net energy available for the formation of the radical species. Therefore, in combination with our simulated analysis, our studies suggest that doubly oxidized species are Fe 3+ + porphyrin radical and nanowires are photoreduced by ultrafast light-induced heme-to-heme charge transfer. Proposed mechanism for ultrafast photoconductivity in OmcS nanowire Based on above results, we propose the following model for the origin of photoconductivity in OmcS nanowires (Fig. 4 ). This model is focused on the singlet states and not triplet states because these states are spectroscopically dark, and would be less pronounced due to their lower energies. As these nanowires transport charges through seamless stacking of hemes (Fig. 1c ), our prior experiments have shown that they can be treated as redox conductors, with the long-range charge transfer governed by a theoretically-predicted hopping mechanism with negligible carrier loss over micrometers 54 . All hemes in the nanowires are initially oxidized and in their ground state as confirmed by UV-Vis spectroscopy (Fig. 2b ). Upon applying a bias, electrons are injected from the electrode into the nanowire, creating a reduced state that travels through the nanowire (Fig. 4b ). The photoexcitation triggers an ultrafast charge transfer resulting in an additional reduced state that persists for picosecond timescale, without any applied bias, far away from the electrode (Fig. 4c ). This newly formed reduced state will have a mobility similar to the electrode-injected state as they both are present in the same nanowire with identical structure. Therefore, upon photoexcitation, the density of reduced states is increased, thus increasing the carrier density of the OmcS to generate photoconductivity in nanowires. The photoreduction observed in our fs-TA is consistent with this model. In addition to the higher carrier density due to photogenerated electrons, it is likely that the mobility of electrons increases upon photoexcitation due to increased driving force for charge transfer in the excited state of hemes 43 . Upon photoexcitation an electron is promoted from the ground state to an excited state. The ultrafast charge transfer between neighboring hemes creates a reduced-state heme in the excited state and a doubly oxidized heme (Fig. 4c, d ). The reduced-state heme can then relax from the excited to the ground state. Upon photoexcitation, the uniformly oxidized nanowire is thus partially reduced and partially double oxidized (Fig. 4c ). The generated doubly oxidized heme will alter the redox energies of the heme chain, with a more positive redox potential. We have previously found that the redox potential of OmcS hemes becomes substantially positive upon oxidation 43 . The OmcS nanowires transport charges via a hopping mechanism 54 —a process in which a charge (electron or hole) temporarily resides at a heme, changing its redox state. The driving force for charge transfer depends on the redox energies of the electron donating and accepting hemes. Therefore, the charge transfer rate is directly related to the mobility. For the fully oxidized (non-excited) state, this process initiates at the electrode surface where injected electrons hop to nanowire redox sites, creating locally reduced hemes. For the photoexcited state, this process is enhanced because transferring an electron to the double-oxidized species, and removing an electron from a reduced heme, are significantly more favorable in the illuminated nanowire than for the oxidized nanowire in the dark. The increased likelihood for charge transfer upon photoexcitation will then result in increased mobility. Furthermore, the initial ultrafast charge transfer between hemes increases the lifetime of the photogenerated state. Both the generation of a “new” mobile charge and the increase in its mobility will contribute to the observed increase in conductivity upon photoexcitation. In summary, we demonstrate, for the first-time, significant photoconductivity in a living system due to ultrafast light-induced charge transfer within protein nanowires. The surprising origin of photoconductivity in these natural systems lies in the higher carrier density and mobility upon photoexcitation. Although ultrafast electron transfer can occur in monomeric cytochromes, it typically requires incorporated dyes as photosensitizers and sacrificial electron donors 36 which can be toxic to cells 1 . In contrast, we find that the protein nanowires intrinsically exhibit robust and ultrafast charge transfer without any need for such site-selective labeling. Our studies thus establish OmcS nanowires as photoconductors intrinsic to cells with capability of ultrafast electron transfer, thus eliminating the need for foreign materials such as molecular dyes or inorganic nanoparticles that limit the catalytic performance 1 . Furthermore, our studies show that sub-ps charge transfer is possible in natural proteins in an excited state. Prior ultrafast electron transfer studies have reported the ground state rates of 15–90 ps in the closest-stacked hemes 36 . This difference is likely because excited-state rates are known to be faster due to higher energy and larger orbital delocalization compared to the ground-state rates 49 . Although many bacterial EET studies remain focused on electrons, protons play a very important role, not only in bacterial energy generation, but also in the electronic conductivity of proteins 55 . For example, through measurements of the intrinsic electron transfer rate, we previously found that both the energetics of a glutamine (proton acceptor) and its proximity to a neighboring tyrosine (proton donor), regulate the hole transport over micrometers in amyloids through a proton rocking mechanism 56 . Therefore, it is very important to couple electron/proton transfer to accelerate EET and for the development of electronically conductive protein-based biomaterials. The high surface area of these nanowires, combined with their biocompatibility and lack of toxicity, make them attractive candidates for an emerging field of light-driven whole-cell bioelectrocatalysis for a wide range of applications such as water splitting, chemical sensing and CO 2 fixation and production of chemicals, fuels and materials 57 . Our studies may also help establish the efficient and stable production of liquid fuels from sunlight using a liquid sunlight approach 5 . Future studies on nanowires with different heme stacking and protein environment 8 or substituting the metals from iron to zinc 58 or tin 59 could vary the interactions between the heme cofactors to alter the electronic and photophysical properties of nanowires for tuneable functionality 57 ."
} | 8,779 |
31057985 | PMC6497424 | pmc | 3 | {
"abstract": "The\nbiological production of FDCA is of considerable value as a potential\nreplacement for petrochemical-derived monomers such as terephthalate,\nused in polyethylene terephthalate (PET) plastics. HmfF belongs to\nan uncharacterized branch of the prenylated flavin (prFMN) dependent\nUbiD family of reversible (de)carboxylases and is proposed to convert\n2,5-furandicarboxylic acid (FDCA) to furoic acid in vivo. We present\na detailed characterization of HmfF and demonstrate that HmfF can\ncatalyze furoic acid carboxylation at elevated CO 2 levels\nin vitro. We report the crystal structure of a thermophilic HmfF from Pelotomaculum thermopropionicum , revealing that the\nactive site located above the prFMN cofactor contains a furoic acid/FDCA\nbinding site composed of residues H296-R304-R331 specific to the HmfF\nbranch of UbiD enzymes. Variants of the latter are compromised in\nactivity, while H296N alters the substrate preference to pyrrole compounds.\nSolution studies and crystal structure determination of an engineered\ndimeric form of the enzyme revealed an unexpected key role for a UbiD\nfamily wide conserved Leu residue in activity. The structural insights\ninto substrate and cofactor binding provide a template for further\nexploitation of HmfF in the production of FDCA plastic precursors\nand improve our understanding of catalysis by members of the UbiD\nenzyme family.",
"discussion": "Results and Discussion Initial Identification,\nExpression, and Characterization of Thermostable FDCA (De)carboxylases It has previously been reported that the thermophilic bacterium Geobacillus kaustophilus HTA426 is capable of degrading\nfurfural. 17 A BLAST search of the G. kaustophilus genome 18 using the C. basilensis Hmf gene\ncluster suggested the presence of a similar Hmf gene cluster located\non plasmid pHTA426. Although there is no mention in the literature\nregarding the ability of G. kaustophilus to degrade HMF, a C. basilensis HmfF\nhomologue (WP_011229502) could be located on pHTA426, possessing 51%\nsequence identity and located adjacent to a HmfG/UbiX homologue. Active\nrecombinant G. kaustophilus HmfF was\nsuccessfully produced in E. coli when\nit was coexpressed with E. coli UbiX\n( Figure S2 ). However, while soluble recombinant G. kaustophilus HmfF could be produced, the protein\nhad a tendency to aggregate, hampering crystallogenesis and other\nbiophysical studies. Other thermophilic HmfF homologues were screened,\nwith the P. thermopropionicum HmfF\nenzyme being the most promising in terms of protein expression levels\nand stability. The purified recombinant HmfF enzymes (from both P. thermopropionicum and G. kaustophilus ) were capable of decarboxylating 2,5-furandicarboxylic acid to furoic\nacid in vitro ( Figure S2b ) but could not\nfurther decarboxylate furoic acid to furan. Expression and Detailed\nCharacterization of P. thermopropionicum HmfF Purified PtHmfF expressed in absence of E. coli UbiX coexpresssion was pale yellow and possessed\na UV–vis spectrum consistent with oxidized FMN binding. In\ncontrast, when it was coexpressed with UbiX, the purified recombinant\nprotein was pale pink, possessing a complex UV–vis spectrum\nwith three main features in addition to the protein peak at 280 nm\n( Figure 2 A). These\ninclude a feature at 390 nm, similar to that observed previously for\nthe model system A. niger Fdc1, 11 a peak at 450 nm (likely corresponding to the\npresence of a subpopulation bound to oxidized FMN rather than prFMN),\nand a broad peak centered around 550 nm. Similar spectral features\nat 550 nm were previously identified as corresponding to the semiquinone\nradical form of the prFMN cofactor. 11 , 13 , 19 Figure 2 HmfF spectral properties, in vitro reconstitution, and\noxygen dependence of activity. (A) UV–vis spectra obtained\nfor heterologous expressed P. thermopropionicum HmfF. Spectra are shown of the WT protein expressed on its own (orange\nline) or coexpressed with UbiX and purified either aerobically (purple)\nor anaerobically (green). Spectra were normalized on the A 280 peak. The inset shows the closeup of the cofactor-related\nspectral features present in the 300–800 nm region. (B) UV–vis\nspectra of single expressed “apo” P.\nthermopropionicum HmfF as isolated (orange), following\nreconstitution with in vitro synthesized prFMN under anaerobic conditions\n(red), and following exposure to air (blue). (C) Activity of reconstituted\nPtHmfF against aerobic or anaerobic substrate before and after exposure\nto air. Assays were performed against 900 μM FDCA at 25 ° C (error bars represent SEM, n = 3). P. thermopropionicum HmfF in Vitro Reconstitution Confirms Oxidative Maturation Is Required\nfor Activity While UbiX produces prFMN in a reduced state,\nthe cofactor must undergo oxidative maturation within UbiD to produce\nthe active prFMN iminium form. 8 − 11 To investigate the requirement\nfor oxidative maturation of the cofactor in HmfF, apo -enzyme was reconstituted in vitro as described previously for AroY\nand UbiD. 13 , 19 Single expressed HmfF lacking decarboxylation\nactivity was reconstituted under anaerobic conditions and revealed\nprominent features at 360 and 530 nm ( Figure 2 B). Exposure to air resulted in an enhancement\nof the spectral features at 360–380, 450, and 530 nm, a range\nof spectral features suggestive of a mixture of normal oxidized FMN,\nprFMN radical , and possibly prFMN iminium , similar\nto that observed in the as isolated coexpressed enzyme. Consistent\nwith this, the anaerobic reconstituted protein displayed low levels\nof decarboxylase activity when it was assayed under anaerobic conditions.\nHowever, the rate of enzymatic decarboxylation was 5-fold higher when\nthe protein was assayed under aerobic conditions ( Figure 2 C). Taken together, these data\nconfirm that, as with Fdc1 and AroY, HmfF requires oxidative maturation\nof prFMN for activity. Pt HmfF Is Light and Oxygen\nSensitive The activity of the as-isolated coexpressed Pt HmfF was found to rapidly decrease over time when it was\nstored on ice. The loss in activity appeared to be partially due to\nlight exposure, with the half-life of Pt HmfF increasing\nfrom 35 to ∼100 min when it was stored in the dark under aerobic\nconditions ( Figure S3 ). Similar observations\nwere made for the A. niger Fdc1 enzyme,\nwhere light exposure was found to induce a complex isomerization of\nthe cofactor leading to inactivation. 20 However, unlike Fdc1, protection from illumination was not sufficient\nto stabilize Pt HmfF activity. In contrast, Pt HmfF stored under anaerobic conditions did not appear\nto lose activity over the course of several hours, suggesting that\ninactivation was also the result of O 2 , as observed for\nAroY. 13 Subsequently, Pt HmfF was either purified anaerobically or purified aerobically and\nthen reconstituted in vitro and assayed under anaerobic conditions.\nThe Pt HmfF enzyme activity was found to have a pH\noptimum between 6 and 6.5, with a temperature maximum of ∼60\n°C ( Figure 3 ).\nHowever, from 55 °C and above the activity decreased rapidly\nover the course of a few minutes, indicating that the enzyme was being\ninactivated, making it difficult to obtain accurate initial rates.\nThermal denaturation of Pt HmfF monitored using CD\nspectroscopy revealed a melting temperature of ∼68 °C\n( Figure S4 ). Thus, all subsequent assays\nwere performed at 50 °C. An Arrhenius plot of the 25–50\n°C data points indicated an activation energy of 80.7 kJ mol –1 . At pH 6 and 50 °C, the apparent K m and k cat values for FDCA\nwere 49.4(±3.7) μM and 2.39(±0.05) s –1 , respectively ( Figure 3 C). The Pt HmfF enzyme was also found to have minor\nactivity with 2,5-pyrroledicarboxylic acid (PDCA). In contrast, no\ndecarboxylation could be detected for 2,3-furandicarboxylic acid,\n5-formyl-2-furoic acid, 5-hydroxymethyl-2-furoic acid, 5-nitro-2-furoic\nacid, 2,5-thiophenedicarboxylic acid, 2,6-pyridinedicarboxylic acid,\nterephthalic acid, isophthalic acid, or muconic acid. Figure 3 Pt HmfF\nenzyme activity. (A) Effect of pH on activity. (B) Effect of temperature\non activity. Inset: Arrhenius plot of data. (C) Steady-state kinetic\nparameters obtained for P. thermopropionicum HmfF against FDCA (blue) and PDCA (red) at 50 °C and pH 6.\nError bars represent SEM, n = 3. Pt HmfF Catalyzes H/D Exchange of a Small Range of\nHeteroaromatic Acids It has previously been shown that UbiD\nenzymes are capable of catalyzing deuterium exchange of substrates\nthat can undergo UbiD-mediated carboxylation. 20 , 21 1 H NMR showed that incubation of furoic acid with Pt HmfF in D 2 O resulted in depletion of the resonance\npeak at 7.6 ppm consistent with exchange of the proton in the 5-position\n(denoted H a ) with a deuteron ( Figure S5 ). This was further supported by a change in splitting of\nthe 6.5 ppm resonance (corresponding to H b ) from a doublet\nof doublets to a doublet resulting from the loss of coupling between\nH b and H a . Partial H/D exchange of the 5-position\nof pyrrole-2-carboxylate (∼30%, Figure S4B ) could also be observed under the conditions tested; however,\nno exchange of thiophene-2-carboxylate was detected ( Figure S5C ). These observations confirm that the level of\nH/D exchange follows the same trend as observed for the level of decarboxylation\nof the corresponding diacids. With this in mind, we used H/D exchange\nto assay Pt HmfF against substrates where the corresponding\ndiacids were not commercially available. The proton in the 5-position\nof 2-oxazolecarboxylic acid could only be readily exchanged for deuterium\nin the presence of enzyme ( Figure S5D ).\nIn contrast, no enzyme-dependent exchange could be observed for position\n2 of 5-oxazolecarboxylic acid ( Figure S5E ). Pt HmfF and Gk HmfF Catalyze\nFuroic Acid Carboxylation at Elevated [CO 2 ] The\nHmfF reverse reaction, carboxylation, has been demonstrated to occur\nin vivo for distinct UbiD members that function as dedicated carboxylases, 22 − 24 while those family members that act as decarboxylases under physiological\nconditions (such as AroY and Fdc1) can catalyze carboxylation in vitro\nat elevated levels of CO 2 . 11 , 13 , 14 To investigate the ability of HmfF enzymes to catalyze\nthe reverse reaction, carboxylation of furoic acid to produce FDCA,\npurified Pt HmfF and Gk HmfF enzymes\nwere incubated with 50 mM furoic acid and 1 M bicarbonate at 50 °C\novernight. HPLC analysis of the reaction mixtures revealed a peak\nwith retention time of 2.3 min that comigrates with an FDCA standard\n( Figure 4 A). Mass spectrometry\nconfirmed that this species had a mass of 154.99 Da, consistent with\nthe expected mass for FDCA. We sought to determine whether performing\nthe reaction under pressurized CO 2 could increase the amount\nof carboxylated product. Reaction mixtures containing 50 mM furoic\nacid were incubated overnight with HmfF at 50 °C. In the presence\nof 1 M KHCO 3 , there was no significant difference between\nreaction mixtures incubated under N 2 at atmospheric pressure\nor under CO 2 at 32 bar with ∼2 mM FDCA produced.\nIn the absence of bicarbonate, ∼150 μM FDCA was produced\nunder 32 bar CO 2 , whereas no FDCA was detectable under\nN 2 ( Figure 4 B). While HmfF presents an attractive route to the production of\n2,5-furandicarboxylic acid, a potential bioreplacement for polymer\nprecursors, yields remain low even under high [CO 2 ]. Given\nthe unfavorable equilibrium for the carboxylation reaction, future\nefforts aimed at increasing the yield for this reaction will likely\nrequire in situ conversion of FDCA. Figure 4 Pt HmfF-catalyzed carboxylation\nof furoic acid to FDCA. (A) HPLC chromatogram demonstrating enzymatic\nproduction of FDCA by carboxylation of furoic acid by P. thermopropionicum HmfF. Chromatograms of FDCA\n(red) and 50 mM furoic acid in 1 M KHCO 3 solution incubated\nin the absence (blue) and presence (purple) of the Pt HmfF enzyme. Mass spectrometry confirmed that the species that comigrated\nwith the FDCA standard also possessed a mass consistent with FDCA.\n(B) Furoic acid carboxylation under CO 2 pressure. Assays\nwith or without 1 M KHCO 3 were incubated overnight either\nunder N 2 at atmospheric pressure or under CO 2 at 32 bar. Error bars represent SEM, n = 3. Pt HmfF\nCrystal Structures Reveal FMN Binding Mode To aid crystallization,\nthe Pt HmfF was expressed without affinity tag. The\nbest crystals obtained belonged to the P 2 1 space group and diffracted to 2.7 Å. The procedure was repeated\nwith Se-Met-substituted enzyme, and the structure was solved using\nSe-Met SAD, revealing a Pt HmfF hexamer ( D 3 symmetry) in the asymmetric unit. Although the UV–vis\nspectra of the purified enzyme used for crystallization trials indicated\nthe presence of cofactor, no electron density corresponding to the\ncofactor could be detected in preliminary electron density maps. Final\nrefinement was done using data collected to 2.7 Å on Pt HmfF crystals soaked with FMN (as a stable analogue of\nthe prFMN cofactor) in the presence of K + and Mn 2+ , revealing clear electron density for both the FMN and the associated\nmetal ions in the prFMN binding site ( Figure S6 ). The Pt HmfF structure is similar to other UbiD\nfamily member structures with a Z score of 47 with\nthe bacterial protocatechuate decarboxylase AroY (rmsd 1.6 Å\nover 441 C-αs), 13 45 with the E. coli UbiD (rmsd 2.5 Å over 440 C-αs), 19 40 with the fungal cinnamic acid decarboxylase\nFdc1 (rmsd 3.1 Å over 440 C-αs) 11 and 38 with recently solved TtnD decarboxylase involved in polyketide\nbiosynthesis (rmsd 2.7 Å over 414 C-αs). 25 The Pt HmfF monomer consists of an N-terminal\nprFMN binding domain connected via an α-helical linker to the\noligomerization domain ( Figure 5 ). The C-terminus consists of an extended loop region with\nsome α-helical character that interacts with the prFMN binding\ndomain of an adjacent Pt HmfF monomer. An overlay\nof the six Pt HmfF monomers reveals that minor variation\noccurs in the respective positions of the N-terminal prFMN binding\ndomain and the oligomerization domain, suggestive of domain motion\nvia the hinge region connecting both domains ( Figure 5 B). As the active site (vide infra) is located\nat the interface between both domains, this could be relevant to catalysis.\nThe phosphate moiety of the bound FMN is coordinated by Mn 2+ and K + ions (the identity of these was derived from the\nfact they were added to the crystal and was not independently verified),\nwhile the isoalloxazine is positioned directly adjacent to the conserved\nE(D)-R-E ionic network of residues conserved in UbiD ( Figure 5 C). In the case of Pt HmfF, the active site is only partly occluded from solvent\nas a consequence of the relatively open conformation of the N-terminal\nprFMN binding domain; this is similar to what has been observed for\nthe canonical UbiD and AroY enzymes ( Figure S7 ). To achieve full occlusion from solvent, as is observed for the\nfungal Fdc1 enzyme, a hinge motion (akin to that observed by comparison\nof the various Pt HmfF monomers) leading to a closed\nconformation would be required. Figure 5 Pt HmfF crystal structure.\n(A) Pt HmfF hexamer (D3 symmetry) shown in two orientations,\nrepresented in cartoon depiction, with the prFMN binding domain in\nblue, the connecting helix in magenta, the hexamerization domain in\ngreen, and the C-terminal helix in red. The bound FMN is shown as\nyellow spheres. Arrows indicate the interfaces disrupted by mutagenesis\n(vide infra). (B) Overlay of the six Pt HmfF monomers\npresent in the asymmetric unit with the C-α traces depicted\nin ribbon using a color coding similar to (A). (C) Side-by-side comparison\nof the Pt HmfF active site with other structurally\ncharacterized UbiD family members. Key residues are show in atom color\nsticks, with carbons colored according to domain structure as used\nin (A). In the case of the Aspergillus niger Fdc1 enzyme, the α-fluorocinnamic acid complex is shown, with\nthe substrate shown in cyan carbons. In the case of the TtnD enzyme,\nthe loop containing residues E272–E277 is not ordered in the\nFMN-bound structure. The conserved E(D)-R-E motif is highlighted by\nthe use of red labels. Pt HmfF Active Site Contains a Furoic Acid Binding\nMotif All attempts to acquire a crystal structure of the Pt HmfF in complex with substrate through either soaking\nor cocrystallization failed, a possible consequence of the open configuration\nof the enzyme. Guided by the structure of the related Fdc1 in complex\nwith cinnamic acid substrates, 11 FDCA can\neasily be placed into the active site of Pt HmfF in\na similar position with respect to the prFMN cofactor. This positions\nthe substrate furan oxygen approximately within hydrogen-bonding distance\nof His296 and locates the distal substrate carboxylate adjacent to\nArg304 and Arg331. All three putative substrate binding residues are\nconserved in the HmfF branch of the UbiD family tree ( Figure S1 ). To support our hypothesis regarding\nthe role of H296, R304, and R331 in substrate binding, we made Pt HmfF H296N, R304A, and R331A variants. All variants possess\nUV–vis spectra similar to that of the WT with the exception\nof the H296N variant ( Figure 6 A). In the latter case, cofactor related features between\n300 and 400 nm are less intense, indicating lower cofactor content.\nProlonged incubation of large quantities of protein with substrate\nresulted in complete decarboxylation of FDCA by WT Pt HmfF and the R304A variant (assayed by HPLC). Under these conditions,\nH296N was able to decarboxylate ∼87% of the substrate, in comparison\nwith 30% using the R331A variant ( Figure 6 B). Using PDCA as a substrate, only the WT\nand H296N were able to perform 100% decarboxylation ( Figure 6 C). A continuous spectrophotometrically\nbased assay (using 1 mM substrate) revealed that all three variants\nwere severely compromised in activity in comparison to the wild type\nenzyme, with k cat values 30–400\nfold lower than the WT against FDCA ( Figure 6 D). Interestingly, the H296N variant displays\na preference for 2,5-pyrroledicarboxylic acid over FDCA and has a\nslightly higher activity for PDCA in comparison to the WT enzyme ( Figure 6 D). Michaelis–Menten\nkinetics revealed that both R304A and H296N variants were not saturated\nat 1 mM substrate (the maximum possible under the experimental conditions),\nwhile no reliable data could be obtained for R331A ( Figure 6 E). These data clearly indicate\nthe FDCA affinity has been compromised by substitutions at positions\nH296, R304, and R331. Figure 6 Characterization of Pt HmfF variants.\n(A) UV–vis spectra of Pt HmfF variants including\nWT (blue), H296N (green), R304A (magenta), R331A (orange), and L403A\n(red). Spectra were normalized on the A 280 peak. The inset shows a closeup of the cofactor-related spectral\nfeatures present in the 300–800 nm region. (B) Decarboxylation\nof 10 mM 2,5-furandicarboxylic acid (FDCA) to furoic acid after overnight\nincubation with Pt HmfF variants. (C) Decarboxylation\nof 10 mM 2,5-pyrroledicarboxylic acid (PDCA) to pyrrole-2-carboxylate\nafter overnight incubation with Pt HmfF variants.\n(D) Rate of decarboxylation of 1 mM FDCA (blue) or 1 mM PDCA (red)\nby Pt HmfF variants. (E) Steady-state kinetics of Pt HmfF variants. Error bars represent SEM, n = 3. A Dimeric Pt HmfF Variant Binds prFMN but Is Compromised for Activity The resolution of the hexameric Pt HmfF structures\nobtained is limited and is in sharp contrast to the atomic resolution\nroutinely achieved for the dimeric A. niger Fdc1. A structural alignment of Pt HmfF hexamer\nwith the related A. niger Fdc1 dimer\nstructure demonstrates that Pt HmfF A315, N348, F351,\nT355, A388, F393, V395, and M399 form key hydrophobic interactions\nbetween the individual Pt HmfF dimers. In Fdc1, the\nequivalent positions are D343, R382, D385, N389, P424, T429, F431,\nand R435, respectively: i.e. generally larger and/or charged residues.\nWe created a Pt HmfF variant by substituting for the\ncorresponding A. niger Fdc1 amino acids\n(i.e., A315N, N348R, F351D, T355N, A388P, F393T, V395F, and M399R)\nto disrupt dimer–dimer interactions. SEC-MALLS of the purified Pt HmfF dimer variant indicated a native mass of 110 kDa,\nbroadly consistent with the expected mass of a dimer. Similarly to\nthe WT protein, the purified dimer variant possesses a complex UV–vis\nspectrum with three main features in the 300–800 nm region\n( Figure S8 ). This suggests the presence\nof prFMN, in addition to minor populations of FMN and the radical\nprFMN. Despite the presence of prFMN, the dimer variant display weak\nactivity, and incubation of 10 mM FDCA against 20 μM Pt HmfF dimer mutant only resulted in 30% decarboxylation\nfollowing overnight incubation ( Figure 6 B). Crystal Structure of Dimeric Pt HmfF Suggests a Key Role for a Conserved Leu in Activity The P. thermopropionicum HmfF dimer\nvariant was crystallized and the structure solved to 2.3 Å using\nmolecular replacement with the WT Pt HmfF monomer.\nUnlike the wild type enzyme, the dimer variant crystals contain prFMN\nin the active site. Despite the extensive mutation of the WT dimer–dimer\ninterface, the Pt HmfF dimer variant is very similar\nin structure to an individual dimer module from the WT hexamer. The\nprFMN is bound in a similar position and configuration as the FMN\nin the Pt HmfF hexamer, with little difference in\nthe position of the majority of active site residues. A notable exception\nis Leu403, which is located on a loop region that is disordered in\nthe Pt HmfF dimer variant and therefore absent from\nthe active site ( Figure 7 ). The 398–410 region including Leu403 is disordered in both\nthe Pt HmfF dimer variant monomers, a likely consequence\nof the M399R mutation and/or the disruption of the WT dimer–dimer\ninterface. As a consequence, the Pt HmfF dimer variant\nactive site is exposed to the solvent. To confirm whether the absence\nof Leu from the dimer active site contributed to the low activity\nof the dimer mutant, a L403A Pt HmfF was created.\nWhile the UV–vis profiles of both WT and the L403A variant\nare comparable, the latter had a k cat value\n∼40-fold lower than that of the WT ( Figure 6 D,E). However, unlike the H296 and R304/331\nvariants, the L403A K m value for FDCA\nwas not significantly different from the WT, suggesting that L403\ndoes not contribute to substrate binding. The hydrophobic nature of\nthe carboxylic acid binding pocket has been implicated in the mechanism\nof other decarboxylases, 26 , 27 and it is plausible\nthat the highly conserved Leu403 fulfils a similar role. Furthermore,\nLeu403 is one of the residues most affected by the proposed domain\nmotion ( Figure 5 B and Figure S7 ) that occludes the Pt HmfF active site from solvent. Figure 7 Pt HmfF dimer variant\ncrystal structure. (A) Pt HmfF dimer variant shown\nin cartoon representation, with color coding as in Figure 5 a. The mutations disrupting\nthe hexamer formation interface are indicated by cyan spheres for\nthe corresponding Cα positions. (B) Overlay of the two Pt HmfF monomers present in the asymmetric unit with the\nCα traces depicted in ribbon using a color coding similar to\nthat in (A). In addition, a single monomer of the Pt HmfF hexamer structure is shown in gray. (C) Position of the Leu403\nregion (in red) at the prFMN-domain/multimerization domain interface\nthat is disordered in the Pt HmfF dimer structure.\nMutations at the hexamer formation interface are shown as cyan spheres,\nexcept for M399R, which is shown in red. (D) Overlay of the respective Pt HmfF hexamer active site (in complex with FMN, in gray)\nand the Pt HmfF dimer variant in complex with prFMN iminium . For comparison, the position of the α-fluorocinnamic\nacid substrate of A. niger Fdc1 with\nrespect the prFMN cofactor is shown (in cyan), as well as the corresponding\nposition of L439 (homologous to Pt HmfF L403). Proposed Mechanism for\nHmfF Previous studies for the Fdc1 enzyme have suggested\nthat the reversible decarboxylation occurs via a 1,3-dipolar cycloaddition\nbetween substrate and prFMN. In principle, a similar reaction scheme\ncan be proposed for any of the UbiD substrates. However, for those\nsubstrates where (de)carboxylation occurs directly on an aromatic\nring system, cycloaddition also requires dearomatization. An alternative\nproposal has been put forward for AroY, on the basis of formation\nof a quinoid intermediate that allows formation of a substrate–prFMN\nadduct. In view of the modest aromatic nature of the furan ring, an\nFDCA or furoic acid adduct with prFMN could be formed through either\ncycloaddition (Ib in Figure 8 ) or via formation of an oxonium ion (Ia in Figure 8 ) in the case of HmfF. Chemical\nprecedent exists for the 1,3-cycloaddition of furans to 1,3-dipoles. 28 It is unclear at present which route is preferred\nfor HmfF, and this will require further investigation. However, it\nis interesting to note that HmfF-catalyzed H/D exchange can be readily\nobserved for weakly aromatic heteroaromatic acids only at those positions\nthat are adjacent to a carbon, hinting at the possibility that species\nIb is indeed formed during the enzyme reaction. Furthermore, neither\ndecarboxylation of the more aromatic thiophenedicarboxylic acid nor\nH/D exchange of thiophene-2-carboxylate was observed. In fact, thiophene\ncompounds are not known to readily undergo cycloaddition reactions,\nin contrast to furan. While HmfF is able to catalyze furoic acid carboxylation\nunder ambient conditions, this requires a mechanism for increasing\n[CO 2 ]. Other decarboxylases have been found to catalyze\npyrrole carboxylation in supercriticial CO 2 , 29 and we intend to explore whether HmfF (natural\nor evolved variants) can be used under these conditions. Further studies\nwill also need to address cofactor stability and homogeneity to ensure\na robust biocatalyst for the carboxylation of furoic acid. Figure 8 Proposal for\nthe HmfF mechanism. The HmfF substrate is bound by polar interactions\nwith the HmfF specific R304/R331 and H296, in addition to the UbiD\nfamily conserved R152 (part of the UbiD Glu-Arg-Glu motif). Substrate\nbinding is possibly linked to domain motion, affecting the relative\nposition of L403 and R331. Formation of a covalent prFMN iminium –substrate adduct can occur either through nucleophilic attack,\nleading to species Ia, or through cycloaddition, leading to species\nIb. Decarboxylation of either species leads to intermediate II, following\nE260/CO 2 exchange; protonation of the substrate via E260\nleading to product release occurs through either intermediate IVa\nor IVb."
} | 6,651 |
39759078 | PMC11700627 | pmc | 4 | {
"abstract": "Summary Largely varied anti-icing performance among superhydrophobic surfaces remains perplexing and challenging. Herein, the issue is elucidated by exploring the roles of surface chemistry and surface topography in anti-icing. Three superhydrophobic surfaces, i.e., gecko-like, petal-like, and lotus-like surfaces, together with smooth hydrophobic and hydrophilic surfaces, are prepared and compared in ice nucleation temperature under both non-condensation and condensation conditions. As a result, in non-condensation condition, water droplet freezing is caused by interfacial heterogeneous nucleation, wherein both surface chemistry and surface topography contribute to deferring freezing, and the former is dominant. In condensation condition, the freezing strongly correlates to condensation frosting. Surface chemistry maintains as a strong deterrent, whereas surface topography has two competing effects on the freezing. The paper deepens the understanding of water freezing on superhydrophobic surfaces, unravels the correlation between superhydrophobicity and anti-icing, and provides design guidelines on application-oriented anti-icing surfaces.",
"introduction": "Introduction Icing, a ubiquitous phenomenon, however brings about devastating disasters to air and road traffic, malfunction of solar cells and wind turbines, and plummeted crop production. 1 To this end, much effort has been devoted to anti-icing, among which superhydrophobic strategy is a widely adopted recipe to lower ice nucleation temperature (INT) and prolong freezing delay time (FDT) in undercooling conditions. 2 , 3 , 4 Although the anti-icing performance of superhydrophobic surfaces has been intensively studied over the last decade, it is still in dispute, which can be reflected from a wide INT gap of over 10°C and two orders of magnitude deviation for FDT. 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 Apparently, the varied anti-icing performance was derived from different superhydrophobic genres. Superhydrophobic surfaces are realized via regulating both surface chemistry and surface topography. 13 , 14 Employment of low-surface-energy materials or coatings is necessary for superhydrophobicity. Designs of surface topographies owning multi-scale hierarchical roughness and trapped air pockets are also indispensable. Since the revelation of dual-scale nano-/micro-textures for lotus leaves, a plethora of biomimetic superhydrophobic surfaces have sprouted up, possessing distinctive topographies, e.g., lotus-like, petal-like, and gecko-like. 15 , 16 , 17 , 18 , 19 Generally, such characters of superhydrophobic surfaces benefit the anti-icing performance by lifting the energy barrier for ice nucleation and/or lowering the heat transfer between surfaces and water droplets. 20 , 21 , 22 , 23 , 24 Moreover, the varied anti-icing performance of superhydrophobic surfaces was tested in different conditions. The effects of droplet size and cooling rate can be well explained by classical nucleation theory (CNT). 25 In addition, relative humidity (RH) has a profound influence on the freezing of water droplets. On one hand, high RH probably destabilizes the trapped air pockets and increases surface-droplet contact in a direct manner. 26 , 27 , 28 On the other hand, the ambient water vapor may condense on surfaces to interfere the freezing of water droplets in an indirect manner. 29 , 30 Therefore, to find out the origin of anti-icing performance variation among superhydrophobic surfaces, the roles of surface chemistry and surface topography in anti-icing under different conditions should be traced and explicated, which though is absent to the best of our knowledge. Herein, five samples, i.e., smooth hydrophilic surface, smooth hydrophobic surface, gecko-like, petal-like, and lotus-like superhydrophobic surfaces, are prepared. And their anti-icing properties are compared in two test conditions, which are condensation and non-condensation conditions. Briefly, the respective INT of gecko-like, petal-like, and lotus-like superhydrophobic surfaces is −30.5, −30.5, and −30.4°C in non-condensation condition, which goes to −15.0, −12.8, and −16.1°C in condensation condition accordingly. The nearly same INT values in non-condensation condition deviate from each other in condensation condition, and the INT values in non-condensation condition are far superior to the counterparts in condensation condition. The reason is that the freezing mechanism of water droplets experiences a change when condensation occurs, and thus the roles of surface chemistry and surface topography vary, resulting in such distinct anti-icing performance.",
"discussion": "Discussion Herein, a huge INT gap up to 18°C was discerned among different kinds of superhydrophobic surfaces using different test environments. In non-condensation condition, the freezing of water droplets on surfaces is caused by heterogeneous nucleation at droplet-surface interface, which instead strongly correlates to condensation frosting in condensation condition. Owning to such distinct freezing mechanisms, the roles of surface chemistry and surface topography of superhydrophobic surfaces in anti-icing vary significantly. In non-condensation condition, both surface chemistry and surface topography contribute to anti-icing. The former raises the energy barrier for ice nucleation by employing low-surface-energy materials, and the latter lowers the nucleation kinetics by reducing the contact between droplets and surfaces. In condensation condition, the role of surface chemistry remains positive, which defers the frost propagation by forming dropwise condensation. But the role of surface topography becomes complex: on one hand, it speeds up the frost propagation due to the increasing condensate density; on the other hand, it slows down the frost propagation via the fast coalescence of condensates that is evident in lotus-like surface. As a result, low surface energy benefits the anti-icing in both non-condensation and condensation conditions. Nano-sized textures are promising in non-condensation condition, whereas frost-free/frost-delay textures enabling fast coalescence, jumping removal, 29 , 45 or large vapor pressure gradients 66 have great potential in condensation condition. Limitations of the study We have revealed two different freezing mechanisms of water droplets on superhydrophobic surfaces and thus the role variation of surface chemistry and surface topography in anti-icing. However, we set several preconditions: there was no heat transfer between surfaces and water droplets in non-condensation environment; there was no jumping removal of condensates in condensation environment; and all surfaces were placed horizontally. Therefore, future investigation is needed to explore more situations."
} | 1,691 |
40227506 | PMC11996751 | pmc | 5 | {
"abstract": "As an emerging memory device, memristor shows great potential in neuromorphic computing applications due to its advantage of low power consumption. This review paper focuses on the application of low-power-based memristors in various aspects. The concept and structure of memristor devices are introduced. The selection of functional materials for low-power memristors is discussed, including ion transport materials, phase change materials, magnetoresistive materials, and ferroelectric materials. Two common types of memristor arrays, 1T1R and 1S1R crossbar arrays are introduced, and physical diagrams of edge computing memristor chips are discussed in detail. Potential applications of low-power memristors in advanced multi-value storage, digital logic gates, and analogue neuromorphic computing are summarized. Furthermore, the future challenges and outlook of neuromorphic computing based on memristor are deeply discussed.",
"conclusion": "Conclusion and Perspectives Overall, memristors represent progress in neuromorphic computing architectures, bringing significant advantages with their inherent physical properties and operational characteristics. First, non-volatile resistance state allows them to store information without additional data transmission power consumption. Second, many memristors achieve stable switching characteristics at feature sizes below 10 nm, with great potential for expansion. In-memory computing eliminates the traditional von Neumann bottleneck and greatly reduces the energy consumption associated with data movement between independent processing and storage units. The adjustable multi-level storage state enables matrix multiplication and weight updates for neuromorphic computing. With excellent CMOS compatibility, memristors facilitate integration into existing semiconductor manufacturing workflows, while supporting new computing paradigms such as logic-in-memory and brain-inspired neuromorphic computing. Recent demonstrations of memristor-based neural networks have achieved remarkable energy efficiencies below 1 fJ per synaptic operation, which is orders of magnitude better than conventional digital implementations and biological computing. The development of new materials remains key to improving the performance of memristors. Researchers are exploring 2D materials such as graphene and transition metal dichalcogenides, which have unique electrical properties and atomic-level thickness. These materials can achieve more precise resistance modulation and lower power consumption. Research on metal oxides continues, focusing on designing defect states and interface properties to achieve better switching characteristics and reliability. Array structure optimization is to minimize sneak current and improve read/write margins. Advanced selectors can be developed, including volatile switch selectors and engineered tunnel barriers. At the same time, three-dimensional integration strategies are explored to increase storage density while maintaining low power consumption levels. For storage applications, researchers are developing more complex programming schemes and error correction methods. Research on new switching mechanisms such as phase change and magnetoresistance effects may produce hybrid devices that combine the advantages of different storage mechanisms. For the digital logic computing, future research focuses on optimizing device characteristics for logic operations, developing more efficient programming schemes, and creating new circuit topologies that exploit the unique properties of memristors. In neuromorphic computing, future research will focus on developing ultra-low-power devices and systems, achieving extremely low programming currents to achieve ultra-low-power-consumption pulse generation and transmission. In terms of training schemes, future development schemes need to take into account the non-ideality of the device and optimize the power-performance balance through approximate computing techniques and multi-device architectures. The development of multi-functional memristor is also advancing, which can perform synaptic and neural functions at the same time, thereby achieving more compact and efficient neuromorphic systems. In response to the challenges of neuromorphic applications, researchers are improving energy efficiency through innovative programming schemes and adaptive precision techniques. Future work will also implement online learning algorithms under low-power operation and explore the use of complementary memristor devices to simplify the weight update process. In addition, the integration of memristor neuromorphic systems with CMOS circuits is also being optimized, especially interface circuits operating at low voltages.",
"introduction": "Introduction Von Neumann architecture is the basic architecture of modern computers, proposed by mathematician John von Neumann in 1945. Its core idea is to store program instructions and data in the same memory block and process the data by reading and executing these instructions through a central processing unit (CPU). This architecture's primary benefit lies in its adaptability and malleability, allowing the computer to undertake various tasks by altering programs stored in its memory [ 1 ]. However, von Neumann structure has its inherent flaws, where data storage and computing share the same channel. Such working mode limits processing speed of computer, especially if it uses dynamic random access memory (DRAM) as its primary memory. DRAM access not only requires high energy consumption, but also requires periodic refreshing. During data processing, the processor has to run continuously even while waiting for data, leading to additional energy consumption. As a result, the so-called “energy wall” and “speed wall” are formed. As internet technology rapidly evolves, the demand for artificial intelligence is experiencing exponential growth. Artificial intelligence has achieved numerous breakthroughs in various domains, including image processing, natural language processing, and big data analysis [ 2 – 4 ]. The amount of data that need to be trained and processed are also increasing daily. To address this problem, complex hardware systems consisting of numerous CPUs and graphics processing units (GPUs) have been developed. As semiconductor technology is approaching its physical limits, Moore's law is also facing failure [ 5 , 6 ], and researchers must examine the constraints of von Neumann architecture through the lens of computer architecture and software algorithms. In this regard, researchers have proposed various approaches, such as the introduction of multi-level caches [ 7 ], the introduction of data streaming [ 8 ], and the proposal of in-memory computing. Among emerging technologies, in-memory computing, first conceptualized by W.H. Kautz in 1969 [ 9 ], seamlessly integrates computational functions within storage, drastically reducing the delay for data transfer. This integration further leads to reduced power consumption and improved efficiency and is hailed as the next-generation computer architecture poised to transcend the barriers of von Neumann architecture. In recent years, there has been a swift advancement in the development of novel non-volatile memory and in-memory computing technology. With high speed, low power consumption and high-density integration capability, memristor is becoming a research hotspot in in-memory computing fields. Inspired by human brain, memristors with weights updating functions are considered ideal for developing in-memory computing and artificial intelligence [ 10 ]. This paper summarizes the research progress of memristors in the field of in-memory computing and artificial intelligence from the perspective of power consumption, covering the aspects of the device structure, mechanism, and key performance parameters of memristors, as well as the introduction of memristor arrays. Then, the low-power functional materials applied in memristors are categorized and discussed. Afterward, the review focuses on discussion of reducing power consumption in several compelling application areas of memristors, especially in multi-bit memories, logic gates, and neuromorphic computing. By summarizing the principles of memristors applied therein, the low-power implementation mechanism is well analyzed. Furthermore, the existing research progress, future challenges and outlook are discussed in detail. Figure 1 shows the overview of this review article. Figure 2 a shows the von Neumann architecture diagrams mentioned above. Figure 2 b shows a schematic diagram of the “energy wall” and the “speed wall”. Fig. 1 Overview of memristors for low-power storage and computing: including devices, materials, artificial synapses and neurons, and neural networks. From the device level, resistive random access memory (RRAM), phase change random access memory (PCRAM), magnetoresistive random access memory (MRAM) and ferroelectric device are potential low-power neuromorphic computing electronics. From materials system level, ion transport materials, phase change materials, magnetoresistive materials and ferroelectric materials are main functional material layers for low-power memristors. These novel memristors could be used to act as artificial synapses and neurons for low-power neuromorphic computing, including artificial neural network (ANN), spiking neural network (SNN) and convolutional neural network (CNN) Fig. 2 a Schematic illustration of the segregation structure. b Schematic representation of the “energy wall” and “speed wall” facing the von Neumann structure. c Schematic diagram of RRAM device structure. d Schematic diagram of PCRAM device structure. e Schematic diagram of MRAM device structure. f Schematic diagram of ferroelectric device structure Memristor The concept of memristor was first proposed by Professor Chua in 1971, which was the fourth basic passive circuit element after resistance, capacitance, and inductance, filling the gap in the description of the relationship between electric charge and magnetic flux [ 11 ]. Its mathematical model is expressed as the ratio of magnetic flux to electric charge, \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$M = {\\text{d}}\\varphi /{\\text{d}}q$$\\end{document} M = d φ / d q , the resistance is determined by the magnetic flux. It is a nonlinear resistance element with memory characteristics. However, in actual physical systems, direct coupling of magnetic flux and charge is not easy to achieve, and ideal memristors remain more at the theoretical level. Although many devices do not strictly meet the definition of ideal memristors, they exhibit similar characteristics, especially the non-volatile characteristic and adjustability. The realization of generalized memristors is usually based on ion migration, the formation and breaking of conductive filaments (CFs), phase change or magnetic spin effects, etc. RRAM, PCRAM, MRAM and ferroelectric memristor have emerged. RRAM is one of the typical representatives of memristors. Its resistance state is determined by the distribution of oxygen vacancies or CFs inside the material. The resistance can be changed by voltage pulses and be retained after removing pulses. The structure is usually divided into electrodes and functional layers, presenting a sandwich structure of electrode-functional layer-electrode, as shown in Fig. 2 c. PCRAM uses phase change materials between crystalline and amorphous states to achieve resistance change. The material can be heated to different states under different current pulses, with low resistance in the crystalline state and high resistance in the amorphous state, thereby achieving data writing and storage. The PCRAM device structure is generally mushroom-shaped, with a wider top electrode, a narrower bottom electrode, and a layer of phase change material in the middle. The device structure is shown in Fig. 2 d. MRAM uses the non-volatile magnetic materials and spin electronics for storage. It stores data through a magnetic tunnel junction (MTJ), which consists of two layers of magnetic material and an insulating layer. One magnetic layer is fixed, and the magnetization direction of the other free layer can be changed by current. The resistance state of the MTJ represents the data, with low resistance corresponding to parallel magnetization and high resistance corresponding to antiparallel magnetization. The device structure is shown in Fig. 2 e. Different from early MRAM relying on magnetic field induced switching, spin-transfer torque (STT) technology directly changes the magnetization direction of the free layer through current, reducing power consumption and suitable for high-density storage. Spin-transfer torque random access memory (STT-RAM) is developed based on STT technology. Similarly, there is spin–orbit torque random access memory (SOT-RAM) that uses the spin–orbit torque (SOT) effect. Ferroelectric memristor uses the polarization characteristics of ferroelectric materials to regulate the resistance state of the device. Ferroelectric materials have reversible polarization direction. When an external electric field is applied, the polarization direction of ferroelectric materials can be flipped, thereby changing the barrier height or interface charge distribution. This change affects the tunneling behavior of the current and the conductivity characteristics and ultimately manifests as different resistance states, as shown in Fig. 2 f. Functional Materials According to the common memristor types, memristor functional layer materials can be divided into ion transport materials, phase change materials, magnetoresistive materials and ferroelectric materials, as shown in Fig. 3 . Ion transport materials are mainly targeted at RRAM. In recent years, research in this area has mainly focused on inorganic and organic materials, specifically oxides, perovskites, two-dimensional (2D) materials and organic materials. Inorganic oxides have excellent performance and mature preparation technology and are currently widely used, but traditional binary oxides still have problems such as large leakage current and large power consumption. By doping or constructing multi-layer oxide heterojunctions, the formation and dissolution of conductive filaments can be improved for low-power-consumption storage. Perovskites and two-dimensional materials have unique structures, so they have excellent ionic conductivity and low-voltage operation [ 12 , 13 , 14 , 15 ]. Organic materials are regarded as strong competitors for the next generation of memory due to their flexibility, adjustability and low-cost potential, especially in flexible devices [ 16 ]. Typical research performance reports are summarized in Table 1 . Phase change materials are mainly chalcogenide alloys, with Ge–Sb–Te (GST) as the core. Recent PCRAM devices are also based on GST for heterogeneous doping and proportion alloying. When evaluating the impact of phase change materials on the performance of PCRAM devices, crystallization temperature, thermal conductivity, etc. are key indicators [ 17 ]. Khan et al. introduced GeTe/Sb 2 Te 3 superlattice structure in PCRAM, reducing heat loss and power consumption by 25–30 times [ 18 ]. Yang et al. introduced a conductive bridge phase change mechanism into a heterogeneous Ge-Sb-O alloy, which achieved fJ-level energy consumption (43 fJ) [ 19 ]. These works provide evidence for low-power-consumption applications of PCRAM. Magnetoresistive materials with spin polarization characteristics are mainly used for MRAM, where the free layers and fixed layers are made of ferromagnetic materials. As a king of typical ferromagnetic material, CoFeB can form a good interface with the insulating layer and has a low magnetization reversal energy. MgO usually acts as an insulator in the magnetic tunnel junction and can achieve a high tunnel magnetoresistance ratio. Most applications require MTJ to have perpendicular magnetic anisotropy (PMA), that is, the magnetization direction of the material is more likely to be arranged in a direction perpendicular to the plane of the film. PMA is related to the interface effect, lattice structure and stress of the material. The general methods to improve PMA include stacking materials with strong spin–orbit coupling such as ruthenium, cobalt or platinum in the buffer layer, or using an external voltage to regulate the magnetic anisotropy of the magnetic material. STT-RAM has been partially commercialized, but due to high current requirements and material degradation, researchers introduce SOT-RAM to reduce power consumption and increase write speed through the spin–orbit torque effect. The most studied SOT materials are heavy metal materials and topological insulators with strong spin Hall effect or Rashba effect [ 20 ]. Heavy metal materials such as Ta, W and Pt are used for the SOT layer, which have a high spin Hall angle and can efficiently generate spin currents. The surface states of topological insulators (such as Bi 2 Se 3 , Bi 2 Te 3 ) have high spin polarization rates and can achieve efficient spin injection at low currents.\n Table 1 Summary of the characteristics of the four functional materials of RRAM related to device research Structure Thickness Operating voltage Programming power consumption Endurance Year of publication Inorganic oxides and heterojunctions ITO/Bi:SnO₂/TiN [ 21 ] 20 nm − 0.5 V/0.4 V The SET operating power is 16 µW 10⁷ 2020 Ag/SiO₂/Ta₂O₅/Pt [ 22 ] 6.5 nm 0.14 V to 0.24 V/− 0.06 V to − 0.14 V N/A > 1000 2020 Pd/BaTiO 3 :Nd 2 O 3 /La 0.67 Sr 0.33 MnO 3 (LSMO)/STO [ 23 ] BNO: 34 nm LSMO:12 nm − 1 V/2 V 0.45 fJ per synaptic event > 10 10 2024 Two-dimensional materials Au/h-BN/Ti [ 24 ] 5 nm − 0.5 V/0.5 V 1.2 pJ/pulse, 30 ns pulse width and 45 µA current > 6000 2023 Ti /h-BN/Au [ 25 ] ~ 2.3 nm 2.75 V < 2 pJ 600 2024 Pt/WSe 2 /Hf x Zr 1−x O 2 (HZO)/TiN [ 26 ] WSe 2 : ~ 0.7 nm HZO:10 nm − 1.2 V / 1.5 V N/A > 2000 2025 Au/CuInS2/Cu [ 27 ] N/A 0.6 V 10 nW 1000 2025 Perovskite materials Ag/CH 3 NH 3 PbI 3 /FTO [ 28 ] 350 nm − 0.2 V/0.2 V ~ 47 fJ μm −2 > 10 3 2020 Ag/BA 2 MA 5 Pb 6 I 19 /Pt [ 29 ] ~ 300 nm − 0.15 V/0.15 V ~ 150 μW, I cc = 1 mA > 5 × 10 6 2024 Organic Materials Al/Cu-doped pMSSQ/Al [ 30 ] ~ 80 nm < 0.9 V < 0.5 pJ per pulse 500 2017 Ag/PFC-73/ITO [ 31 ] 114 nm 0.86 V N/A 60 2023 ITO/PEDOT:PSS/D:A/PDINN/Ag [ 32 ] The light intensity used (ranging from 0.51 to 194.01 mW cm −2 ) 2023 Fig. 3 Schematic diagram of memristor classification of different functional materials, including ion transport, phase change, magnetoresistive and ferroelectric. Among them, ion transport materials include organic and inorganic types [ 33 ]. Copyright (2014) American Chemical Society [ 34 ]. Copyright (2019) Wiley‐VCH, phase change materials are mainly chalcogenide alloys [ 35 ]. Copyright (2022) The Authors [ 36 ]. Copyright (2020) The Authors, magnetoresistive materials mainly constitute MTJ [ 37 ]. Copyright (2023) Science China Press [ 38 ]. Copyright (2024) The Authors, and ferroelectric materials mainly have spontaneous polarization characteristics [ 39 ]. Copyright (2020) The Authors [ 40 ]. Copyright (2024) Wiley‐VCH Ferroelectric materials can achieve reversible polarization reversal under electric field, thereby regulating tunneling current or interface charge distribution and realizing resistance state storage. Classical ferroelectric materials include bismuth titanate (BTO) and barium strontium titanate (BST), which are widely used in ferroelectric tunneling junctions due to their high remanent polarization and low leakage current. Because of excellent complementary metal–oxide–semiconductor (CMOS) compatibility, hafnium oxide-based materials (such as doped HfO 2 ) have become a research hotspot in recent years, especially in low-power and high-density memories. Two-dimensional ferroelectrics is a kind of emerging ferroelectric materials, such as In 2 Se 3 and MoTe 2 , which have ultra-thin thicknesses and are suitable for high-density integration and flexible electronics. Figure 4 summarizes the power consumption of various memristors when completing synaptic operations. RRAM and ferroelectric memristors can reach a lower level than biological levels of 10 fJ. The reported lowest power consumption is 4.28 aJ of HfAlOx-based RRAM, indicating that RRAM exhibits great potential in low-power neuromorphic computing. Therefore, the following content will be expanded on low-power-consumption RRAM. Fig. 4 Power consumption of different low-power memristors when performing synaptic plasticity [ 40 – 58 ], where biological synaptic power consumption is ~ 10 fJ. The reported power consumption of novel memristors range from 5 nJ to 4.28 aJ, exhibiting great potential in neuromorphic computing Memristor Array Two typical structures of memristor array are the 1 transistor 1 resistor (1T1R) array and the crossbar array. As illustrated in Fig. 5 a, 1T1R arrays are active arrays where each memristor is connected in series with a transistor. The word lines connect to the gate electrode of transistor, and the source lines connect to the source of the transistor. The bit lines connect to the top electrode of the memristor, and the bottom electrode connects to the drain of the transistor. The cell area of a 1T1R array is typically 12F 2 (F is the minimum feature size). As illustrated in Fig. 5 b, crossbar arrays are passive arrays with 4F 2 , consisting of perpendicular word lines and bit lines that form a crossbar structure. Memristors are arranged at the cross-points, which is more suitable for integration than a 1T1R and has no quiescent power dissipation. However, crossbar structure is prone to latent path currents. The latent path currents will flow through the other path resistors, thus causing inaccurate readings in the calculations, as well as additional power losses. In contrast, the 1T1R array, with its larger cell area and better isolation of neighboring cells, has no risk of sneak currents, which has higher computational read accuracy. For the crossbar array, a common approach to solving this problem is increasing the I–V nonlinearity by connecting a selector in series with one end of each memristor cell. The selector can use either a diode a resistor (1D1R) for unipolar memristors or a two-terminal selector device for bipolar memristors (1S1R). The combined device effectively suppresses the leakage currents caused by the unipolar memristor’s reverse bias or bipolar memristor’s low bias, resulting in much lower currents [ 59 – 61 ]. In recent years, prototype chips based on memristor arrays have been widely developed. Figure 5 c–h shows recent studies of memristors arrays, which summarize the structures, the types, the sizes and the realized functions. Fig. 5 Physical diagram based on 1T1R and crossbar memristor arrays. a Schematic diagram of a basic 1T1R array [ 62 ]. Copyright (2023) The Authors. b Schematic diagram of a basic crossbar array [ 63 ]. Copyright (2019) The Authors. c 128 × 64 1T1R array for handwritten digit classification [ 64 ]. Copyright (2018) The Authors. d 32 × 32 1T1R reconfigurable memristor array for analog computing tasks [ 65 ]. Copyright (2022) The Authors. e 2K memristor chips and an FPGA board, which mainly uses memristor arrays to achieve high-precision medical image reconstruction [ 62 ]. Copyright (2023) The Authors. f Schematic diagram of 32 × 32 WO x memristor array realize temporal information processing and handwritten digit recognition [ 66 ]. Copyright (2017) The Authors. g SEM image of a 20 × 20 crossbar array, used for neuromorphic computing with each memristor acting as a synapse [ 67 ]. Copyright (2018) The Authors. h 12 × 12 crossbar memory array composed of self-selective van der Waals heterostructure memory cells [ 63 ]. Copyright (2019) The Authors"
} | 6,040 |
29430725 | null | s2 | 6 | {
"abstract": "Vast potential exists for the development of novel, engineered platforms that manipulate biology for the production of programmed advanced materials. Such systems would possess the autonomous, adaptive, and self-healing characteristics of living organisms, but would be engineered with the goal of assembling bulk materials with designer physicochemical or mechanical properties, across multiple length scales. Early efforts toward such engineered living materials (ELMs) are reviewed here, with an emphasis on engineered bacterial systems, living composite materials which integrate inorganic components, successful examples of large-scale implementation, and production methods. In addition, a conceptual exploration of the fundamental criteria of ELM technology and its future challenges is presented. Cradled within the rich intersection of synthetic biology and self-assembling materials, the development of ELM technologies allows the power of biology to be leveraged to grow complex structures and objects using a palette of bio-nanomaterials."
} | 262 |
24119078 | PMC3815454 | pmc | 7 | {
"abstract": "Spider dragline silk is considered to be the toughest biopolymer on Earth due to an extraordinary combination of strength and elasticity. Moreover, silks are biocompatible and biodegradable protein-based materials. Recent advances in genetic engineering make it possible to produce recombinant silks in heterologous hosts, opening up opportunities for large-scale production of recombinant silks for various biomedical and material science applications. We review the current strategies to produce recombinant spider silks.",
"introduction": "Introduction Spider silks have been a focus of research for almost two decades due to their outstanding mechanical and biophysical properties. Spider silks are remarkable natural polymers that consist of three domains: a repetitive middle core domain that dominates the protein chain, and non-repetitive N-terminal and C-terminal domains. The large core domain is organized in a block copolymer-like arrangement, in which two basic sequences, crystalline [poly(A) or poly(GA)] and less crystalline (GGX or GPGXX) polypeptides alternate. At least seven different types of silk proteins are known for one orb-weaver species of spider (Lewis, 2006a ). Silks differ in primary sequence, physical properties and functions ( Hu et al ., 2006 ). For example, dragline silks used to build frames, radii and lifelines are known for outstanding mechanical properties including strength, toughness and elasticity ( Gosline et al ., 1984 ). On an equal weight basis, spider silk has a higher toughness than steel and Kevlar ( Vepari and Kaplan, 2007 ; Heim et al ., 2009 ). Flageliform silk found in capture spirals has extensibility of up to 500%. Minor ampullate silk, which is found in auxiliary spirals of the orb-web and in prey wrapping, possesses high toughness and strength almost similar to major ampullate silks, but does not supercontract in water. Figure 1 depicts the location and structural elements of MaSp, MiSp and Flag silks. Figure 1 A. An adult female orb weaver spider Nephila clavipes and her web. B. Schematic overview of N. clavipes web composed of three different spider silk proteins and their structures. The coloured boxes indicate the structural motifs in silk proteins. An empty box marked ‘?’ indicates that the secondary structure of the ‘spacer’ region is unknown. Note: MaSp1 or MaSp2: major ampullate spidroin 1 or 2; MiSp1 and 2: minor ampullate spidroin1 and 2; Flag: flagelliform protein. The photo was taken by Olena and Artem Tokarev in the Florida Keys. Finally, there are other silk types such as aciniform, pyriform, aggregate and tubuliform (egg case) with unusual primary structure, composition and properties. Diverse and unique biomechanical properties together with biocompatibility and a slow rate of degradation make spider silks excellent candidates as biomaterials for tissue engineering, guided tissue repair and drug delivery, for cosmetic products (e.g. nail and hair strengthener, skin care products), and industrial materials (e.g. nanowires, nanofibres, surface coatings). Recent advances in genetic engineering have provided a route to produce various types of recombinant spider silks ( Prince et al ., 1995 ; Fahnestock and Bedzyk, 1997 ; Rabotyagova et al ., 2009 ; Xu et al ., 2007 ). However, production of spider silk proteins at a larger scale remains challenging. Moreover, recombinant silk threads do not recapitulate the full potential of native fibres in terms of mechanical properties. Different heterologous host systems have been investigated to develop suitable production systems. In this review, we discuss recent advances in the production of recombinant spider silks in heterologous host systems with the main focus on microbial production. In particular, we focus on dragline silks. Current cloning strategies, expression systems and purification strategies will be discussed to help researchers to engineer customized synthetic spider silk-like proteins for various needs, including biomaterials and material science applications. Structure of silk proteins Spider silks are fascinating polymers, as is the spinning process that members of Araneidae family use to make these exceptional materials. Spiders use complex spinning to rapidly transform water soluble, high molecular weight, silk proteins into solid fibres at ambient temperature and pressure, giving rise to an environmentally safe, biodegradable and high performance material ( Asakura et al ., 2007 ; Lewicka et al ., 2012 ; Teulé et al ., 2012a ). The details on anatomy and physiology of the spider spinning apparatus ( N. clavipes ) can be found elsewhere (Knight and Vollrath, 2001 ; 2002 ; Eisoldt et al ., 2011 ; Rising et al ., 2011 ). In order to understand the challenges and needs associated with biotechnological production of recombinant spider silks, primary protein motifs, composition and secondary structural elements must be discussed. As mentioned earlier, one spider is capable of producing up to seven different types of silks with varying mechanical properties. In spite of different mechanical and physiological properties, the majority of spider silks share a common primary structural pattern comprised of a large central core of repetitive protein domains flanked by non-repetitive N- and C-terminal domains. The most investigated silk is dragline silk, which shows a remarkable combination of strength and elasticity. The golden orb-weaver spider, N. clavipes , produces dragline silk in the major ampullate gland ( Knight and Vollrath, 2001 ). Dragline silk is the protein complex composed of major ampullate dragline silk protein 1 (MaSp1) and major ampullate dragline silk protein 2 (MaSp2). Both silks are approximately 3500 amino acid long. MaSp1 can be found in the fibre core and the periphery, whereas MaSp2 forms clusters in certain core areas. The large central domains of MaSp1 and MaSp2 are organized in block copolymer-like arrangements, in which two basic sequences, crystalline [poly(A) or poly(GA)] and less crystalline (GGX or GPGXX) polypeptides alternate in core domain. The main difference between MaSp1 and MaSp2 is the presence of proline (P) residues accounting for 15% of the total amino acid content in MaSp2 ( Hu et al ., 2006 ), whereas MaSp1 is proline-free. By calculating the number of proline residues in N. clavipes dragline silk, it is possible to estimate the presence of the two proteins in fibres; 81% MaSp1 and 19% MaSp2 ( Brooks et al ., 2005 ). Different spiders have different ratios of MaSp1 and MaSp2. For example, a dragline silk fibre from the orb weaver Argiope aurantia contains 41% MaSp1 and 59% MaSp2 ( Huemmerich et al ., 2004 ). Such changes in the ratios of major ampullate silks can dictate the performance of the silk fibre ( Vollrath and Knight, 1999 ). Specific secondary structures have been assigned to poly(A)/(GA), GGX and GPGXX motifs including β-sheet, 3 10 -helix and β-spiral respectively ( Humenik et al ., 2011 ). The primary sequence, composition and secondary structural elements of the repetitive core domain are responsible for mechanical properties of spider silks; whereas, non-repetitive N- and C-terminal domains are essential for the storage of liquid silk dope in a lumen and fibre formation in a spinning duct ( Ittah et al ., 2006 ). The primary amino acid sequence, composition and secondary structural elements of other silk types are reviewed elsewhere (Lewis, 2006b ; Humenik et al ., 2011 ). Production of recombinant silk proteins Spiders cannot be farmed, in contrast to silkworms, due to their aggressive behaviour and territorial nature ( Kluge et al ., 2008 ). Collecting silk from webs is a time-consuming task. It took 8 years to make a golden spider silk cape from 1.2 million golden orb webs ( Chung et al ., 2012 ). Therefore, biotechnological production of recombinant spider silks is the only practicable solution to harvest silks on a larger scale and to meet growing needs of medicine and biotechnology. A variety of heterologous host systems have been explored to produce different types of recombinant silks ( Table 1 and Table 2 ). Recombinant partial spidroins as well as engineered silks have been cloned and expressed in bacteria ( Escherichia coli ), yeast ( Pichia pastoris ), insects (silkworm larvae), plants (tobacco, soybean, potato, Arabidopsis), mammalian cell lines (BHT/hamster) and transgenic animals (mice, goats). Table 1 Summary of recombinantly expressed spider silks in bacteria and yeasts. Spider silk origin, number of monomers, molecular weight, cloning and expression plasmids as well as restriction enzymes and purification strategies used to produce recombinant silks are shown Type Host Origin Protein Number of monomers MW (KDa) Cloning plasmid RE Expression plasmid Purification Strategy Yield (mg L −1 ) References Bacteria E. coli Nephila clavipes 16, 32, 64, 96 55, 100, 193, 285 pET30a(+) Nhe I/ Spe I pET30a(+) Ammonium sulphate Xia et al ., 2010 6; 15 16, 39.5 pETNX PA 96.8 mg L −1 ; 200 mg L −1 Dams-Kozlowska et al ., 2012 MaSp 1 8; 16 65–163 pBR322 derived Pst I pFP202 (pET9a + pET11a) 300 mg L −1 Fahnestock and Irwin, 1997 8; 16 65–163 pFP202, pFP204, or pFP207 IMAC Fahnestock and Irwin, 1997 16, 24 46, 70 pBSSKII+ AvrII, Nhe1 pET19k N A −1 An et al ., 2011 poly(A) and GGX 1, 2, 3, 6 10, 18 pET30a(+) Spe I/Nde I pET30a(+) 25 mg ml −1 Rabotyagova et al ., 2009 E. coli Nephila clavipes 8; 16 65–163 pBR322 derived Pst I pFP202, pFP204, or pFP207 300 mg L −1 Fahnestock and Irwin, 1997 8, 16, 32 31, 58, 112 pBBSK Sca/Xma/BspEI pET19b 10 mg g −1 Lewis et al ., 1996 Argiope aurantia MaSp2 16 63 IMAC Brooks et al ., 2008 12 71 pET30a(+) N A −1 pET30a(+) N A −1 Brooks et al ., 2008 8 67 Brooks et al ., 2008 E. coli Nephila clavipes Masp1/Masp2 24/16 62/47 pBSSKII+ Xma1/Sca1/BspE1 pET19K IMAC 120 mg L −1 An et al ., 2011 1x-18x 15, 23, 32, 41 pUC18 Spe I/Nde I pQE-9 15, 7, 3, 2 mg L −1 Prince et al ., 1995 E. coli Argiope trifasciata AcSp1 2, 3, 4 19, 38, 51.7, 76.1 pET32 BamHI/ BsgI/ BseRI pET32 IMAC 80 mg L −1 ; 22 mg L −1 Nephila antipodiana TuSp1 11 190 Xma1/Pvu1/BspE1 40 mg L −1 Xu et al ., 2007 Salmonella Araneus diadematus ADF1 1x-3x 30–56 pJ2 HindIII/XbaI pTRC99a_Cm SEC N A −1 Widmaier et al ., 2009 ADF2 1x-3x Widmaier et al ., 2009 ADF3 1x-3x Widmaier et al ., 2009 Yeast Pichia Pastori Nephila clavipes Masp 1 8, 16 65 pBR322 derived Pst I pFP684 Ammonium sulphate 663 mg L −1 Fahnestock and Bedzyk, 1997 Table 2 Summary of recombinantly expressed spider silks in insects, plants and mammalians. Spider silk origin, number of monomers, molecular weight, cloning and expression plasmids as well as restriction enzymes and purification strategies used to produce recombinant silks are shown Type Host Origin Protein Number of monomers MW (KDa) Cloning plasmid RE Expression plasmid Purification strategy Yield (mg L −1 ) References Insects B. mori Nephila clavipes Masp1 2 83 pSLfa1180fa Spe1/Nde1 pBac[3xP3-DsRedaf] IMAC N A −1 Wen et al ., 2010 4 70 pSL1180 pFastBacHT-C 6 mg/larva Zhang et al ., 2008 Masp1+Flag multiple 75–130 pBSSKII+ and pSLfa1180fa Sca1/Xma1/BspE1 pBAC[3xP3-DsRedaf] N A −1 Teulé et al ., 2012b Plants Nicotiana tobaccum Solaum tubercum Masp1 multiple 12.9–99.8 pUC19 NgoMIV/HindIII pRTRA7/3 (NH 4 ) 2 SO 4 10–50% saturation 0.5-2 % total protein Scheller et al ., 2001 Nicotiana benthamiana Nephila clavipes Flag (intein) 4; 10 47, 72, 100, 250 pRTRA15 splicing events pCB301-Kan IMAC 1.8 mg/50 g leaft material 0.34%; 0.03% in leaves, 1.2%; 0.78% in seeds Hauptmann et al ., 2013 Arabdopsis thaliana Glycine max Masp1 8, 16 64, 127 pBSSK+ BglII/BamH1 Cong' + Pha3' ammonium sulphate 1% in somatic embryos Barr et al ., 2004 Mammalians Trangenic mice Nephila clavipes MaSP1 6 31–66 pGEM-5zf BamHI/NcoI pBC1 centrifugation 11.7 mg L −1 Xu et al ., 2007 COS-1 cells Euprosthenops sp. 25, 22 pER1-14 BamH1/EcoRV pSecTag2/Hygro A N A −1 N A −1 Grip et al ., 2006 Baby hamster kidney Nephila clavipes Masp1/ Masp2 N A −1 59, 106/ 59 pBSSK+ ApaI/SapI CMV promoter ammonium sulfate Baby hamster kidney Araneus diadematus ADF3 63, 60, 110, 140 pSecTag-C MscI/PvuII 25–50 mg L −1 Lazaris et al ., 2002 Unicellular organisms as heterologous host systems Unicellular organisms, such as bacteria and yeast, have been investigated as host systems for recombinant silks. A gram-negative, rod-shaped bacterium E. coli is a well-established host for industrial scale production of proteins. Therefore, the majority of recombinant spider silks have been produced in E. coli (Lewis et al ., 2011 ; Fahnestock and Irwin, 1997 ; Wang et al ., 2006 ; Rabotyagova et al ., 2009 ; Rabotyagova et al ., 2010 ; An et al ., 2011 ; An et al ., 2012 ; Teulé et al ., 2012a ). E. coli is easy to manipulate, has a short generation time, is relatively low cost and can be scaled up for larger amounts protein production. The recombinant DNA approach enables the production of recombinant spider silks with programmed sequences, secondary structures, architectures and precise molecular weight ( Rabotyagova et al ., 2011 ). There are four main steps in the process: (i) design and assembly of synthetic silk-like genes into genetic ‘cassettes’, (ii) insertion of this segment into a DNA vector, (iii) transformation of this recombinant DNA molecule into a host cell and (iv) expression and purification of the selected clones. Figure 2 summarizes the recombinant DNA approach used to prepare silk-like proteins. Figure 2 Recombinant DNA approach used to prepare silk-like proteins. The monomeric silk-like gene sequences can be synthesized as short single-stranded oligonucleotides (up to 100 bp) by commercial oligonucleotide synthesis or used directly as polymerase chain reaction products from cDNA libraries. Large repetitive sequences can be constructed by using concatemerization, step-by-step directional approach and recursive ligation ( Fig. 3 ). Concatemerization is a useful method when a library of genes of different sizes is desired but has limitations in the preparation of genes with specific sizes ( Meyer and Chilkoti, 2002 ). To overcome limitations of concatemerization, recursive directional ligation or a step-by-step ligation is employed ( Meyer and Chilkoti, 2002 ; Wright and Conticello, 2002 ). Recursive directional ligation allows for facile modularity, where control over the size of the genetic cassettes is achieved. Moreover, recursive directional ligation eliminates the restriction sites at the junctions between monomeric genetic cassettes without interrupting key gene sequences with additional base pairs that makes it different from the step-by-step ligation approach ( Higashiya et al ., 2007 ). Figure 3 Gene multimerization approaches. Note: RE Site stands for a restriction enzyme site. For example, we have employed step-by-step directional ligation to produce various partial recombinant spider silks as well as engineered silk-like proteins based on the sequences of dragline silk originated from N. clavipes ( Prince et al ., 1995 ; Wang et al ., 2006 ; Huang et al ., 2007 ; Rabotyagova et al ., 2009 ; 2010 ; Mieszawska et al ., 2010 ; Gomes et al ., 2011 ; Numata et al ., 2012 ). As one example, spider silk block copolymers were generated in E. coli (Rabotyagova et al ., 2009 ; 2010 ). In the first cloning step, a commercially available pET30a(+) vector (Novagen, San Diego, CA, USA) was modified with an adaptor sequence, carrying NheI and SpeI restriction sites. The adaptor was inserted into XhoI and NcoI sites of a pET30a(+) to generate pET30L. The coding sequences of two spider silk-like monomers A (hydrophobic block) and B (hydrophilic) were designed to carry SpeI and NheI restriction sites at the ends of the sequences. This allowed ligation of the domains into a pET30L vector. By using a step-by-step directional ligation approach, direct control over the assembly of monomeric genes into complex sequences was achieved. Six different constructs were cloned and transformed into the bacterial host for expression. An N-terminal His-tag was used for protein purification by immobilized metal affinity chromatography ( Rabotyagova et al ., 2009 ). Another genetic engineered strategy has been proposed by Lewis Laboratory to assemble long repetitive spider silk genes ( Teule et al ., 2009 ). This cloning strategy employs a one-step head-to-tail ligation that can produce large inserts in precise manner ( Lewis et al ., 1996 ; Brooks et al ., 2008 ; Teule et al ., 2009 ; Teulé et al ., 2012a ). The spider silk synthetic genes were optimized for codon usage in E. coli and were cloned into a plasmid vector pBluescriptII SK(+) (Stratagene). Each silk module was carrying compatible XmaI and BspEI restriction sites at the ends on the coding sequences. The vector also contained a unique restriction site ( ScaI ) in the ampicillin resistance gene. By simultaneously performing two double digestion reactions ScaI – XmaI and ScaI – BspEI two fragments each containing a copy of a silk monomer gene were obtained. The fragments were ligated together using T4 ligase resulting in the doubling of the size of silk genes and restoring the ampicillin resistance of the plasmid ( Fig. 4 ). Several round of cloning were performed to obtain repetitive sequences of a desired size. Next, the multimeric synthetic genes were subcloned into an expression pET19b vector using NdeI and BamHI restriction sites. Since the expression vector was carrying NdeI and BamHI sites, the liberated inserts were cloned in-frame with pET19b. Similar to pET30L, silk genes in pET19b are under control of the T7 promoter and require the addition of isopropyl-β-D-1-thiogalactopyranoside to initiate protein expression. The expressed proteins can be purified by immobilized metal affinity chromatography (IMAC) due to the presence of an N-terminal His-tag. Several recombinant spider silk proteins from different species were produced using this genetic engineering strategy including silks from N. clavipes ( Teule et al ., 2009 ) Argiope aurantia ( Brooks et al ., 2008 ). Recombinant spider silk proteins from Nephylengys cruentata , Parawixia bistriata and Avicularia juruensis were produced employing this cloning strategy ( Leopoldo et al ., 2007 ) (US patent 20 100 311 645). Figure 4 summarizes the strategy. Figure 4 Cloning strategy used by the Lewis group to engineer long repetitive spider silk sequences (in green). A. Cloning of a silk monomer into the vector pBluescript II SK+. B. The resulting plasmid is double digested and fragments containing silk monomers are ligated again to produce longer sequences. C. The synthetic spider silk multimer is ligated into pET19b expression vector. Note: Restriction digestion sites are indicated by star. Adapted from reference (Teule et al ., 2009 ). A three module cloning strategy based on the sequences of ADF-3 and ADF-4 was developed by Scheibel research group ( Huemmerich et al ., 2004 ), designed so that multiple modules can be combined. Moreover, additional coding sequences such as N- or C-terminal domains can be added if needed. The purification protocol is based on heat resistance of silk proteins followed by an ammonium sulphate precipitation that is different from Ni-NTA IMAC. Different purification strategies have been employed recently to optimize small and large-scale production of recombinant silks. Most of the spider silk proteins are produced with an N- or C-terminal His-tags to make purification simple and produce enough amounts of the protein. However, the presence of this tag can affect protein secondary structure and interfere with the process of spider silk fibre formation. Dams-Kozlowska et al . (2012 ) proposed two strategies to purify spider silks from lysates without the use of a His-tag. These protocols are based on thermal treatment and organic acid resistance of silk proteins and do not require the presence of the His-tag. After purification, silk proteins based on MaSp1 gene sequence were formed into films that subsequently were used to grow murine fibroblast cell culture. The results demonstrated that silk films were non-toxic to the cells ( Dams-Kozlowska et al ., 2012 ). Because of the highly repetitive core sequence of spider silk genes, frequent homologous recombination, deletions, transcription errors, translation pauses, accumulation in inclusion bodies and low yields were observed during the production of recombinant silks in E. coli . Moreover, when the protein size was increased from 43 kDa to higher (the size of native spidroins is between 300 and 350 kDa), protein yields decreased dramatically. Codon optimization for the specific host expression system helped maximize the translation of the foreign gene transcripts and thus, improved protein yields ( Fahnestock and Bedzyk, 1997 , Lewis, 2006b ). It was also suggested that depletion of tRNA pools upon protein expression resulted in transcription and translation errors ( Rosenberg et al ., 1993 ). Recently, Xia et al . (2010 ) employed a metabolic engineered strategy to enhance the production of recombinant spider silks. The authors reported production of full length (284.9 kDa) recombinant N. clavipes dragline silk proteins that were rich in glycine (43–45%). Production of these silk proteins was enhanced by the use of the metabolically engineered expression host within which the glycyl-tRNA pool was elevated. The fibres spun with the native-sized recombinant spider silk protein showed tenacity, elongation and Young's modulus of 508 MPa, 15% and 21 GPa, respectively, comparable to those of native spider dragline silk ( Xia et al ., 2010 ). Through extensive proteomic analysis, serine hydroxymethyltransferase (GlyA) and β-subunit of glycly-tRNA synthetase (GlyS) were found to be upregulated to meet the high cellular demand for glycly-tRNA when expressing glycine-rich silk proteins. Increased glycine biosynthetic flux by overexpressing glycyl-tRNA synthetase elevated the total tRNAGly pool and resulted in enhanced production of high molecular weight recombinant spider silks. Recently, large spider recombinant egg case silk protein from Nephila antipodiana , 378 kDa, was engineered using E. coli , where gene multimers were chemically linked by cysteine disulfide bonds. The recombinant silk sequence consisted of two silk proteins: tubuliform spidroin 1 (TuSp1) and C-terminal domain of MisP1. Non-repetitive C-terminal domain of MiSp1 was chosen due to its higher water solubility and stability compared with the C-terminal domain of TuSp1. A disulfide linkage between two C-terminal domains was formed by introducing a point mutation (S76 to S76C). This link allowed the formation of a hybrid DNA construct that was expressed in E. coli (DE3). The recombinant protein was expressed in E. coli . Moreover, the artificial fibres spun from this protein showed higher tensile strength and Young' modulus than natural egg case protein ( Lin et al ., 2013 ). The highly repetitive silk gene arrangement and the unusual mRNA secondary structure result in inefficient translation that limits the size of the silks produced in E. coli . To minimize the presence of truncated silk proteins and allow the extracellular secretion of silks, the mythylotropic yeast P. pastoris has been used. Fahnestock and Bedzyk (1997 ) produced N. clavipes spider dragline silks in yeast P. pastoris . Synthetic genes were expressed at high levels under control of the methanol-inducible AOX1 promoter. Transformants containing multiple gene copies produced elevated levels of silk protein. Results demonstrated that P. pastoris can be used to successfully produced produce long repetitive proteins ( Fahnestock and Bedzyk, 1997 ). Spider silks from Araneus diadematus (ADF-1, 2 and 3) have also been expressed using the type III secretion system of a gram-negative, non-spore-forming, enterobacterium Salmonella . The authors reported yield values range from 90 to 410 nmol L −1 h −1 that is similar to 10 mg L −1 h −1 for a protein the size of ADF-2. The results demonstrated the feasibility to use Salmonella for the large-scale spider silk production ( Widmaier et al ., 2009 ). Mammalian cell lines, such as bovine mammary epithelial alveolar and baby hamster kidney cells, were used to express MaSp1 and MaSp2 ( Lazaris et al ., 2002 ). The cells expressed recombinant proteins; however, as size of silk gene increased, the yield decreased dramatically due to inability of mammalian cells to cope with large repetitive sequences. Several factors have attributed to the decreased yields including, but not limited to, inefficient transcription, insufficient secretion, low copy numbers and translational limitations. The produced silk proteins were spun into fibres, and their mechanical properties were tested. It was noted that those recombinant silks that were produced without a His-tag demonstrated better mechanical properties compared with fibres made of silk proteins with a His-tag (i.e. fibres were brittle). Similar problems (i.e. transcription and translation limitations) have been reported when green monkey kidney fibroblast-like cell lines (COS-1) were used to express a 636-base pair gene fragment of MaSp1 from the African spider Euprosthenops sp . ( Grip et al ., 2006 ). Table 1 summarizes genetic engineering approaches, cloning strategies, and production yields of recombinant silk proteins produced in unicellular heterologous host systems. Multicellular organisms as heterologous host systems Due to the low production rate and instability (i.e. frequent homologous recombination, deletions, transcription errors, translation pauses) of spider silk repetitive genes in unicellular organisms, multicellular organisms such as insects, plants and mammals have been studied for production of recombinant spider silk proteins. Silkworms ( B. mori ) can be farmed and produce cocoons containing large quantities of silkworm silk known as fibroin ( Vepari and Kaplan, 2007 ; Hu and Kaplan, 2011 ). Moreover, to produce a solid thread, silkworms employ a spinning process that is similar to that used by spiders to make dragline silk. The presence of a natural silk production system in silkworms makes them excellent candidates to investigate as heterologous hosts for spider silk production. There have been several reports of the transfer of silk genes from spiders to silkworms ( Motohashi et al ., 2005 ; Zhang et al ., 2011 ; Teulé et al ., 2012b ). Baculovirus-based expression systems have been used to introduce silk genes into a heterologous host. Baculovirus infects silkworms and allows for production of large quantities of heterologous proteins in a short period of time ( Motohashi et al ., 2005 ). Using this expression system, MaSp1 from N. clavipies linked with an enhanced green fluorescent protein (EGFP) fusion protein was cloned and expressed in the B. mori cell line (BmN) and larvae ( Zhang et al ., 2008 ). The authors reported successful production of a recombinant EGFP-MaSp1 fusion protein in both systems. In the silkworm larvae, a total of 6 mg of fusion protein was expressed, whereas in the BmN cells, 5% of the cell total protein was occupied by this recombinant silk. The major limitations of this expression system were low solubility of silk proteins and inability to assemble spider silk fibres. It was shown that more than 60% of the fusion proteins formed aggregates via self-assembly. To overcome solubility issues, MaSp1 C-terminal domain is to be incorporated due to its role to prevent aggregate formation. To produce fibres, germline-transgenic silkworms ( B. mori ) were produced by injecting silkworm eggs with a piggyBac transformation vector carrying MaSp1 sequence ( Wen et al ., 2010 ). The insects were capable of spinning fibres and forming cocoons containing recombinant spider silk. However, the mechanical properties of the fibres were lower than dragline MaSp1 silk due to the low ratio of MaSp1 in the total silk protein. In a recent effort to develop tough fibres, transgenic silkworms encoding chimeric silkworm/spider silk proteins were produced using piggyBac vectors (Teulé et al ., 2012b ). The vector, used previously by the Tamada group ( Kojima et al ., 2007 ) included the B. mori fibroin heavy chain promoter and enhancer, a genetic sequencing encoding a 78 kDa synthetic spider silk protein, and an EGFP tag. Strong EGFP signals were observed by fluorescence ( Fig. 5 ). The composite fibres were tougher than the parental silkworm silk fibres and as tough as native dragline spider silk fibres. Figure 5 Expression of the chimeric silkworm/spider silk/EGFP protein in (A) cocoons, (B and C) silk glands and (D) silk fibres from spider 6-GFP silkworms. Reproduced with permission from (Teulé et al ., 2012b ). These results demonstrate that silkworms can be engineered to generate composite silk fibres containing stably integrated spider silk protein sequences, which significantly improved overall mechanical properties. Transgenic plants have also been investigated as heterologous host systems to produce recombinant spider silks. Advances in genetic engineering technology and transformation methods make it possible to produce non-plant proteins in plants ( Yang et al ., 2005 ; Rech et al ., 2008 ). Moreover, one plant offers several different expression systems, such as seeds, leaves, tubers and roots with potential for organelle-specific accumulation of recombinant proteins ( Scheller and Conrad, 2005 ). Stable transgenic tobacco and potato lines were engineered to express MaSp1 genes from N. clavipes ranging from 420 to 3600 bp ( Scheller et al ., 2001 ). Recombinant spider silk proteins were found in the endoplasmic reticulum (ER) of tobacco and potato leaves at the accumulation of 2% of total soluble protein. Moreover, the production levels were independent of the size of silk genes. Purification was performed using high temperature treatment followed by acidification and ammonium sulphate precipitation. Additionally, recombinant MaSp1-like proteins were also produced in the leaves and seeds of Arabidopsis (small flowering plants related to cabbage) as well as in somatic soybean embryos ( Barr et al ., 2004 ). The expression of recombinant silks was driven by the 35S promoter in leaves and the β-conglycinin α' subunit promoter in seeds and somatic soybean embryos. The results demonstrated that recombinant spider silk proteins had higher accumulation levels in seeds than in the leaves. Recently, a native-sized FLAG protein from N. clavipes was cloned and expressed in the ER of tobacco plant ( Nicotiana benthamiana ) leaf cells using an intein-based posttranslational protein fusion technology ( Hauptmann et al ., 2013 ). This method avoids the need for highly repetitive transgenes resulting in a higher genetic and transcriptional stability. Additional details on production of fibrous proteins in plants can be found elsewhere ( Scheller and Conrad, 2005 ). Transgenic production of recombinant silk proteins in mammary glands and secretion of them into milk has been investigated in mice and goats ( Williams, 2003 ; Xu et al ., 2007 ). In case of transgenic mice production, MaSp1 and MaSp2 synthetic genes (40 and 55 kDa) were synthesized and cloned into the pBC1 expression vector (Invitrogen, Carlsbad, CA, USA) together with a goat β-casein signal sequence. The chimeric gene construct was microinjected into pronuclei of fertilized eggs of Kunming white mice ( Xu et al ., 2007 ). Southern blot analysis was used to identify mice containing transgene construct as well as a copy number of transgene. The expression of dragline silk in milk was confirmed by Northern blot followed by Western blot analysis. The results revealed that transgenic mice were capable of expressing recombinant silk proteins in their milk. Genetically engineered (transgenic) goats capable of expressing spider silk proteins based on the sequences of MaSp1 and MaSp 2 were produced by Nexia Biotechnologies, and later by the Lewis group ( Lazaris et al ., 2002 ; Service, 2002 ). Silk protein expression was controlled by the β-casein promoter and was expressed in the milk of transgenic goats. Silk proteins were observed only in mammary tissues as confirmed by Western blot ( Steinkraus et al ., 2012 ). Maximum yields observed for the recombinant silk production in transgenic animals were low (11.7 mg l −1 ) when compared with bacterial expression ( Table 1 and Table 2 ). Today, the large-scale production of recombinant silk proteins from transgenic animals is relatively expensive and challenging in terms of animal breeding."
} | 8,217 |
30944413 | null | s2 | 8 | {
"abstract": "Quorum sensing is a process of bacterial cell-to-cell chemical communication that relies on the production, detection and response to extracellular signalling molecules called autoinducers. Quorum sensing allows groups of bacteria to synchronously alter behaviour in response to changes in the population density and species composition of the vicinal community. Quorum-sensing-mediated communication is now understood to be the norm in the bacterial world. Elegant research has defined quorum-sensing components and their interactions, for the most part, under ideal and highly controlled conditions. Indeed, these seminal studies laid the foundations for the field. In this Review, we highlight new findings concerning how bacteria deploy quorum sensing in realistic scenarios that mimic nature. We focus on how quorums are detected and how quorum sensing controls group behaviours in complex and dynamically changing environments such as multi-species bacterial communities, in the presence of flow, in 3D non-uniform biofilms and in hosts during infection."
} | 265 |
36505976 | PMC9720699 | pmc | 9 | {
"abstract": "Abstract Spider silk is the toughest fiber found in nature, and bulk production of artificial spider silk that matches its mechanical properties remains elusive. Development of miniature spider silk proteins (mini‐spidroins) has made large‐scale fiber production economically feasible, but the fibers’ mechanical properties are inferior to native silk. The spider silk fiber's tensile strength is conferred by poly‐alanine stretches that are zipped together by tight side chain packing in β‐sheet crystals. Spidroins are secreted so they must be void of long stretches of hydrophobic residues, since such segments get inserted into the endoplasmic reticulum membrane. At the same time, hydrophobic residues have high β‐strand propensity and can mediate tight inter‐β‐sheet interactions, features that are attractive for generation of strong artificial silks. Protein production in prokaryotes can circumvent biological laws that spiders, being eukaryotic organisms, must obey, and the authors thus design mini‐spidroins that are predicted to more avidly form stronger β‐sheets than the wildtype protein. Biomimetic spinning of the engineered mini‐spidroins indeed results in fibers with increased tensile strength and two fiber types display toughness equal to native dragline silks. Bioreactor expression and purification result in a protein yield of ≈9 g L −1 which is in line with requirements for economically feasible bulk scale production.",
"conclusion": "3 Conclusion Using biological principles, we employed protein engineering to design mini‐spidroins with predicted increased β‐sheet propensities and increased inter‐β‐sheet binding strengths. Prokaryotic expression, protein purification, and biomimetic fiber spinning resulted in four different types of fibers with significantly improved tensile strength compared to the original mini‐spidroin. Using this strategy, we successfully produced the first biomimetic fibers with toughness values matching those of native dragline silk fibers. Finally, we show that these fibers can be produced at very high yields in bioreactors, vouching for feasible large‐scale production.",
"introduction": "1 Introduction Spider silk is nature's high‐performance fiber. Its unique combination of high tensile strength and extensibility results in an unsurpassed toughness which makes it very attractive for many industrial applications. [ \n \n 1 \n , \n 2 \n , \n 3 \n \n ] Due to limited availability of the natural material, large scale production must involve the expression of the silk proteins (spidroins) in heterologous hosts. [ \n \n 4 \n \n ] \n Spiders have up to seven different types of silk glands in which the spidroins are being produced, stored, and processed. [ \n \n 5 \n \n ] The major ampullate gland makes the strongest silk, which is used in the dragline and for making the framework of the web. [ \n \n 6 \n , \n 7 \n , \n 8 \n , \n 9 \n , \n 10 \n , \n 11 \n \n ] The spidroins are synthesized by epithelial cells lining the major ampullate gland and are stored in the gland lumen as a highly concentrated dope. [ \n \n 9 \n , \n 12 \n , \n 13 \n \n ] Changes in the microenvironment along the gland, [ \n \n 14 \n \n ] for example, ion exchange, drop in pH from 8.0 to at least 5.7, [ \n \n 15 \n \n ] increased shear forces, [ \n \n 16 \n \n ] and dehydration [ \n \n 7 \n \n ] lead to conformational transitions of the spidroins and fiber formation. [ \n \n 15 \n , \n 17 \n , \n 18 \n , \n 19 \n , \n 20 \n \n ] \n Spidroins are composed of an N‐terminal domain (NT), [ \n \n 21 \n \n ] a repetitive region that often is extensive [ \n \n 22 \n \n ] and a C‐terminal domain (CT). [ \n \n 18 \n \n ] The terminal domains are crucial for solubility of the spidroins during storage and regulate the assembly of the spidroins into a solid fiber. [ \n \n 17 \n , \n 18 \n , \n 19 \n , \n 20 \n , \n 23 \n \n ] The repetitive region of most major ampullate spidroins (MaSps) contain up to 100 tandem repeats of poly‐Ala blocks and Gly‐rich motifs. [ \n \n 22 \n , \n 24 \n \n ] In the soluble dope, the spidroins are mostly in random coil and helical conformations, [ \n \n 25 \n , \n 26 \n , \n 27 \n , \n 28 \n , \n 29 \n \n ] whereas the solid silk fiber contains nanosized crystals made up by stacked antiparallel β‐sheets embedded in amorphous structures. [ \n \n 30 \n , \n 31 \n , \n 32 \n , \n 33 \n , \n 34 \n \n ] This heterogeneous structure of the silk fiber is important as the β‐sheet crystals confer the strength while the amorphous structures confer the extensibility to the fiber. [ \n \n 10 \n , \n 35 \n , \n 36 \n \n ] The amorphous matrix, containing β‐turns and ordered structures with conformational similarities to collagen and poly‐proline helices, are dominated by the glycine‐rich regions. The β‐sheets, formed by the poly‐Ala blocks, orient with the β‐strands parallel to the fiber axis, [ \n \n 37 \n , \n 38 \n , \n 39 \n , \n 40 \n \n ] and the Ala side chain of a given β‐strand fill the space close to an α‐carbon in a neighboring β‐stand, analogous to a tightly packed steric zipper. [ \n \n 41 \n , \n 42 \n , \n 43 \n \n ] \n There are two main strategies for producing artificial silk fibers; one being expression of insoluble spidroins with subsequent solubilization and fiber processing using organic solvents, [ \n \n 44 \n , \n 45 \n , \n 46 \n , \n 47 \n , \n 48 \n , \n 49 \n \n ] and another being a biomimetic approach involving only aqueous solutions throughout the purification and spinning procedures and in which the molecular mechanisms and triggers for fiber formation are replicated. [ \n \n 50 \n , \n 51 \n , \n 52 \n , \n 53 \n \n ] The first approach enables expression of large spidroins that can be spun into fibers with high tensile strength, but the protein yields are far from what is required for industrial production. [ \n \n 54 \n , \n 55 \n \n ] Using the second approach, mini‐spidroins composed of an NT, a short repeat region consisting of two poly‐Ala/Gly‐rich blocks and a CT, have been developed. Such mini‐spidroins are extremely water‐soluble and can be spun into fibers using biomimetic spinning set‐ups. [ \n \n 51 \n , \n 52 \n , \n 53 \n , \n 56 \n \n ] Moreover, one of these mini‐spidroins, NT2RepCT, can be produced at a yield of 14.5 g L −1 in bioreactor cultivations which vouch for economically feasible bulk production. [ \n \n 55 \n , \n 56 \n \n ] Fibers spun from NT2RepCT are superior compared to previously published as‐spun fibers, but still, the fibers only reach about 15% of the native silk fiber's tensile strength. [ \n \n 1 \n , \n 51 \n \n ] NMR spectroscopy revealed that the mini‐spidroin's two poly‐Ala blocks are in an α‐helical conformation in the soluble state and convert to β‐sheet conformation in the as‐spun wet fiber, as expected. However, upon drying the fiber, the poly‐Ala blocks are transitioning back to α‐helical conformation, [ \n \n 57 \n \n ] which could lead to the inferior mechanical properties of dried NT2RepCT fibers compared to the native silk fiber. We therefore hypothesize that the mechanical properties of recombinant fibers could be improved by increasing the β‐strand propensity and inter‐β‐sheet interactions of the poly‐Ala blocks, [ \n \n 58 \n \n ] as it has been suggested by replacing the poly‐alanines with amyloidogenic sequences. [ \n \n 59 \n \n ] \n Notably, Ala residues have a low propensity to form β‐strands, whereas more hydrophobic residues like Val, Cys, Ile, and Phe show a higher β‐strand propensity, [ \n \n 60 \n \n ] and thus could be considered better candidates for forming stable β‐sheets in the silk fiber. However, being secretory proteins, the spidroins need to pass through the translocon when produced by the gland epithelium. [ \n \n 61 \n \n ] If the nascent polypeptide chain contains segments that are rich in Val, Ile, Cys, or Phe the translocon will mediate insertion into the endoplasmic reticulum membrane, [ \n \n 62 \n , \n 63 \n \n ] and thus any spidroin segment rich in these amino acid residues would be trapped in the cell. In fact, Ala is the most hydrophobic residue that allows passage through the translocon, which suggests that the spidroins have evolved to optimize hydrophobicity in their β‐sheet forming segments to the extent possible for a secretory protein. [ \n \n 58 \n , \n 60 \n \n ] Intracellular expression in prokaryotes will bypass the restrictions imposed by the secretory pathway that native spidroins must adhere to since translation and accumulation of the target protein takes place in the cytosol. These fundamental biological principles led us to use rational design and protein engineering to generate mini‐spidroins that potentially can be produced at high yields in prokaryotic hosts and be used to generate stronger biomimetic artificial spider silk fibers ( Figure \n 1 A,B ). The Zipper database [ \n \n 64 \n \n ] was used to screen a large panel of mini‐spidroins with designed modifications of the poly‐Ala blocks and candidates with low Rosetta energies were chosen for heterologous expression. Soluble target proteins were identified, characterized biochemically, and spun into fibers using a biomimetic spinning device. The mechanical performance of the fibers reveals that engineering of the repeat domain of mini‐spidroins is possible and can result in fibers with increased tensile strength. Figure 1 Schematic representation of the designed constructs. A) NT2RepCT (A 15 ‐A 14 ) is composed of an N‐terminal domain (NT, red; PDB: 4FBS), a repeat region with two poly‐Ala blocks (green and yellow), and a C‐terminal domain (CT, blue, PDB 3LR2). Both subunits of the soluble NT2RepCT dimer are shown (one is shaded). B) Protein sequence alignment of the repetitive region from A 15 ‐A 14 and engineered constructs thereof. Note that all constructs contain NT, a repeat part, and CT. Substitutions in the poly‐Ala blocks are indicated in orange.",
"discussion": "2 Results and Discussion Based on the β‐strand/α‐helix propensity ratios of amino acid residues as well as their hydrophobicity, Ile and Val were chosen to design 13 different constructs with substitutions in the poly‐Ala blocks of the original NT2RepCT sequence (referred to as A 15 ‐A 14 to reflect the composition of the two poly‐Ala blocks), (Figure 1 ). Additionally, the less hydrophobic residue Thr was used since it is branched at the β‐carbon and hence favors β‐strand conformation. [ \n \n 65 \n , \n 66 \n \n ] \n Figure 1B shows the amino acid sequences of the repetitive regions from A 15 ‐A 14 and engineered constructs with substitutions indicated (complete sequences can be found in Table S1 , Supporting Information). Substitutions were mainly introduced at every second position resulting in β‐strands with mutated side chains on the same side. Mutations were introduced in either both (e.g., (AV) 7 ‐(AV) 7 ) or only in one of the poly‐Ala blocks (e.g. (AV) 7 ‐A 14 ). The number of substitutions varied between 15 (e.g., V 15 ‐A 14 , in which all Ala are replaced by Val in the first poly‐Ala block) and 3 as in, for example, (A 3 V) 3 ‐(A 14 ), which contains Val substitution at every fourth position in the first poly‐Ala block. A few additional constructs were designed to analyze the impact of the position of the substituted residues, for example, (A 3 I) 3 ‐A 14 , A 15 ‐(A 3 I) 3 and IA 6 IA 6 I‐A 14 that all have three Ile substitutions but in different locations. The packing of β‐sheets in amyloid‐like fibrils involve steric zippers, [ \n \n 41 \n , \n 67 \n \n ] which are also found in spider silk β‐sheet crystals. [ \n \n 36 \n , \n 43 \n \n ] Steric zippers are formed by tightly bound β‐strands with high complementarity of the involved side chains. [ \n \n 41 \n , \n 67 \n \n ] The Zipper database predicts the stability and propensity of hexapeptides in a given amino acid sequence to form steric zippers by calculating the energies of the interstrand interactions. Rosetta energies equal or below −23 kcal mol −1 suggest a high propensity to form steric zippers. [ \n \n 64 \n \n ] \n \n Figure \n 2 A shows the Rosetta energies estimated for constructs A 15 ‐A 14 and (A 3 I) 3 ‐A 14 (corresponding profiles for all engineered mini‐spidroins are shown in Figure S1 , Supporting Information, and summarized in Table S2 , Supporting Information, and Figure 2B ). As expected, the hexapeptides in the poly‐Ala region of the A 15 ‐A 14 construct have low Rosetta energies (−24.6 kcal mol −1 ) and thus should be able to form steric zippers (Figure 2C ). All designed constructs contain at least one hexapeptide with a Rosetta energy lower than that of A 15 ‐A 14 (Table S2 , Supporting Information), ranging from −24.9 to −29.4 kcal mol −1 (for (AT) 7 ‐(AT) 7 and V 15 ‐A 14 , respectively). Generally, the effect on the Rosetta energies increased with an increasing number of hydrophobic replacements in the poly‐Ala region. Figure 2 Rosetta energy profiles of A) A 15 ‐A 14 and (A 3 I) 3 ‐A 14 (profiles for all designed proteins are found in Figure S1 and Table S2 , Supporting Information). Bars show Rosetta energies for moving hexapeptides (indicated at the first residue of each hexapeptide), red bars indicate Rosetta energies equal or below −23 kcal mol −1 (dashed line). Green bars indicate Rosetta energies above the threshold and are unlikely to form steric zippers ( https://services.mbi.ucla.edu/zipperdb/ ). [ \n \n 64 \n \n ] B) Bars indicate the Rosetta energy of the hexapeptide with the lowest predicted energy from A 15 ‐A 14 and the engineered mini‐spidroins (all hexapeptides are shown in Table S2 , Supporting Information). C) Hypothetical zipper structure of two β‐sheets composed of hexapeptides AAAAAA from A 15 ‐A 14 and AIAAI derived from (A 3 I) 3 ‐A 14 , respectively. Of the 15 designed proteins, seven were overexpressed and six were highly overexpressed in E. coli BL21 cells ( Table \n 1 \n and Figures S2 and S3 , Supporting Information). Constructs with Val substitutions had lower expression levels than corresponding constructs with Ile substitutions, but the number of substitution and the hydrophobicity did not have any general impact on expression levels (Figure S4 , Supporting Information). The (AT) 7 ‐(AT) 7 construct did not express well which could be due to that this repeat was designed to resemble a “CAT tail” which is known to lead to aggregation of the nascent polypeptide chain and to degradation by the proteasome. [ \n \n 68 \n \n ] \n Table 1 Summary of number of substitutions, expression levels, solubility after cell lysis, protein yield, and spinnability into fibers of the engineered proteins. Expression levels, solubility after cell lysis, and spinnability into fibers are rated from very high (+++), intermediate (++), low (+), and not at all (0). Rating of expression level and solubility after cell lysis were estimated by appearance of the target band on SDS‐PAGE (Figures S2 and S3 , Supporting Information). (−) indicates not tested. ( 1 ) indicates degradation during expression. (*) marks purification using gravity columns instead of FPLC Construct Number of substitutions Expression levels Solubility after cell lysis Average protein yield [mg L −1 culture] Spinnability into fibers 1. A 15 ‐A 14 \n 0 +++ +++ 250 +++ 2. (AT) 7 ‐(AT) 7 \n 14 + − − − 3. (A 3 T) 3 ‐(A 3 T) 3 \n 6 ++ +++ 58* +++ 4. (AV) 7 ‐(AV) 7 \n 14 +++ 0 − − 5. (AV) 7 ‐A 14 \n 7 +++ 0 − − 6. V 15 ‐A 14 \n 15 +/++ 1 \n 0 − − 7. (A 3 V) 3 ‐(A 3 V) 3 \n 6 +++ +++ 139* +++ 8. (A 3 V) 3 ‐A 14 \n 3 ++ +++ 216 +++ 9. (AI) 7 ‐(AI) 7 \n 14 + + 4* − 10. A 15 ‐(AI) 7 \n 7 + + − − 11. (AIA 2 ) 3 ‐(AIA 2 ) 3 \n 8 ++ + − − 12. (A 3 I) 3 ‐(A 3 I) 3 \n 6 +++ +++ 94* + 13. (A 3 I) 3 ‐A 14 \n 3 +++ +++ 207 +++ 14. A 15 ‐(A 3 I) 3 \n 3 +++ +++ 233 +++ 15. (A 2 I) 4 ‐A 14 \n 4 ++ +++ 243 +++ 16. IA 6 IA 6 I‐A 14 \n 3 ++ ++ 139 − John Wiley & Sons, Ltd. In addition to A 15 ‐A 14 , seven of the constructs were found mainly in the soluble fraction after cell lysis in 20 m m Tris‐HCl, and four constructs were in both the soluble and insoluble fraction (Table 1 and Figure S3 , Supporting Information). Increased hydrophobicity, number of substitutions, and lower Rosetta energies correlated with lower solubility after cell lysis (Figure S4 , Supporting Information). 9 of the 15 designed constructs plus the control A 15 ‐A 14 yielded sufficient soluble protein for purification. Nondenaturing immobilized metal affinity chromatography yielded between 4 and 243 mg of pure target protein per 1 L shake flask culture (average of 10 × 1 L cultures). Notably, six of the engineered mini‐spidroins gave very high yields (>100 mg L −1 Table 1 ). (AV) 7 ‐(AV) 7 , (AV) 7 ‐A 14 , and V 15 ‐A 14 expressed well but were insoluble after lysis, likely due to high hydrophobicity of the engineered segments. Expression and purification of the A 15 ‐(AI) 7 and (AIA 2 ) 3 ‐(AIA 2 ) 3 constructs did not result in enough soluble protein for further characterization. The constructs that showed intermediate to high expression levels but were insoluble after cell lysis were treated with 8 m urea but could not be solubilized to the extent needed for enabling purification of enough protein for fiber spinning (not shown). The position of the Ile replacements within one Ala block had an impact on the protein yield but whether these were located in the first or second poly‐Ala block did not matter. For example, (A 3 I) 3 ‐A 14 and A 15 ‐(A 3 I) 3 both have three Ile substitutions in the first and second poly‐Ala block, respectively, and show comparable yields. In contrast, (A 3 I) 3 ‐A 14 and IA 6 IA 6 I‐A 14 have the same number of Ile replacements in the first block, but their location differ as does the yield (207 vs 139 mg L −1 culture for (A 3 I) 3 ‐A 14 and IA 6 IA 6 I‐A 14 , respectively). Next, we investigated the secondary structure content and the thermal stability of the purified constructs by circular dichroism (CD) spectroscopy ( Figure \n 3 \n ). We found that all constructs had an overall α‐helical secondary structure (Figure 3A ) which indicates that the amino acid substitutions did not affect the secondary structure of the soluble proteins to any large extent. Heating to 90 °C led to a decreased signal for all constructs and concomitant transition to β‐sheet dominated secondary structures (Figure 3C ). The heat‐induced conformational changes were irreversible upon cooling of the samples (Figure 3D ). Melting curves for all constructs showed that the proteins unfolded around 46–50 °C, which is in line with reports on the isolated terminal domains, [ \n \n 15 \n \n ] and means that the substitutions in the repetitive region of the mini‐spidroins only had a minor effect on the thermal stability of the proteins (Figure 3B ). Figure 3 CD spectroscopy of purified engineered mini‐spidroins. A) Initial spectra at 20 °C and B) molar ellipticity measured at 222 nm from 20 to 90 °C was converted to fraction natively folded (%) and then normalized. CD spectroscopy of different constructs C) heated to 90 °C and D) after cooling to 20 °C. Out of the nine engineered mini‐spidroins that were successfully purified (excluding A 15 ‐A 14 ), eight could be concentrated to at least 200 mg mL −1 to generate spinning dopes, while (AI) 7 ‐(AI) 7 yielded too little protein (Table 1 ). The dopes made from the eight constructs were transferred to syringes and extruded through a thin glass capillary into a low pH aqueous buffer according to a previously described biomimetic spinning procedure. [ \n \n 50 \n , \n 51 \n \n ] Seven engineered mini‐spidroins could be spun into fibers, and only the IA 6 IA 6 I‐A 14 protein aggregated prematurely in the syringe. One of the mini‐spidroins, (A 3 I) 3 ‐(A 3 I) 3 , formed fibers that were too fragile to be retrieved. The reason for the poor integrity of the (A 3 I) 3 ‐(A 3 I) 3 fibers is not known but was not related to premature aggregation in the dope. The other six engineered fiber types, plus the A 15 ‐A 14 fibers, were successfully collected onto a motorized wheel at the end of the spinning bath ( Figure \n 4 A and Video S1 , Supporting Information). There was no difference in the appearance of the spun fibers (Figure 4B ) and the diameter of the different fiber types, determined by light microscopy, varied between 4 and 19 µm (Figure S5H and Table S3 , Supporting Information). The reason for the differences in diameter between the different fiber types is not known but is likely linked to differences in the properties of the proteins since the spinning conditions were kept constant. Figure 4 Mechanical properties of spinnable engineered mini‐spidroins in comparison with A 15 ‐A 14 . A) Photographs of the biomimetic spinning set‐up; a video of the spinning can be found in Video S1 , Supporting Information. B) Photographs of spun fibers. C) Strength, D) strain at break, E) toughness modulus, dashed line indicates toughness modulus of a native dragline silk, [ \n \n 10 \n \n ] and F) Young's modulus. Whiskers show standard deviation. * p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001. Representative stress–strain graphs for all spinnable engineered mini‐spidroins are shown in Figure S5A–G , Supporting Information. The diameters of the fibers are shown in Figure S5H , Supporting Information. The values and corresponding standard deviations are shown in Table S3 , Supporting Information. The tensile strength of all fibers spun from engineered proteins increased significantly compared to A 15 ‐A 14 except for (A 3 V) 3 ‐A 14 and (A 2 I) 4 ‐A 14 (Figure 4 and Table S3 , Supporting Information). The two similar fiber types (A 3 I) 3 ‐A 14 and A 15 ‐(A 3 I) 3 displayed the highest increase in strength, the former reaching 131 MPa, which is almost three times higher than that of A 15 ‐A 14 (Figure 4C ). This indicates that rational protein engineering of the spidroin poly‐Ala blocks indeed can result in increased fiber tensile strength and stiffness. Unexpectedly, the introduced amino acid substitutions also had a high impact on the extensibility of the fiber, as the strain at break varied from 0.03 to 2.0 mm mm −1 (Figure 4D and Table S3 , Supporting Information). The two strongest fiber types ((A 3 I) 3 ‐A 14 and A 15 ‐(A 3 I) 3 ) displayed an exceptional increase in strain (to 1.6 and 2.0 mm mm −1 , respectively), while (A 3 V) 3 ‐A 14 , (A 2 I) 4 ‐A 14 fibers showed moderately increased strain (0.79 and 0.85 mm mm −1 , respectively) compared to A 15 ‐A 14 (0.45 mm/mm). (A 3 T) 3 ‐(A 3 T) 3 and (A 3 V) 3 ‐(A 3 V) 3 fibers were the least extensible (0.03 and 0.08 mm mm −1 , respectively). These two proteins contain substitutions in both poly‐Ala blocks and, possibly, the reason for the inferior strain of these fibers could be an increased propensity of the engineered segments to interact intra‐molecularly over forming intermolecular contacts. Apparently, the mechanical properties of artificial spider silk fibers can be significantly improved by introducing Ile in every fourth position in the first or second poly‐Ala block. These two mini‐spidroins, (A 3 I) 3 ‐A 14 and A 15 ‐(A 3 I) 3 , formed fibers with a toughness modulus that is comparable to native dragline silk (146 and 125 MJ m −3 , respectively, compared to 136 MJ m −3 for a native dragline silk from Argiope argentata ), (Figure 4E ). [ \n \n 10 \n \n ] Fibers formed by (A 3 V) 3 ‐A 14 and (A 2 I) 4 ‐A 14 also reached a significantly higher toughness modulus than A 15 ‐A 14 (50 and 37 MJ m −3 , respectively, compared to 18 MJ m −3 ). To investigate the link between fiber secondary structure content and mechanical properties, we used attenuated total reflection Fourier‐transform infrared (ATR‐FTIR) spectroscopy. The results, shown in Figure \n 5 \n and Figure S7 and Table S4 , Supporting Information, indicate that no large differences in secondary structure content between fibers were detected, but (A 3 V) 3 ‐A 14 , (A 3 I) 3 ‐A 14 and A 15 ‐(A 3 I) 3 had a slightly increased β‐sheet content, along with decreased α‐helix/random coil content compared to A 15 ‐A 14 fibers. However, the (A 3 V) 3 ‐(A 3 V) 3 and (A 2 I) 4 ‐A 14 fibers failed to show increased β‐sheet content compared to A 15 ‐A 14 fibers and we could detect no strong correlations between secondary structure content and mechanical properties of the fiber (Figure S6 , Supporting Information). Thus, ATR‐FTIR spectroscopy of the different fiber types did not detect any significant differences in secondary structure content. Therefore, we decided also to use solid‐state NMR spectroscopy to investigate the unmodified fibers (A 15 ‐A 14 ) and the best performing engineered fibers, (A 3 I) 3 ‐A 14 . As expected, more Ala residues were found in a β‐sheet conformation in (A 3 I) 3 ‐A 14 compared to A 15 ‐A 14 fibers ( Figure \n 6 \n ). Figure 5 FTIR spectroscopy of engineered fibers. Normalized and baseline‐subtracted absorbance spectrum in the amide I region of A) A 15 ‐A 14 , (A 3 V) 3 ‐(A 3 V) 3 , (A 3 V) 3 ‐A 14 , and (A 3 T) 3 ‐(A 3 T) 3 and B) A 15 ‐A 14 , (A 3 I) 3 ‐A 14 , A 15 ‐(A 3 I) 3 , and (A 2 I) 4 ‐A 14 . C) Percent secondary structure content determined by cofitting the absorbance spectrum and the second derivative. Horizontal line indicates β‐sheet content of A 15 ‐A 14 . Fits of absorbance spectra and second derivative of fibers spun are shown in Figure S7 , Supporting Information. Figure 6 Solid‐state NMR 13 C‐ 13 C correlation spectra (aliphatic region) of A 15 ‐A 14 (blue) and (A 3 I) 3 ‐A 14 (red) fibers. The Cα/Cβ correlations of Ala and Ile in α‐helical and β‐sheet conformation are indicated. The altered mechanical properties of the fibers made from the engineered spidroins indicate that intermolecular interactions in the spidroins are affected. In the native dragline silk fiber, pulling the fiber first results in reversible deformation of the amorphous regions up until the yielding point, after which the hydrogen bonds in the amorphous region break, resulting in softening of the material. [ \n \n 36 \n , \n 43 \n \n ] When the amorphous protein chains are extended, the load is transferred onto the β‐sheet crystals leading to a stiffening of the fiber. Upon further increased load, the β‐sheet crystals undergo stick‐slip deformation and the fiber breaks. [ \n \n 36 \n , \n 43 \n , \n 69 \n \n ] The increased tensile strength of the fibers made from engineered proteins suggests that our strategy to increase the β‐strand propensity and inter‐β‐sheet interactions indeed can result in stronger fibers, although some of the engineered fibers concomitantly displayed a decreased strain. Theoretically, increased β‐sheet formation and intermolecular interactions in the stacked β‐sheets could not only result in increased fiber strength, but also increased extensibility, since the amorphous region would be allowed to extend fully before the load is transferred to the crystalline region. In lack of poly‐Ala β‐sheet crystals, as in the A 15 ‐A 14 fibers, the intermolecular contacts may be too weak to allow a full extension of the amorphous protein chains before fiber failure. At the same time, it may be disadvantageous that all β‐sheets stack in crystals since only about 40% of the Ala residues in the native dragline silk are found in this conformation and the rest form less ordered β‐sheets. [ \n \n 70 \n \n ] In this study, introducing replacements in both poly‐Ala blocks resulted in fibers with dramatically reduced strain which suggest a suboptimal packing of the proteins in the fiber. Since the (A 3 I) 3 ‐A 14 fibers displayed superior mechanical properties, these fibers are attractive candidates for bulk‐scale production. Previously, A 15 ‐A 14 has been shown to express at very high levels (≈21 g L −1 ) in a bioreactor‐based E. coli fed‐batch culture. [ \n \n 56 \n \n ] Following the same protocol, the expression level of (A 3 I) 3 ‐A 14 amounted to 13 g L −1 and the final yield after purification using an automated purification protocol was 8.9 g L −1 (Figure S8 A,B, Supporting Information). To our knowledge, these yields are the second highest reported for any recombinant spidroin produced in E. coli and line with what is required for economically viable bulk production. [ \n \n 55 \n , \n 56 \n \n ] After purification, (A 3 I) 3 ‐A 14 was concentrated to 300 mg mL −1 and could easily be spun into fibers. Notably, 8.9 g recombinant silk protein is enough to produce an ≈18 km long fiber. When comparing (A 3 I) 3 ‐A 14 fibers produced from proteins recovered from bioreactor and shake flask fermentations, respectively, the former had slightly lower strength (Figure S9D , Supporting Information). However, the bioreactor produced (A 3 I) 3 ‐A 14 fibers still had a significantly higher tensile strength and strain compared to A 15 ‐A 14 fibers (Figure S9 , Supporting Information)."
} | 7,180 |
38343096 | PMC10934265 | pmc | 11 | {
"abstract": "Despite the considerable\ninterest in the recombinant production\nof synthetic spider silk fibers that possess mechanical properties\nsimilar to those of native spider silks, such as the cost-effectiveness,\ntunability, and scalability realization, is still lacking. To address\nthis long-standing challenge, we have constructed an artificial spider\nsilk gene using Golden Gate assembly for the recombinant bacterial\nproduction of dragline-mimicking silk, incorporating all the essential\ncomponents: the N-terminal domain, a 33-residue-long major-ampullate-spidroin-inspired\nsegment repeated 16 times, and the C-terminal domain (N16C). This\ndesigned silk-like protein was successfully expressed in Escherichia coli , purified, and cast into films from\nformic acid. We produced uniformly 13 C– 15 N-labeled N16C films and employed solid-state magic-angle spinning\nnuclear magnetic resonance (NMR) for characterization. Thus, we could\ndemonstrate that our bioengineered silk-like protein self-assembles\ninto a film where, when hydrated, the solvent-exposed layer of the\nrigid, β-nanocrystalline polyalanine core undergoes a transition\nto an α-helical structure, gaining mobility to the extent that\nit fully dissolves in water and transforms into a highly dynamic random\ncoil. This hydration-induced behavior induces chain dynamics in the\nglycine-rich amorphous soft segments on the microsecond time scale,\ncontributing to the elasticity of the solid material. Our findings\nnot only reveal the presence of structurally and dynamically distinct\nsegments within the film’s superstructure but also highlight\nthe complexity of the self-organization responsible for the exceptional\nmechanical properties observed in proteins that mimic dragline silk.",
"conclusion": "Conclusions In this study, we have\ndemonstrated the successful cloning and\nexpression of a designed, synthetic spider dragline silk based on\nthe amino acid sequence of the dragline silk proteins of Nephila clavipes . The construct contained all the\ncharacteristic building blocks of natural spider silks, including\nthe nonrepetitive N- and CTDs, as well as the repetitive core domains\nrepeated 16 times (N16C), resulting in a final protein size of 68.1\nkDa. We achieved the straightforward directional assembly of the modules\nusing Golden Gate assembly, which can be easily utilized to further\nincrease the repetition size. Additionally, we developed an\nefficient recombinant protein production\nand purification strategy and characterized the structure and dynamics\nof the recombinantly produced, 13 C, 15 N-labeled\nspidroin mimic with solution- and solid-state NMR spectroscopy. By\nanalyzing the 1 H, 13 C, and 15 N chemical\nshifts, we evaluated the secondary structure of the core repetitive\ndomain of N16C dissolved in DMSO- d 6 . Furthermore,\nwe analyzed the structure and dynamics of hydrated N16C film cast\nfrom formic acid solution using proton-detected solid-state MAS methods,\nemploying both cross-polarization and J -coupling-based\nmagnetization transfers. Comparing the solution and solid-state chemical\nshifts, as well as the 15 N and 13 C relaxation\nrate constants measured in the solid state, we identified three structurally\nand dynamically distinct segments in the film superstructure. These\nsegments include a rigid, strongly hydrogen-bonded β-nanocrystalline\ncore surrounded by a solvent-exposed, dynamic α-helical shell\nin exchange with fully solubilized, flexible repeat units with random\ncoil characteristics. These findings highlight the complexity of the\nhierarchical organization responsible for the remarkable mechanical\nproperties of dragline-silk-mimicking proteins. By harnessing the\npower of recombinant silk production and advanced spectroscopic techniques,\nwe are one step closer to understanding the correlation between molecular\nstructure and mechanical response of silk-based high-performance materials.",
"introduction": "Introduction Spider dragline silks, or spidroins, are\nnatural fibers with exceptional\nproperties including their unique lightweight, impressive extensibility,\nhigh tensile strength, durability, and biocompatibility. 1 − 7 These unparalleled qualities make spidroins valuable materials for\nvarious industrial and biomedical applications. Throughout history,\nthey have been utilized in diverse fields such as clothing, fishing,\npainting, medicine, and weaponry. 4 , 8 Despite significant\nefforts in replicating the mechanical features of spidroins, artificial\nfibers often fall short in terms of versatility and overall performance\ncompared to those of natural spider silks. 9 − 11 However, recent\nadvancements in producing spidroins with previously unattainable molecular\nweights through heterologous bacterial expression 12 or utilizing transgenic silkworms 13 are bringing us closer to the commercialization of tailor-made bioengineered\nsilk fibers. Spider dragline silk has attracted considerable attention\ndue to its exceptionally high tensile strength, exceeding that of\nsteel, and toughness that is three times higher than that of Kevlar. 4 , 14 Most dragline silks are made from two major ampullate spidroin proteins\n[major ampullate spidroin protein 1 (MaSp1) and major ampullate spidroin\nprotein 2 (MaSp2)], which share a common structure, characterized\nby a low-complexity, highly repetitive core with up to a few hundred\nrepeats of 20–200 amino acids. This core is flanked by highly\nconserved, nonrepetitive α-helical N- and C-terminal domains\n(NTD and CTD). The repeat regions have a block copolymer structure\nin which hydrophilic “soft” glycine (Gly)-rich and hydrophobic\n“hard” alanine (Ala)-rich segments alternate. These\nsegments are composed of shorter consensus motifs, including poly(A),\npoly(GA), GGX, GSG, QQ, and GPGXX stretches, where X = Y (Tyr, tyrosine),\nL (Leu, leucine), or Q (Gln, glutamine). Solid-state nuclear\nmagnetic resonance (NMR) data suggested that\nthe Gly-rich segments predominantly adopt semiextended 3 1 -helices, while the GPGGX and GPGQQ elements of MaSp2 form elastin-like\ntype II β-turns. 15 − 24 Alternatively, these segments can be incorporated into β-sheet\nstructures, and thus, they are part of the hard segments. 19 , 20 , 24 , 25 The low density of hydrogen bonds in the soft Gly-rich regions grants\nextensibility to the dragline fiber, 7 , 26 while the\nsuccessive β-turns are associated with the supercontractive\nproperty of the silk fiber. 21 Upon sheer-induced\nstress, the majority of the alanine residues arrange into β-sheets\nand form β-nanocrystals, whereas the alanines in poly(GGA) adopt\n3 1 -helix structures. 15 , 18 , 24 , 25 In the β-nanocrystals,\nthe poly(A) β-sheets tightly interlock, resembling the steric\nzippers of amyloid crystals, thereby impeding water penetration between\nthe poly(A) sheets. 17 The relatively small\ncrystalline domains 27 − 31 embedded within the semiamorphous matrix act as intermolecular cross-links,\ncontributing to the high tensile strength of the spun fibers. 7 , 24 , 26 Additionally, the large number\nof repeated core units enhances interchain interactions and reduces\nchain-end defects, further augmenting the unprecedented tensile strength\nof the silk fibers. 7 , 12 , 13 , 32 The importance of the high repeat\nnumber in recombinant spider\nsilk proteins was recognized early on, 33 − 35 and attempts have been\nmade to increase the repeat number all the way up to 192. 12 , 32 Achieving such a feat required the implementation of extensive metabolic\nengineering and synthetic biology approaches, which have not yet been\nadapted for large-scale biotechnological production. In addition to\nthe repeat number, the terminal domains play fundamental roles in\ndetermining the properties of silk proteins, especially with respect\nto protein gland solubility and initiation of fiber assembly through\nsalt- and pH-dependent dimerization. 36 − 38 Previous studies have\nestablished a correlation between recombinantly produced silk’s\ntensile properties, such as the Young’s modulus, strength,\nand toughness of the fibers, with the presence of the CTD and the\nNTD. 9 , 39 For such reasons, the inclusion of the highly\nconserved globular terminal domains in the final spidroin product\nis necessary, especially when assembly and solubility properties are\nof interest. 36 Repetitive DNA sequences\nprovide challenges for standard cloning\ntechniques, for example, due to their inherent ability to recombine\nand hence instability. Typically, recombinant tandem-repeat DNA sequences\nof fibrous proteins, including mimics of spider silks (reviewed in\nrefs ( 5 , 11 , 40 – 42 )), collagen, 43 elastin, 44 , 45 keratin, 46 and resilin, 47 are constructed via stepwise concatenation, 48 , 49 recursive directional ligation, 12 , 44 , 45 or step-by-step directional approaches, 50 − 52 which involve several rounds of plasmid amplification, digestion,\nligation, and possible sequencing between repeat extensions. 52 − 54 In these stepwise techniques, at each oligomerization step, the\ncloning efficiency is greatly reduced by the possibility of obtaining\nempty vectors or self-ligated and circularized inserts. Aside from\nthe time and material costs associated with these traditional approaches,\na major drawback of a few of these methods is the scars that remain\nat the recombinant sites in the final constructs. These scars are\nthen translated into extraneous amino acids in the primary protein\nsequence, compromising the accuracy and potentially the structural\nintegrity of the expressed protein. Nevertheless, seamless stepwise\ncloning has been successfully applied in a few instances to produce\nrecombinant fibrous proteins. 44 , 55 − 59 To overcome these shortcomings, in this study, we used Golden\nGate\nassembly 60 to generate an expression construct\nfor a spidroin mimic, called N16C ( Figure 1 ), which includes both the N- and C-terminal\ndomains and 16 repeats of the repetitive core units. Golden Gate assembly\nrelies on type II restriction endonucleases, which cut double-stranded\nDNA outside their recognition sequence and leave a short, single-stranded,\nuser-defined overhang that guarantees ordered gene assembly of multiple\nconstituents. The restriction sites are eliminated during subcloning,\nwhich allows for simultaneous digestion and ligation in a one-pot\nreaction and facilitates the seamless assembly of gene constituents.\nPreviously, Golden Gate assembly has been successfully used to straightforwardly\nassemble repetitive DNA sequences encoding elastin-like proteins (ELPs). 59 Figure 1 Schematic representation and amino acid sequence of the\nN16C protein\nconstruct: NTD (blue), derived from the N. clavipes MaSp2, the repetitive central domain (S16, black), based on N. clavipes MaSp1 and MaSp2, and CTD (maroon) from N. clavipes MiSp1. Amino acids labeled with green\naccount for the His-tag used for purification purposes. High repeat numbers and the presence of terminal\ndomains\nare necessary\nbut are not sufficient requirements for reproducing the properties\nof natural silk fibers. The failure in natural silk reproduction arises\nfrom the incomplete understanding of the driving forces behind the\nprotein’s self-assembly across multiple spatial scales. Open\nquestions include the exact structures of the involved molecules and\nsequence motifs, the types of interactions and transformations they\nundergo, the kinetics and thermodynamics of their interactions, and\nthe dynamics that affect the local energetics and the macroscale mechanics\nof the silk fibers. Addressing these questions requires atomic-level\ninsights into the protein assembly both experimentally and computationally.\nSolid-state magic-angle spinning (MAS) NMR spectroscopy is arguably\nthe most suitable experimental technique to investigate complex heterogeneous\nsystems as it has been demonstrated for natural 17 − 21 , 23 − 25 , 61 − 65 and genetically engineered spider silks, 4 , 56 , 66 − 70 or for selectively isotope-enriched silk-mimicking\nmodel peptides as reviewed in refs ( 71 – 74 ). To fully exploit the advantages of recombinant spider silk\nproduction,\nwe expressed and purified a uniformly 13 C– 15 N-labeled version of our designed spidroin mimic N16C and measured\nits solid-state MAS NMR spectra to gain insights into the local structure,\nhydrogen-bonding, nanosecond and microsecond time-scale dynamics,\nand hydration-induced macroscopic organization of each amino acid\ntypes in the cast film. Fast MAS (55.55 kHz) and high magnetic fields\n(700 MHz) combined with proton detection significantly improved the\nsensitivity and the resolution of the acquired spectra, allowing for\nthe almost complete amino-acid-specific assignment of the repetitive\ncore, which then facilitated the analysis of site-specific dynamics\nacross multiple time scales. The uncovered correlation between the\natomic-level structure and hydration-induced local dynamics of the\ndesigned synthetic spidroins provides critical insight into the dynamic\norganization of natural silks and engineered silk-like proteins.",
"discussion": "Discussion Spidroins\nhave attracted significant attention in the fields of\nbiotechnology and material science due to their remarkable properties,\nincluding the ability to form strong, stable, and tough fibers. However,\ntheir expression in alternative hosts like E. coli presents challenges, primarily due to the inherent instability and\nrepetitive nature of the encoding DNA sequences, as well as their\nlimited solubility in the expression host. 93 Consequently, extensive efforts have been dedicated to improving\nprotein production and purification processes, optimizing cloning\nstrategies, and designing genetic circuits to regulate gene expression\nresources. Despite these endeavors, achieving a high yield in fibrous\nprotein production remains an unresolved issue. Encouragingly, we\nmanaged to successfully express a 68 kDa model spidroin protein, N16C,\ncomprising 16 repeat units, with a comparatively high yield in TB\nmedium, without the need for additional metabolic or genetic engineering\ninterventions. The Golden Gate assembly technique employed in constructing\nthe silk gene sequence in this study can be readily applied to generate\nDNA sequences encoding 32, 64, or even 128 repeat units, thereby enabling\nthe expression of N32C, N64C, or N128C spidroin mimics, respectively.\nHowever, it is important to note that producing such large repeat\nproteins will undoubtedly present several challenges associated with\nheterologous protein expression. An extra benefit of being able\nto express silk-like proteins in E. coli is that they can be easily produced in isotopically\nenriched forms for downstream NMR investigations. Solid-state NMR\nprovides unprecedented insights into the atomic-resolution structure,\ndynamics, organization, and interaction of semiamorphous semicrystalline\nsystems, like silk. The sensitivity enhancement achieved by uniform 13 C and 15 N labeling combined with proton detection\nand fast solid-state MAS NMR gave us access to a wide range of multidimensional\nNMR experiments that we customized to focus on specific aspects of\nthe material property. For example, we used two fundamentally different\nmagnetization transfers (cross-polarization and INEPT) to access either\nthe rigid core or the solubilized flexible segments of the hydrated\nN16C film, and we built these transfer steps into both assignment\nexperiments and spin relaxation measurements. The resulting chemical\nshifts and relaxation rate constants allowed for the extraction of\nsite-specific information on hydrogen bonding, secondary structure,\nand nanosecond-to-microsecond time scale dynamics of both rigid and\nsoluble states. Since Ala and Gly are the major constituents of the\nprotein sequence, we focused more on their analysis as they selectively\nreport on the properties of the hard and soft segments. Since\nin our designed spidroin sequence, only one type of Ala exists\nin the repetitive segment, the observed three different sets of Ala\nresonances must stem from the same poly(A) sequence in different chemical\nenvironments. Based on their chemical shifts and relaxation properties,\nwe assigned them to belong to Ala (i) inside the β-nanocrystals,\n(ii) at the water–protein interface, (iii) and in solution.\nIn the solid-state MAS spectra of native dragline silk fibers, Holland\net al. observed β-stranded as well as α-helical Ala and\nGly environments, but they associated the separate shifts with sequentially\ndifferent residues, e.g., with Ala that is part of the poly(A) segments\nor Ala in GGA elements flanking the poly(A) regions. 19 , 20 , 63 The structural and dynamic\ninformation on the Ala resonances suggested\na model for the repetitive region of the N16C film, where a densely\npacked β-nanocrystalline core is surrounded by a solvent-exposed\nlayer with α-helical conformation that is in dynamic exchange\nwith fully solubilized repetitive units of N16C, which are highly\nflexible and unstructured ( Figure 8 ). Figure 8 Schematic representation of the conformational states\nin the hydrated\nN16C film. The enlargement shows a structural model of the repetitive\nregion in which the poly(Ala) and some of the Gly-rich segments form\nβ-sheets and arrange into randomly oriented β-nanocrystals\n(blue), whereas the rest of the Gly-rich segments are unordered or\nadopt a 3 1 -helical structure. At the interface of the β-nanocrystals\nand amorphous segments, the poly(Ala) repeats form α-helices\n(red). Some protein chains are fully solubilized and adopt a random\ncoil conformation (green). The majority of Gly in the soft segment adopted\nan extended β-sheet\nconformation and became part of the β-nanocrystals, while a\nsmaller fraction was found in a different conformation that we tentatively\nassociated with the semiextended 3 1 -helical conformation.\nThe large fraction of Gly in β-sheet could be a consequence\nof casting the film from formic acid. Formic acid is known to initiate\nβ-sheet formation in regenerated Bombix mori silk fibroin and in recombinant spider silk films. 69 , 94 − 96 The freshly prepared dry N16C film was brittle but\nbecame elastic after isopropanol treatment and incubation in water\novernight. Stiffness and brittleness are associated with a high number\nof β-sheets and long-range order. 84 , 94 As water interacts\nwith the ordered β-crystalline structures, the crystallinity\ndecreases and the film becomes more elastic, as demonstrated on B. mori fibroin films. 94 , 97 In the hydrated\nN16C film, even though most Gly was incorporated into the rigid β-crystalline\ndomain, it actively contributed to the elasticity of the film as it\nshowed extensive μs time scale flexibility. Our findings align\nwith the results of other X-ray and NMR studies of spider silk fiber\nthat suggested a three-phase model, where the rigid β-crystalline\ncore and the elastic amorphous matrix are interconnected by a semiordered\nphase. 3 , 98"
} | 4,713 |
19229199 | null | s2 | 12 | {
"abstract": "The extreme strength and elasticity of spider silks originate from the modular nature of their repetitive proteins. To exploit such materials and mimic spider silks, comprehensive strategies to produce and spin recombinant fibrous proteins are necessary. This protocol describes silk gene design and cloning, protein expression in bacteria, recombinant protein purification and fiber formation. With an improved gene construction and cloning scheme, this technique is adaptable for the production of any repetitive fibrous proteins, and ensures the exact reproduction of native repeat sequences, analogs or chimeric versions. The proteins are solubilized in 1,1,1,3,3,3-hexafluoro-2-propanol (HFIP) at 25-30% (wt/vol) for extrusion into fibers. This protocol, routinely used to spin single micrometer-size fibers from several recombinant silk-like proteins from different spider species, is a powerful tool to generate protein libraries with corresponding fibers for structure-function relationship investigations in protein-based biomaterials. This protocol may be completed in 40 d."
} | 271 |
33458606 | PMC7797931 | pmc | 13 | {
"abstract": "Summary The expeditious development of information technology has led to the rise of artificial intelligence (AI). However, conventional computing systems are prone to volatility, high power consumption, and even delay between the processor and memory, which is referred to as the von Neumann bottleneck, in implementing AI. To address these issues, memristor-based neuromorphic computing systems inspired by the human brain have been proposed. A memristor can store numerous values by changing its resistance and emulate artificial synapses in brain-inspired computing. Here, we introduce six types of memristors classified according to their operation mechanisms: ionic migration, phase change, spin, ferroelectricity, intercalation, and ionic gating. We review how memristor-based neuromorphic computing can learn, infer, and even create, using various artificial neural networks. Finally, the challenges and perspectives in the competing memristor technology for neuromorphic computing systems are discussed.",
"conclusion": "Conclusion and perspectives The massive data produced by numerous edge devices and sensors have supported the striking development of AI. Several research institutes have estimated that the total global data amount will grow to 175 zettabytes by 2025 ( Reinsel et al., 2018 ). This has necessitated higher-performance computing power to process big data. However, it has been predicted that the conventional CMOS-based computing systems will reach their physical and performance limits by 2024, according to the 2018 International Roadmap for Devices and Systems. These facts necessitate a paradigm shift to post-CMOS technology to overcome volatility, multi-state parallel operations, and von Neumann bottlenecks. Intensive research has demonstrated that memristor-based hardware achieves groundbreaking performance in neuromorphic computing, which is inspired by the human brain. They are considered excellent candidates for alternate conventional computing systems. However, despite their considerable potential, there are still material-, device-, architecture-, and algorithm-level challenges impeding their commercialization. Compared to the development of AI software, learning-based AI hardware is lagging far behind. Thus far, researchers on memristor materials have found it difficult to manipulate reliable multi-level states, where the memristor unit can be sufficiently distinguished. Therefore, the commercialization of memristors for AI will be accelerated through the development of high-performance and reliable materials, the standardization of synaptic properties, such as conventional CMOS technologies, and road mapping of related technologies. Recently, memristor-CMOS hybrid AI hardware has yielded good performance ( Thakur et al., 2018 ). These hybrid computing systems are considered to be effective for boosting the development of neuromorphic computing systems because CMOS technology has been well developed. Besides developing memristor devices based on a specific material system, hybridizing various memristive materials on a system can be an alternative path to actualizing earlier commercialization. Furthermore, an architecture that completely controls the sneak current in the CBA structure is required. Large-scale size studies and the existing computing architectures remain hindered. In addition, neuroscience needs to clarify the learning mechanism of the human brain to develop high-performance neural network algorithms. In summary, significant research has been conducted to address these limitations, and the memristor technology, which is a building block of neuromorphic computing systems, is expected to give rise to new computing systems. In the near future, we hope that computers inspired by the human brain will outperform the human brain.",
"introduction": "Introduction Unlike other living things, the greatest reason for mankind's prosperity is not physical strength but a highly developed intelligence. Human beings are thought to have the highest level of intelligence among intellectual creatures; this is presumed to be the result of having the most sophisticated brain. Based on these perceptions, human beings have been ultimately pursuing an understanding of the brain and intelligence. Since Turing pioneered artificial intelligence (AI), researchers have attempted to invent machines that can “learn” at a similar level as humans so that they can deal with the complex problems they faced. Consequently, the first programmable computers were invented beyond basic arithmetic operations. Turing suggested a conceptual machine that can solve any calculation using only proper rules ( Turing, 1936 ). This conceptual architecture inspired the design of the von Neumann architecture, the foundation of modern computing architecture. The von Neumann architecture is a structure in which the processor and memory regions are physically separated, and data are transferred via bridged buses. Owing to the drastic advancement in complementary metal-oxide-semiconductor (CMOS) technology, computer performance has improved year by year, following Moore's law ( Moore, 1965 ) and Dennard scaling ( Dennard et al., 1974 ). Based on advances in computing performance and massive data explosions, recently developed AI algorithms based on deep learning have exhibited outstanding performance, especially in recognition problems. However, big data produced by numerous edge devices and sensors require smaller, faster, and low power-consuming computing performance, which has become a challenge with conventional CMOS-based computing systems. In addition, the high (energy and speed) cost of the von Neumann architecture has been touted as a problem, which is well known as the von Neumann bottleneck ( Zidan et al., 2018 ). The latency between the processor and memory is inevitably generated in the von Neumann architecture because it separates the process units from memory units. To address these issues, researchers have returned their focus to the human brain and proposed brain-inspired neuromorphic computing. Although the brain's learning mechanisms are not fully understood, the consistent development of neurobiology has revealed the operation principles of the brain, neurons, and synapses. Through a neurotransmitter, the synapses process and store information in a way that strengthens or weakens the connection between pre- and post-synaptic neurons, depending on the degree of involvement between neurons. Compared to the conventional CMOS-based computing that saves Boolean data type (0 or 1) in a single unit, the brain's neural networks save multiple states in a single synapse by adjusting the “synaptic weight,” which is known as the synaptic plasticity. To apply these features to the new computing system, one memory cell needs to gradually change the information ( Burr et al., 2017 ). The “memristor” (portmanteau of “memory” and “resistor”) technology, proposed by Chua ( Chua, 1971 ), has been regarded as an emerging technology for realizing neuromorphic computing systems. Compared with the CMOS, the memristor exhibits an I-V hysteresis ( Chua, 2014 ), which makes it possible to change conductance states gradually, similar to the biological synaptic weight. The memristor is also a nonvolatile memory technology that reads the nonvolatile resistive states, instead of the volatile capacitance states ( Vourkas and Sirakoulis, 2015 ). In addition, these devices can be integrated into a large-scale crossbar array (CBA) architecture, where memristive devices are sandwiched between the crossed upper and lower electrode lines ( Xia and Yang, 2019 ). Therefore, the theoretical density can be reduced to 4F 2 (F is the feature size) and even higher with 3D structure ( Lin et al., 2020 ), which is expected to overcome the scaling limit of the CMOS technology (typically exceeds 6F 2 ). In this review, we first introduce the six types of memristors (ionic migration, phase change, spin-based, ferroelectric, intercalation, and ionic gating devices) competing to be the representative next-generation neuromorphic device. We describe the operation principle and mechanism at the material nanoscale and illustrate how memristors act as artificial synapses in neuromorphic computing systems, similar to the human nervous system. Furthermore, we explore studies in which memristive neuromorphic computing has performed AI functions such as pattern recognition, classification, and even creation. Finally, perspectives and challenges in true brain-inspired computing are discussed."
} | 2,136 |
33305176 | PMC7718163 | pmc | 14 | {
"abstract": "Summary Memristive devices share remarkable similarities to biological synapses, dendrites, and neurons at both the physical mechanism level and unit functionality level, making the memristive approach to neuromorphic computing a promising technology for future artificial intelligence. However, these similarities do not directly transfer to the success of efficient computation without device and algorithm co-designs and optimizations. Contemporary deep learning algorithms demand the memristive artificial synapses to ideally possess analog weighting and linear weight-update behavior, requiring substantial device-level and circuit-level optimization. Such co-design and optimization have been the main focus of memristive neuromorphic engineering, which often abandons the “non-ideal” behaviors of memristive devices, although many of them resemble what have been observed in biological components. Novel brain-inspired algorithms are being proposed to utilize such behaviors as unique features to further enhance the efficiency and intelligence of neuromorphic computing, which calls for collaborations among electrical engineers, computing scientists, and neuroscientists.",
"introduction": "Introduction Artificial intelligence (AI) has made great progress in recent years with the help of the advances of deep learning (DL) technologies ( LeCun et al. 2015 ). However, owing to the high volume of data needed to be frequently transferred between processing units and memories, the performance of deep learning algorithms is limited by the von Neumann bottleneck in conventional computers. The existing von Neumann bottleneck can be overcome by in-memory computing with memristive devices, where the computation takes place in the analog domain in the memory, right at the data location. Memristive devices, existing in several forms, such as resistive switching random access memory (RRAM), phase-change memory (PCM), magnetic random access memory (MRAM), and ferroelectric random access memory (FeRAM) ( Figure 1 C), have tunable conductance states, similar to the plasticity of biological synapses ( Ielmini and Wong, 2018 ; Wang et al., 2020c ), and thus can enable in-memory computing, in analogy to the biological neural system. Thanks to their scalability, stacking-ability, simple device structure, and other intriguing properties, the memristive devices have been considered as leading candidates for synaptic devices for hardware implementation of neural networks and machine learning, providing an energy-efficient and low-latency solution for future AI ( Yang et al., 2013 ; Ielmini and Wong, 2018 ). Figure 1 Integration of Learning Algorithms and Memristive Devices for Memristive Neuromorphic Computing (A) Various algorithms of deep learning (DL) neural networks, including simple perceptron, deep (multiplayer) neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), and restricted Boltzmann machine (RBM). (B) Adaptions and performance enhancements of memristive devices for their synaptic application in DL algorithms: linear read of the memristive device, realization of both positive and negative weight via differential pair, mapping vector matrix multiplication and accumulation (VMMA) to 2D and 3D memristive array, and linear conductance update for identical pulses. Reproduced from ( Lin et al., 2020 ), copyright © 2020, Springer Nature. (C) Various memristive devices, including resistive random-access memory (RRAM), phase-change memory (PCM), spin-torque transfer random access memory (STT-RAM), and ferroelectric random access memory (FeRAM). (D) Illustration of a memristive neuromorphic computing system integrated within a monolithic chip. (E) Unique memristive device features for emerging algorithms, such as the capability of spike-timing dependent plasticity (STDP), nonlinearity for filtering dendrites, integration and firing functions for memristive neurons, and stochasticity. (F) Emerging algorithms and architectures can be implemented by memristive devices, including spiking neural network (SNN), reservoir computing, Hopfield neural network (Hopfield NN), cellular neural network (Cellular NN), and a probabilistic bit (p-bit). With their ionic transport mechanisms similar to the molecular activities in biological intelligent systems, the memristive devices also exhibit rich dynamics, resembling synaptic and neuronal dynamics found in biological systems. For instance, stochastic switching ( Mizrahi et al., 2018 ), pulse-pair facilitation ( Wu et al., 2018 ), and short-term plasticity ( Ohno et al., 2011 ) have been observed in memristors and are suitable to reproduce the dynamics of biological synapses. This may eventually lead to computational systems capable of faithfully emulating the information representation and processing in the brain with much-improved energy efficiency and fidelity over the conventional systems. Previous efforts to implement these algorithms that take advantage of memristive devices have met with limited success for two reasons: (1) Current state-of-the-art DL technologies use digitalized values (floating-point in software solutions or integers in CMOS-based hardware accelerators) for connection weights. Representing these weights using memristive conductance has suffered from non-ideal behaviors. (2) There are currently no complete algorithms to exploit the bio-plausible behaviors of the memristive devices. As a result, there exist substantial mismatches between algorithms and device properties, and therefore the integration of learning algorithms and memristive devices via co-design is imperative. Specifically, mismatches between memristive devices and learning algorithms can be addressed not only from the device side to engineer the materials for “expected” properties, which has been pursued intensively so far, but also from the algorithm side to either compensate the nonidealities of the devices or exploit some of the “unexpected” properties as valuable features for new types of computing, which has been relatively less explored. In this perspective, recent achievements in the co-design of memristive devices and learning algorithms are reviewed, to provide a comprehensive overview of the status and remaining challenges for future explorations. We first give an overview of the co-design efforts of learning algorithms and memristive devices (section “ The Integration of Learning Algorithms and Memristive Devices ”). We then review such co-design efforts in depth from three standpoints: (1) compensating the device nonidealities at the single synapse level and memristive array level in vector-matrix multiplication and accumulation (VMMA) (section “ DL Accelerators by Memristive Hardware “); (2) exploiting unique memristive device features in various bio-inspired learning algorithms (section “ Exploiting Memristive Properties for Brain-Inspired Algorithms ”); and (3) constructing brain-like computing systems with the bio-inspired algorithms enabled by unique memristive features (section “ Toward a Brain-like Computing System ”). The Integration of Learning Algorithms and Memristive Devices The artificial neural network (ANN) has a lot of variants, such as the simple perceptron, deep neural network (DNN), deep convolutional neural network (CNN), recurrent neural network (RNN), and deep belief neural network based on restricted Boltzmann machines (RBMs) ( Figure 1 A), all belonging to the large family of DL techniques ( LeCun et al., 2015 ). The operation of the ANN can be divided into two stages, inference and learning. During the inference phase, each layer of the neural network transforms the input signals by multiplication with the synaptic weights, summation at each output neuron, and activation according to a non-linear function. During the learning phase, the network is trained with data to adjust the synaptic weights for correct inference. Most of the learning tasks can be technically divided into three categories, supervised learning, unsupervised learning, and reinforcement learning, depending on the type and availability of feedback. In most approaches, an objective function or loss function is defined for weight training, that is, how good (or bad) the current weight configuration is, to fulfill the application (e.g., classification or decision) requirement, and the goal of training is to minimize the loss function or maximize the reward. The most successful learning rule so far is error backpropagation (BP), where the loss function (or error) in the last layer is back-propagated to the preceding layers via the synaptic networks ( Rumelhart et al., 1986 ). BP solves the credit (or blame) assignment problem, i.e., the weight updates to decrease the error, by calculating the gradients of the objective function with respect to the network parameters. In the DL networks of Figure 1 A, the massive synaptic connections and matrix-vector multiplications can be implemented physically within the memristive devices as illustrated in Figure 1 B. Memristive devices perform the in-memory computation of the synaptic weighting functions in both forward inference and backward error propagation. Thanks to their ability to nanoscale miniaturization ( Figure 1 C), hardware acceleration with high efficiency for DL techniques can be achieved ( Figure 1 D). However, several technical issues still need to be addressed to commercialize such memristive DL accelerators at large scales. Spiking neural networks (SNNs) take inspirations from neuroscience ( Maass 1997 ) ( Figure 1 F), by preserving a more biological behavior of the spiking neuron, which can emit spikes in response to a spiking input stimulation. Spike-timing dependent plasticity (STDP) ( Bi and Poo 1998 ) is a widely discussed synaptic property that has been realized by memristive devices and extensively reported ( Figure 1 E). A major limitation of the local learning rules such as STDP is that they can only be applied to shallow networks with only one layer. Recent research works aim at pushing the network deeper by introducing backpropagation approximation in spike domain ( Shrestha et al., 2019 ), using synthetic local gradient ( Kaiser et al., 2020 ) and adopting backward residual connections and stochastic SoftMax functions ( Panda et al., 2020 ). Artificial spiking neurons can mimic closely the biological neurons' behaviors, for example, reproducing the exact waveform of the membrane potential in different phases of a fire. However, these bio-realistic models are rarely used in software-simulated SNNs owing to a heavier computational burden. More typically, spiking neurons in SNNs are designed to have a binary response, i.e., “1” when the neuron receives or generates a spike or “0” when it does not. Memristive devices with non-linearity, state-variable accumulation (integration), volatility, as well as stochasticity ( Figure 1 E) can be used to construct spiking neurons and dendrites, thus combining bio-realistic responses with low computational burdens via the physical properties of the memristor. There are various methods to encode the information into the spikes, such as using the frequency of a train of spikes (rate coding) or the precise arriving time of each spike (temporal coding) ( Masquelier et al., 2009 ). Different encoding methods running on different neural networks require different training rules, which, however, are usually more computationally expensive than ANNs. Temporal-encoded SNNs are more efficient than rate-encoded SNNs, as the information can be contained in just one spike. Owing to the time-related nature, temporal-coded SNNs are more suitable to process time-related data, such as speech, sound, and vision. Memristive devices with state-variable accumulation, or short-term volatility, ( Figure 1 E) can be used to process these temporal-encoded spiking patterns in reservoir computing networks ( Moon et al., 2019 ) ( Figure 1 F). Other advanced methods have been proposed to employ a pseudo-gradient to overcome the non-differentiability of a spiking neuron in feedforward networks. For instance, the e-prop training method in recurrent SNN ( Bellec et al., 2020 ) has been proposed to approximate the backpropagation through time (BPTT) algorithm for traditional RNNs ( Werbos 1988 ), which removes biologically unrealistic computational requirements and makes it possible to build on-chip hardware learning units. Besides the SNNs as a signal morphological approach, another approach for brain-like computation is the collective-state computation that emulates brain-like computation at a high level ( Csaba and Porod 2020 ). This computational model originated from the Hopfield neural network ( Hopfield 1984 ), then extended to cellular neural networks ( Chua and Yang 1988 ), coupled oscillators ( Ignatov et al., 2017 ), and adiabatic annealing machines with probabilistic bits ( Borders et al., 2019 ) ( Figure 1 F). In these networks, nonlinear synaptic and neural behaviors as well as probabilistic behaviors are needed, which can be provided by the memristive devices ( Figure 1 E). Similarities between memristive devices and biological components, such as synapses, dendrites, and neurons, have been extensively reported. Some fundamental functionalities or brain-inspired algorithms have also been demonstrated. However, practical computing systems based on these algorithms have rarely been reported. DL Accelerators by Memristive Hardware Memristive synapses can be optimized to accelerate DL algorithms. Assembled into a crossbar array configuration, memristive devices are inherently suitable for efficient VMMA operations, which account for a majority percentage of the computation in artificial neural networks, by directly using Ohm's law for multiplications and Kirchhoff's current law for accumulations. Both passive ( Prezioso et al., 2015 ) and active crossbar arrays ( Li et al., 2018a ) have been reported for VMMA. Three-dimensional (3D) stacking of crossbars provides an additional dimension of parallelism, connectivity, and efficiency for complex neural networks ( Lin et al., 2020 ). The DL neural network has many variants and components, for instance, a fully connected layer, convolutional layer, and recurrent neural network. The mapping between the memristive array and computing layers ( Yao et al., 2020 ) is necessary to achieve the designed topology and is crucial for realistic applications. DL accelerators require both linear current-voltage (I-V) relation and linear weight update characteristics of the memristive synapses. Optimization of material stacks and electrical operation protocols can ensure very stable and linear I-V characteristics for voltage-conductance multiplication ( Hu et al., 2018a ). The differential pair (G+/G-) method ( Suri et al., 2011 ) has been widely used in both RRAM and PCM synapses, to allow for both positive and negative synaptic weights using electrical conductance, and more importantly to compensate for any asymmetry between the two weight update directions. Other methods are also used to improve the linearity of synaptic weight updates, for instance, multi-parallel devices for mapping the least significant and most significant weight components in a single synapse ( Ambrogio et al., 2018 ). So far, memristive DL accelerators have experimentally achieved reasonable accuracy in relatively small-scale tasks such as recognition of the image in the Modified National Institute of Standards and Technology (MNIST) and Canadian Institute For Advanced Research (CIFAR) datasets ( Yao et al., 2020 ; Ambrogio et al., 2018 ). Such demonstrations remain to be seen in larger-scale DL networks like ResNet and other network structures for ImageNet ( He et al., 2016 ). Optimizing Linear Response of the Memristive Device In the inference stage of the neural network, the synapse is essentially carrying out the weight function, that is, scaling the input signal into an output signal. This can be realized by the memristive device through implementing the Ohm's law: I = G·V , where G is the conductance of the memristive device, V is the input signal represented by a voltage, and I is the output signal represented by the current through the memristive device. However, memristive devices do not always follow Ohm's law with a linear I-V relation. Since the electron transport in a low conductance state often involves mechanisms like tunneling or hopping transport ( Ielmini and Zhang 2007 ), it is common to observe nonlinear (e.g., exponential) relation between current and voltage ( Figure 2 A), which compromises the accuracy of directly using Ohm's law for multiplication. This issue is more severe in a passive memristive array where a nonlinear I-V characteristic is deliberately adopted to mitigate the difficulty in selectively programming a target device in the passive array ( Alibart et al. 2013 ). Figure 2 Material- and Device-Level Optimizations to Compensate for the Nonidealities of Memristive Synapses (A) Dealing with non-linear current-voltage (I-V) relation in memristive devices. Left: a typical I-V relation showing exponential dependence of the reading current on applied voltages. Upper right: material modification for linear I-V relation by using composition modulation of localized conduction channels. Reproduced from ( Jiang et al., 2016 ), CC BY. Lower right: pulse number modulation with the number of pulses representing the strength of the input signal and accumulated charge representing the weighted output, resulting in linear behavior between the charge and the pulse number. (B) Differential pair for implementing both positive and negative synaptic weights. (C) Learning errors as the result of the nonlinearity of the weight update. For gradual conductance increases of both G + and G - , potentiation is conducted by applying pulses to the G + part (a → b) and depression is conducted by applying pulses to the G - part (b → c). The same numbers of potentiation and depression should cancel each other; however, it results in different overall weight (G + -G - ) between (a) and (c). (D) An array of memristive devices arranged in crossbar structure for VMMA. The voltage vector was applied to the row bars of the memristive array and the current output in each column bar is sensed as the result of Ohm's law and Kirchhoff's current law, I j = ∑ i G i j V i . (E) and (F) Illustration of each memristive device in series connected with a passive selective device or a transistor, respectively, for access without affected by or affecting other devices. (G) Mapping 2D convolutional kernels to memristive array for convolutional layers. Reproduced from ( Yao et al., 2020 ), copyright © 2020, Springer Nature. The linearity of the I-V characteristic can be improved by optimizing the stacks of the memristive device thus the switching dynamics rely on the composition modulation of a localized conduction channel rather than a tunnel barrier ( Jiang et al., 2016 ). A smaller read voltage is preferred for linear read operation since even an exponential I-V relation can be approximated by a linear behavior at small voltages, which, however, may decrease the signal/noise ratio of the circuits and potentially degrade the accuracy of the network. An alternative way to bypass the nonlinearity of the read operation (for inference) is to use a fixed read voltage while the input signal strength is represented by the duration of the read. The output signal will be the accumulated charge of the output current. The duration can be the width of a single read pulse ( Cai et al., 2019 ) or digitalized into the number of identical pulses, namely, pulse number modulation ( Yao et al., 2017 ; Cai et al., 2019 ). This technique may sacrifice the inference speed. In addition, charge accumulation of the output current requires capacitor-based circuits or additional processes in the digital domain. On the other hand, using the number of identical pulses to represent input signals may simplify or even eliminate the need for analog/digital conversion circuits. As previously mentioned, the mainstream algorithms for artificial neural networks are based on gradient descent of the synaptic weight to minimize the difference between the actual and the target outputs. This results in both negative and positive values of the synaptic weights. The negative synaptic weight also has its biological basis of inhibitory connections in the biological neural system. A practical way to realize negative synaptic weight is to use a differential conductive pair as shown in I =( G + − G − ) ·V , where G + and G − are the conductances of the positive and negative branches of the differential pair, respectively. The minus sign can be easily implemented by Kirchhoff's current law, as shown in Figure 2 B. Although this reduces the synaptic density by two times, it has additional benefits such as a better redundancy for precise representation of the synaptic weight. For example, if one of the memristive device in the pair is stuck to a certain conductance and becomes non-responsive, the pair can still be used to represent any synaptic weight by just programming the other devices in the pair to an appropriate value. Negative synaptic weights can also be realized by subtracting one shared “negative” column (reference) from all “positive” columns ( Milo et al., 2019 ). Compensating Device Non-ideality by Online Learning The synaptic weight can be learned offline by simulating the artificial neural network in traditional computing systems and then transferring the resulting weights to the memristive differential pairs in memristor arrays. This type of weight learning, although consuming a large amount of computational resources on the offline learning platform, is only a one-time cost. To minimize the write errors, it is necessary to utilize iterated program-and-verify steps until the synaptic weight falls into the target range. However, owing to the limited conductance levels and variations of the devices, discrepancies between the original weights and the transferred ones are inevitable. Device failures would also largely degrade network performance ( Li et al., 2018 ). One solution to compensate for the inaccurate program and device failures would be on-line learning or in situ learning. To minimize the output error, the synaptic weight needs to be adjusted according to the gradient descent rule, which can be denoted by Δ w = η·x·δ , where Δ w is the weight changes, η is the learning rate, x is the input value, and δ is the output error caused by this synaptic connection (defined as the partial derivative of the total output error to this node's output, and can be backpropagated from the last layer to all nodes in the network). The output error can be obtained by comparison between the actual output of the memristive network and target output. Then periodical weight update can be performed to the memristive synapses via repeated presentations of input patterns from the dataset. With the output error calculation performed in situ in the memristive network, the weight learning partially adapts to the nonidealities of memristive devices (write noise and device failure), thus mitigating the impact of the memristive nonidealities ( Li et al., 2018 ). A similar approach can also be applied to offline training where the weight update in network simulation takes the nonidealities of the memristive devices into consideration, namely, hardware aware training ( Gokmen et al. 2019 ; Joshi et al., 2020 ). Furthermore, a recently proposed hybrid training scheme takes the advantages of both in situ and ex situ training methods to efficiently realize a memristor-based neuromorphic system ( Yao et al., 2020 ). The weights are initially learned and transferred to memristor conductance as in conventional ex situ training case regardless of device non-ideal characteristics, and only a small part of the weights is trained in situ (e.g., the weights of the last fully connected layer) to adapt to present imperfections and recover the system performance. All these realizations of online learning are clear examples of the strength of the device/algorithm co-design, where device nonidealities are mitigated via system-level algorithm optimization. Realizing Linear Weight-Update Neural network training is a computationally intensive task. Although it can be a one-time cost for some applications, other applications that involve continuous learning or adaption require more frequent re-training of the memristive neural networks. The iterated backpropagation steps in the online training could be a major barrier for training acceleration with memristive networks. It would be ideal if only one write pulse was needed to adjust the memristive synaptic weight in each learning epoch without the need of knowing the current state of the memristive synapse or verifying the updated state. However, this cannot be achieved in most of the memristive devices because their set (increasing the conductance) and reset (decreasing the conductance) switching processes are nonlinear and asymmetric. A gradual (analog) conductance change upon write pulses during switching can only be achieved in either set or reset operation. RRAM devices can only be gradually reset while the set is abrupt. In contrast, PCM devices can only be gradually set while the reset process is abrupt. As a result, the devices for differential pair composed of RRAM devices are prepared in relatively high conductance states to start with. If potentiation is needed, then the negative conductance branch G − will receive reset operating pulses, whereas for depression, reset pulses will be applied to the positive branch G + . The opposite configuration can be used for a PCM differential pair. However, in both cases, the amount of the conductance change per write pulse is not constant, rather it depends on the current conductance state of the devices, thus leading to substantially non-linear weight update. Saturation of conductance change in the low-conductance range of RRAM device during the reset operation and in the high-conductance range of PCM during the set operation is generally observed ( Figure 2 C). During training, numerous weight updates in potentiation and depression should mostly cancel each other and result in only a small amount of net conductance change. The nonlinear changing of conductance in G − and G + , however, prevents such cancellation in the weight updating process and acts as the largest source of accuracy loss in online training ( Burr et al., 2015 ). To linearize the conductance change, material- and device-level optimization has been performed. For instance, by inserting an AlO x barrier layer, a HfO 2 -based RRAM device shows an improved linear potentiation (or depression) behavior of conductance under identical pulses ( Woo et al., 2016 ). Choi et al. limited the conductance changes in a smaller window in Ag + -based memristive devices by modifying the potentiation pulses to only allow a mild amount of Ag participant in the switching dynamics ( Choi et al., 2018 ). A three-terminal memristive device with an electrochemical gate layer can also make the change of source-drain conductance linear with the pulses applied in the gate terminal ( Tang et al., 2018 ; van de Burgt et al., 2017 ). These electro-chemical memristive devices, however, suffer from long pulse duration hence a low operational speed. The gradual linear conductance tuning in both directions also allows the potentiation and depression in both sides of the differential pair, avoiding freeze-out of the conductance tuning in unidirectional update ( Ielmini and Pedretti 2020 ). Capacitor-based synapse has an ideal linear weight update ability but is volatile. A synaptic configuration composed by a major pair of non-volatile PCM devices (G+, G-) and a minor pair of volatile capacitor-based synapses (g+, g-) can thus integrate the advantages of these two parts and lead to linear weight update and non-volatile synaptic weight storage. The minor pair updates its weight frequently in each training cycle and transfers the weight to the major pair periodically. Thanks to the linear weight update ability, a training accuracy comparable with that of a software-based solution can be obtained ( Ambrogio et al., 2018 ). However, this solution results in a bulky individual synapse (2PCM + 3T1C), which occupies a relatively large area on the silicon chip. Utilizing Memristive Arrays for VMMA Mapping the fully connected layers of artificial neural networks to the two-dimensional (2D) memristive arrays is straightforward since the computation in the fully connected layers are essentially VMMA operations ( Ielmini and Pedretti 2020 ), y j = ∑ i w i j x i , where i and j denote the indices of the input vector and output vector, respectively. It can be directly implemented by applying voltages to the rows of a memristive array and sensing the current output at the end of each column as the result of Ohm's law and Kirchhoff's current law, as shown in Figure 2 D. Note that the case shown in Figure 2 D illustrates the ideal case of using a passive memristive array ( Prezioso et al., 2015 ). However, there is a dilemma in a passive array between the desire for a linear I-V relation to directly utilize Ohm's law during inference and the desire for a nonlinear I-V relation to suppress the sneak path currents and program individual devices during learning. Adopting a selector in each cell to form the so-called one-selector/one-resistor (1S1R) structure ( Figure 2 E) increases the nonlinearity for array programming, which, however, makes the pulse duration/number modulation scheme discussed before an only suitable choice for the inference. A more practical implementation of VMMA is to use an active array where each cell of the array is composed of a memristive device and a transistor connected in series ( Figure 2 F). With such transistors in the array, an individual memristor can be selected by activating only the corresponding transistor(s) for reading or programming operations without affecting other devices in the array. With a high on/off ratio of the mature transistor technology, a large memristive crossbar array can be achieved, with the only limitation essentially being the wire resistance. Fully parallel reading during inference can be enabled by turning on all the transistors in the array, and semi-parallel programming (e.g., row by row) is also possible. The cell size is limited by the transistor footprint, which results in a lower synaptic density than the passive crossbar array. The overall area efficiency has been limited by the peripheral circuits for neuronal functions (section “ Co-design with CMOS Peripheral Circuits ”) in most hardware implementations of memristive deep networks so far. When innovative designs are devised to significantly reduce the area of such peripheral circuits, the 1T1R synaptic arrays limited by the size of the transistor will then become the primary concern of the area efficiency. The parasitic effects existing in the array also limit the size of the array. For instance, when the wire resistance is comparable with the device resistance, the voltage drops on the wire connections lower the real voltage applied to the memristive device and might result in program or read errors, i.e., IR drop issues. The actual limit of the size of an array is affected by the wire resistance, the on resistances of the devices, and the on/off ratios. Assuming the memristive device has resistances of 100 kΩ and 10 MΩ in the low-resistance state and the high-resistance state, respectively, and taking the Cu wire interconnection resistance parameter from the ITRS roadmap, Zuloaga et al. predicted that array size up to 256 × 256 can work well after being downscaled to 10 nm technology node ( Zuloaga et al., 2015 ), which is sufficient for most of the demonstrations of deep networks reported so far. If a larger array size for VMMA operation is needed, material-level optimizations to decrease the resistance of the connection wires or to increase the resistance of the memristive devices are needed ( Zhang et al., 2019 ). Alternatively, system reduction schemes that can substantially reduce the weight matrix size can be utilized ( Liu et al., 2014 ). It is also possible to mitigate the IR drop by a compensation scheme retraining the dominant neurons ( Jeong et al., 2018 ). Online adaptive learning ( Li et al., 2018 ), where the automatic weight update will adapt to the mismatch of the actual cell resistance/conductance with the desired synaptic weight, can also mitigate the parasitic effects in array configurations. Convolutional neural networks are more suitable for recognizing static 2D inputs, such as images, thanks to the biologically inspired convolutional kernels similar to the receptive field of complex cells in the visual cortex. Mapping the convolutional layers to memristive arrays needs first converting the convolutional operation of matrix-matrix multiplication of each kernel and its receptive field to a vector-vector multiplication ( Gao et al., 2016 ). The kernel vectors are aligned in columns or rows of the memristive array, and image input is partitioned in receptive fields (patches) and sequentially fed to the kernel vectors, as illustrated in Figure 2 G ( Gokmen et al., 2017 ; Yao et al., 2020 ). This weight sharing methods by the convolutional kernels is also named as spatial weight sharing ( Wang et al., 2020c ). Spatially shared memristive weight strongly reduces the number of memristive devices needed for a deep CNN. For comparable accuracy of recognizing handwritten digits from MNIST database, the three-layer fully connected neural network consumes 329,770 PCM devices (accuracy 97.94% [ Ambrogio et al., 2018 ]), whereas the five-layer convolutional neural network uses 5,629 RRAM devices (accuracy 96.19% [ Yao et al., 2020 ]). Spatially shared kernel vector weight, however, requires repeated vector-vector multiplication between input image patches and the convolutional kernel. The number of patches increases quadratically with the image size ( Gokmen et al., 2017 ), which reduces the throughput and becomes the bottleneck of the system performance. It is also possible to parallelly implement the convolutional operation in fully connected topology ( Lin et al., 2020 ) or partially reduce the spatial share factor. Replicating and transferring identical convolutional kernels to multiple 2D memristor arrays provides a possible solution to boost the parallelism of convolution operations and enhance the throughput accordingly ( Yao et al., 2020 ). However, a careful trade-off between the inference speed and memristive area cost is necessary. Memristive devices are suitable for 3D integration, a powerful solution for high-density storage, and most importantly, for enhancing the neuronal connectivity required in complex neural networks. The 3D integration enables parallel and faster convolutional calculation for neuromorphic application ( Huo et al., 2020 ; Lin et al., 2020 ), bypassing the trade-off between the inference speed and memristive area. As the cross-section of the 3D array is a 2D interface to directly accept 2D image inputs, the image patches and convolutional kernel matrices no longer need to be unrolled to vectors ( Lin et al., 2020 ). Moreover, the additional dimension could enable massive connections and increase the flexibility of memristor topologies. By defining the sliding kernels at each patch as zigzag staircases in 3D space and shaping both the top electrode and bottom electrode as vertical pillars, input signal could be fed into different patches through the same pillar simultaneously. In this manner, the whole image could be presented at the same time and the 3D device structure enables all convolutional operations during the sliding procedure to be processed in parallel, saving the sequential shifting time and improving the system performance. Additionally, unrolling 3D convolutional kernels to 2D matrices, a 3D convolutional neural network for stereoscopic object recognition can be realized ( Huo et al., 2020 ). Other variants of the DL neural network, like long short-term memory (LSTM) ( Li et al., 2019 ; Wang et al., 2019c ) and deep belief neural network composed by RBMs ( Eryilmaz et al., 2016 ), can also exploit the benefits of VMMA capability of memristive arrays, which is very attractive as VMMA consists the major part of the computations in these networks. However, other essential functionalities, like the gate unit controlling the memory time for LSTM cells and probability generations in RBM neurons, are carried out in software. These functionalities can also be implemented by exploiting the unique features of memristive devices more morphologically, which will be covered in the next section. Co-design with CMOS Peripheral Circuits The program and read operations on the memristive array and the weight update calculation need to be carried out by complementary metal–oxide–semiconductor (CMOS) circuits mimicking the behavior of biological neurons, which should be integrated closely with the memristor array on the same chip to further enhance the efficiency. Tailored according to the targeted application and specific hardware architecture, the generally utilized circuitry blocks could include the sample-and-hold (S&H) module and analog-to-digital converter (ADC) to temporarily hold the summed analog currents and transform them to the digital domain, respectively. The digital-to-analog converter (DAC) that converts digital inputs into appropriate voltage amplitudes should be counted if a voltage amplitude-encoding scheme is adopted. Digital control, processing, and routing blocks are also necessary to realize activation functions and monolithically integrate a complete neuromorphic system ( Shafiee et al., 2016 ; Hu et al., 2018 ). Furthermore, extra peripheral circuits need to be considered to realize various kinds of on-chip learning rules. It is worth mentioning that in practical system implementation, the bottom-level device characteristics and the top-level algorithm optimizations would jointly determine the circuit and architecture design to meet the necessary hardware performance. The array size, the precision and speed of ADCs, and other circuit aspects need to be carefully considered with trade-offs between hardware efficiency and cost. CMOS implementation of the peripheral circuits can result in a large area and power consumption. For instance, the DAC and ADC circuits can occupy a much larger area (e.g., 21.57 mm 2 ) than the dense synaptic array (e.g., 0.14 mm 2 ) ( Cai et al., 2019 ). A commonly used strategy to improve the system area efficiency is to temporally share the DAC and ADC elements. State-of-the-art high-precision ADCs consume large area; however, they can have high sampling rates. Thus multiple output nodes of the memristive array can share a single ADC sequentially ( Gokmen and Vlasov 2016 ). An additional multiplexer (MUX) circuit is needed for the sequential selection of output nodes. ADCs of 6–8 bit precision are needed for acceptable accuracy loss of neural networks with a relatively small size while further lowering the precision may induce high accuracy deterioration ( Li et al., 2016 ). Nevertheless, the required precision for large-scale neural networks depends on the exact network structure or dataset. Further reducing the precision can be achieved by separately handling the outlier values and normal values with 4-bit ADCs, showing a relatively small loss of performance ( Park et al. 2018 ). A binary neural network is a possible option to address the area and power inefficiency in ADC- and DAC-based neuronal function realizations, which will be further discussed in section “ Exploiting Bistable Behavior of Memristive Devices .” It is also possible to use neurons working directly in the analog domain ( Krestinskaya et al., 2018 ), thus eliminating the need for the conversion between analog and digital signals. This requires the analog neuronal circuit to perform (in the analog domain) activation functions, like sigmoid or ReLU, which, in most of the hardware deep neural network demonstrations so far, have been implemented by controlling computers or microcontrollers in the digital domain after the conversion of the signal from the analog domain. Using integrate&fire neurons to convert the output of VMMA to spiking trains, where the spike count or frequency denotes the analog value of the neuron output, can also eliminate the need of analog to digital conversion ( Yan et al., 2019 ). For a large-scale integration of the synaptic and neuronal components with the learning algorithms in a practical system, utilizing the above strategies, in-memory computing macros or neuromorphic computing macros have been proposed, such as ISAAC ( Shafiee et al., 2016 ), PRIME ( Chi et al., 2016 ), and Pipelayer ( Song et al., 2017 ). These macro circuits can be tiled together according to the structure of deep neural networks to be constructed. For a more comprehensive review of peripherical circuits and large-scale integration, readers can refer to Yan et al., 2019 . These memristive DL accelerators are projected to be superior to CMOS-based or other solutions in several aspects, such as in performance (operation per second, OPS), area, and power efficiency ( Zhang et al., 2020a ; Sebastian et al., 2020 ). Unfortunately, ideal switching characteristics and linear I-V characteristic are assumed in these designs. In terms of the chip-level demonstration, a fully integrated memristive in-memory computing macro, named as non-volatile computing-in-memory (nvCIM), has been demonstrated ( Chen et al., 2019 ). However, the nvCIM macro works on the binary-input ternary-weight model and will not fully exploit the analog in-memory computing ability of a memristive array. The realization of activation and intra-layer communication is carried out by off-chip field-programmed gate array (FPGA). Recently, this field has been rapidly developing toward monolithically integrated memristive neuromorphic systems, even though the memristive analog behavior has not been fully exploited ( Liu et al., 2020 ; Wan et al., 2020 ). In Table 1 , we summarize the optimization and design efforts in various levels of memristive neuromorphic computing with the purpose of DL accelerators. The color code refers to the degree of optimization/co-design of each implementation. Despite the numerous efforts in Table 1 , a general-purpose memristor accelerator for general neural networks is still missing, partially because device reliability and uniformity issues across multiple arrays are yet to be solved. On the other hand, conventional DAC and ADC solutions consume a large area and energy, becoming the bottleneck of system performance ( Cai et al., 2019 ). Novel routing schemes with the least requirement for the on-chip memory are also needed to make the most of the memristor neuromorphic system. Table 1 A Survey of Optimization and Design Considerations in Various Levels of Memristive Neuromorphic Computing NA: not applicable (or not discussed). a Training within a transfer interval was performed in software with device models and read PCM devices are operated when transfer needed. b Data retrieved from ( Li et al., 2018 ). c Projected for all peripheral circuits integrated on chips with 128 × 128 array and four columns of the memristive array sharing one ADC converter. d ML-CSA: multi-level current-mode sense amplifier; DR-CSA: distance racing current-mode sense amplifier. Exploiting Memristive Properties for Brain-Inspired Algorithms In addition to serving as a static memory of synaptic states for in-memory-computing in deep learning algorithms, memristive devices also have a variety of dynamical properties that share close similarities with biological components, which can potentially lead to computing with augmented efficiency and intelligence. Novel brain-inspired learning algorithms are needed to utilize these intrinsic properties of memristive devices, e.g., the stochasticity of the state ( Yu et al., 2013 ), the dynamics of state transition, and second-order effects ( Du et al., 2017 ). Exploiting the Stochasticity Various random physical phenomena exist in memristive devices, resulting in stochastic variations of conductance levels and switching parameters. For instance, for RRAM and PCM devices, owing to the nature of the ionic-electronic coupled transport mechanism, intrinsic stochasticity exists as random telegraph noise in the reading phase and variation of switching parameters in the weight update phase ( Carboni and Ielmini 2019 ). Stochasticity is a critical problem for memory and storage applications and their usage as synaptic weights in DL accelerators. However, stochasticity as a physical entropy source can be exploited for generating true random numbers or physical unclonable functions for information security. Additionally, they can provide a low-cost solution for implementing some specific neural network algorithms where stochasticity is essential for computation. Under a weak programming condition, the set transition of metal-oxide memristive devices becomes probabilistic. A winner-take-all network can be realized by the competition among post-neurons utilizing the probabilistic switching in synapses ( Yu et al., 2013 ), as illustrated in Figure 3 A. In RBM, the sampling and reconstruction stages heavily rely on the probabilistic of hidden or visible units being activated. A dot product circuit incorporating the stochasticity coming from the intrinsic noise of the memristor array for the RBM has also been demonstrated ( Mahmoodi et al., 2019 ). Using a passive memristor crossbar, a single-layer RBM with ten visible and eight hidden neurons is demonstrated with the energy function minimization ( Figure 3 B). In another work of implementing the Hopfield neural network, the intrinsic noise of a memristive crossbar was used for a combinatorial optimization problem ( Cai et al., 2020 ). A moderate noise level was found useful for the network to escape from local minimum points in the energy landscape better than both the noise-free and the high noise level situations ( Figure 3 C). Figure 3 Various Bioinspired Algorithms Exploiting the Unique Features of Memristive Devices (A) A winner-take-all neural network exploiting the stochasticity of the memristive devices. Reproduced from ( Yu et al., 2013 ), CC BY. (B) A restricted Boltzmann machine exploiting the intrinsic reading noise of the memristive array. Reproduced from ( Mahmoodi et al., 2019 ), CC BY. (C) Exploiting the intrinsic noise of memristive array to avoid trapping in local minimum sites in a Hopfield memristive neural network. Reproduced from ( Cai et al., 2020 ), copyright © 2020, Springer Nature. (D) Memristive dendrites exploiting the non-linearity of memristive devices for filtering and integration functions. Reproduced from ( Li et al., 2020 ), copyright © 2020, Springer Nature. (E) Reservoir computing realized by the short-memory effect of diffusive memristors. Reproduced from ( Midya et al., 2019 ), CC BY. (F) A PCM device-based neuron with the gradual set switching of the PCM device was used to mimic the integration function of neurons. Reproduced from ( Tuma et al., 2016 ), copyright © 2016, Springer Nature. (G) Spike-timing-dependent plasticity (STDP) enabled by the short-term memory (volatility) of the diffusive memristive device. Reproduced from ( Wang et al., 2017 ), copyright © 2016, Springer Nature. (H) A Mott transition memristor device-based Hodgkin-Huxley neuron faithfully reproducing biological spike shape and tunable spiking trains. Reproduced from ( Pickett et al., 2013 ), copyright © 2013, Springer Nature. The stochasticity of magnetic tunnel junctions (or MRAM) was reported to implement three-terminal probabilistic bits (p-bits) ( Borders et al., 2019 ). These p-bits can be viewed as probabilistic neurons that are electrically connected to form an asynchronous network for factorizing integers up to 945 (63 × 15) adapting diabatic quantum computing algorithm. Tunable probability of the random switching of the superparamagnetic tunnel junction also allows the population coding where each neuron embodied by a superparamagnetic tunnel junction is associated with a specific range of inputs, which is then computed as a weighted sum of the rates of each neuron ( Mizrahi et al., 2018 ). Similar tunable probability can be achieved by utilizing the inherent random noises of analog RRAM devices and was used to construct a Bayesian inference neural network that shows high resilience to adversarial testing samples ( Lin et al., 2019 ). Exploiting the Current-Voltage Non-linearity I-V non-linearity of memristive devices could be a major issue when the devices are used as synapses for DL accelerators. However, such non-linearity is an essential synaptic or neuronal behavior in cellular neural networks ( Duan et al., 2015 ; Caravelli et al. 2017 ), where it enriches the dynamics of the system. It has also been shown that the non-linearity can be used to mimic the non-linear integration of biological dendrites ( Lavzin et al., 2012 ), constructing a memristive dendrite ( Li et al., 2020 ). In conventional algorithms of DL neural networks, neurons act as simple elements summing all inputs from synapses. It is found in neuroscience that the non-linear integration of synaptic signals by the dendrites provides primitive processing before the signals reach the neuron body ( Agmon-Snir et al., 1998 ). Recently, a memristive dendrite component has been demonstrated using a Pt/TaO x /AlO δ /Al-based dynamic memristor ( Li et al., 2020 ), exploiting the non-linearity provided by the Schottky-like barrier in the Pt/TaO x interface. By adding non-linearity before the neural summation to realize the non-linear dendritic function can further enhance the performance of the neural network. With memristive dendrites filtering the background signals, the spiking output of the neuron shows more distinction between false patterns and true pattern ( Figure 3 D), and, at the same time, reduces the power consumption. Thus, performance enhancement in both energy efficiency and accuracy is obtained with the addition of non-linear memristive dendrites. Exploiting the State-Variable Accumulation Upon stimulation of weak electrical pulses, the memristive devices sometimes do not show explicit conductance changes. However, explicit conductance may be induced by subsequent pulses only if there are prior pulses. This behavior happens because of the internal state-variable, for instance, the temperature being an internal memory for historical stimuli ( Kim et al., 2015 ). The accumulation of the state-variable may offer an internal timing mechanism and enables an activity-history-dependent modulation of the first-order state, namely, conductance; thus, is also called second-order effect. It has been used to mimic the effect of the Ca 2+ dynamics of biological synapses and enables the temporal learning of timing-encoded information ( Zidan et al., 2017 ). In Ag-based diffusive memristors, the configuration of the transporting ions in the dielectric layer before the threshold switching shares a similar behavior. The accumulation of ions before their final formation of a continuous filament bridging the electrodes can act as internal memory for historical stimuli. The drift of Ag ions under electrical stimuli and the diffusion of Ag ions under zero electrical bias faithfully emulate the ion dynamics that play a critical role in the neuromorphic functions of biological intelligent systems ( Wang et al., 2017 ). For instance, pair-pulsed facilitation and depression were found in the diffusive memristor-based synapse for a high frequency of pulses and low frequency of pulses, respectively ( Wang et al., 2017 ), which also enables the temporal learning naturally. Reservoir computing (RC) can offer efficient temporal processing of recurrent neural networks with a low training cost. RCs based on the second-order effect ( Du et al., 2017 ; Moon et al., 2019 ) and the accumulation of ions in diffusive memristors ( Midya et al., 2019 ) have been explored. The RC exploiting the accumulation or integration of the state-variable (secondary internal variable or ion configurations) acts as a framework extracting features from temporal inputs. Taking advantage of the rich short-term dynamics of the diffusive memristive device, an RC system is constructed with one reservoir layer of a diffusive memristors and one readout layer of a nonvolatile memristor-based trainable perceptron neural network, with which classification of temporally rearranged handwritten digits from the MNIST database is achieved with a much-reduced training workload, as shown in Figure 3 E ( Midya et al., 2019 ). Without the short-term dynamics in the state-variable accumulation, neural networks for processing temporal information should have additional memory gate to control the learning and forgetting of the historical information, resulting in extra costs in circuitry and energy. The state-variable accumulation upon electrical stimuli can be used for the integration function of an artificial memristive neuron ( Tuma et al., 2016 ). The gradual set of PCM device has been utilized to demonstrate an artificial neuron capable of integrating post-synaptic potential at the nanoscale, where the phase configuration (thus the conductance) of the nanoscale PCM device represents the membrane potential, as shown in Figure 3 F ( Tuma et al., 2016 ). With the gradual internal state change upon pulses mimicking the integrating function and its consequential abrupt switching representing the fire behavior, a silicon oxide RRAM cell is reported to emulate a biological neuron ( Mehonic and Kenyon 2016 ). A similar function can also be realized by the accumulation of ion transport in a diffusive memristor ( Hao et al., 2020 ). This approach results in a capacitor-free version of a solid-state neuron; however, it requires a reset of the memristive device back to its original state after each fire. Exploiting the Volatile Memristive Switching Industrial memory storage application of memristive devices requires that the device can retain its conductance state for at least 10 years; thus, it is also called non-volatile memory. In neuromorphic computing for DL accelerators, a similar requirement should be fulfilled, that is, the memristive conductance for synaptic weight should remain stable for a long time to preserve the learned knowledge. However, some memristive devices with Ag as one of its electrode shows short retention time for the high conductance state ( Bricalli et al., 2018 ; Wang et al., 2019b ). The retention time is usually reported to be in the range of sub-microseconds to tens of milliseconds. This can be viewed as short-term synaptic plasticity and is reported to demonstrate some time-related computing functionalities ( Wang et al., 2018a ). The volatilities can be modulated by the strength of the stimuli. An increase in the frequency of applied pulses ( Ohno et al., 2011 ) or using a higher compliance current ( Wang et al., 2019b ) can cause a transition from volatile to non-volatile memory, corresponding to the short-term plasticity and long-term plasticity, respectively. The short-term plasticity of volatile memristive device allows the STDP learning with non-overlapping spikes to be demonstrated in a combined synapse of one volatile memristive device and one non-volatile memristive device ( Wang et al., 2017 ), with the finite delay time of the volatile memristive device bridging the time gap of the non-overlapping spikes ( Figure 3 G). Another approach for memristive neurons is to utilize the abrupt and volatile switching of memristive devices for the fire functionality ( Wang et al., 2018d ; Zhang et al., 2018 ), whereas the integration function is completed by charge accumulation in an external or parasitic parallel capacitor or internal state accumulation before the abrupt switching ( Zhang et al., 2018b ). The volatility of the memristive device enables the artificial neuron to recover its resting state spontaneously after the abrupt switching on, which is obtained by a device reset operation after each firing event in the nonvolatile memristive neurons discussed in section “ Exploiting the State-Variable Accumulation .” Owing to the simple structure and nanoscale-level scalability, these memristive neurons can be much more compact than the bulky CMOS neurons. Moreover, in case long time constants, such as tens of milliseconds, are needed to match the normal time constants of the biological systems, huge capacitors would be required in the CMOS neurons ( Qiao et al., 2015 ). In contrast, a nanoscale diffusive memristor would readily provide such time constant. The implementation of memristive neurons has also enabled fully memristive neuromorphic computing ( Wang et al., 2018 ), further enhancing the integration level of the hardware neuromorphic computing. Volatile memristive switching sometimes accompanies negative differential resistance arising from an insulating-to-conducting phase transition or Mott transition, namely, Mott memory device ( del Valle et al., 2019 ; Zhang et al., 2020b ). Using two Mott memristors with transient memory as ionic channel and two capacitors as charge storage, a neural circuit named as neuristor was built as a hardware Hodgkin-Huxley model ( Hodgkin and Huxley 1952 ) that faithfully mimicked the action potential generation in biological axons, as shown in Figure 3 H ( Pickett et al. 2013 ). In another work, more biologically plausible and intrinsically stochastic neurons were built with vanadium dioxide Mott memristors, which exhibited twenty-three types of biological neuronal behaviors ( Yi et al., 2018 ). The controllable frequency of spikes in these artificial neurons also finds applications in coupled oscillator networks ( Csaba and Porod 2020 ). Exploiting Bistable Behavior of Memristive Devices Without fine material-level and device-level optimization, the memristive device usually shows limited conductance levels other than the capability of analog conductance tuning. With limited conductance states, the conventional artificial neural network needs to be adapted. This can be done by quantizing the analog weight value from offline learning ( Milo et al., 2019 ), which generally results in some loss of recognition accuracy. Many memristive devices only show binary stable states, i.e., high conductance state and low conductance state. For memristive devices embodied as STT-RAM and FeRAM, analog switching is generally more challenging. To exploit the bistable behavior of memristive devices for synaptic applications, a binary neural network was proposed relying on binary synapses (only with two states) and binary node value ( Hirtzlin et al., 2020 ). In the binary neural network, since the weights and inputs from the preceding layer are both binary valued, the weighted outputs are also binary, thus the vector-matrix multiplication becomes an XOR operation ( Luo et al., 2019 ). The accumulation/summation function in the neural nodes degenerates to POPCOUNT operations, i.e., counting the number of “1”s in a series of bits, eliminating the needs of a high-precision current sensor. The activation function afterward is only a sign function, further reducing the computational needs in the neuron nodes. The binary neural network also shows high tolerance to weight bit error ( Hirtzlin et al., 2020 ). Ternary content-addressable memory (TCAM) is another algorithm that intrinsically exploits the bistable behavior of memristive devices ( Yang et al., 2019 ; Ni et al., 2019 ). TCAM can perform in-memory search and pattern matching between the query feature vector and stored vectors of binary bits. In the study by Yan et al., 2019a , Yan et al., 2019b , 2-transistor/2-RRAM TCAM cells were used to store the TCAM vectors. For each TCAM cell, the stored TCAM datum was defined as the bit “1” for RRAM1 in HRS and RRAM2 in LRS, the bit “0” for RRAM1 in LRS and RRAM2 in HRS, the bit “X” (do not care bit) for both RRAMs in HRS. Thus, only two states of the memristive device were required. Within a similar scenario, ferroelectric TCAM with each cell only consisting of two ferroelectric field-effect transistors (FeFETs, three-terminal forms of FeRAM) has also been proposed ( Ni et al., 2019 ). Recently, an analog memristive TCAM was introduced by taking advantage of the analog programming in RRAM devices ( Li et al., 2020a ). Toward a Brain-like Computing System The first and second panels of Figure 4 summarize the projections of various memristive features and the corresponding brain-inspired functions discussed in previous section. Building upon these components, the next step would be the construction of brain-like algorithms and realization of cognitive computations as alternative solutions to the DL techniques. Mainly, two approaches can be seen in recent developments of memristive neuromorphic systems. One is from the signal morphological aspect, to emulate the spiking behavior of the biological neural network ( Roy et al. 2019 ). The other one is from the connection morphological aspect, to emulate the collective state dynamics and evolution of the biological network. The SNN ( Wang et al., 2018b ), closely mimicking the information presentation in biological neural systems, is considered as a viable way to achieve brain-like computing with high energy efficiency and error tolerance. Collective-state computation has several forms, such as Hopfield neural network ( Hopfield 1982 ), cellular neural network ( Chua and Yang 1988 ), and coupled oscillators ( Csaba and Porod 2020 ), mimicking brain activity at a high level, solving the problem by the system automatically finding its stable states in its energy landscape. A hybrid solution of these two approaches is also possible. However, there are no clear projections of which and how many brain-inspired functions can be utilized in these two approaches as illustrated in Figure 4 . Figure 4 Exploiting Memristive Features in Various Brain-Inspired Algorithms and Projections of Using These Brain-Inspired Functions for Cognitive Computations with Brain-Like Algorithms The unique features of memristive devices have been proposed to realize various brain-inspired functions for a neural network (from the first column to the second column); however, how to combine these brain-inspired functionalities to realize brain-like algorithms (from the second to the third) and for practical cognitive functions (from the third to the fourth) has no clear paths. Spike-Timing Dependent Plasticity and Spiking Neural Network SNN is considered as the third generation of neural networks ( Maass 1997 ), following the first generation based on McCulloch-Pitts neurons ( McCulloch and Pitts 1943 ) with digital input and output, and the second generation composed of multiple perceptron layers with gradient descent learning algorithm, which applies activation functions after the weighted sum of the inputs and achieves analog-valued input and output ( Lecun et al., 1998 ). The various DL neural networks with analog-valued input and output in section “ DL Accelerators by Memristive Hardware ” can be in general converted to an SNN, with the spiking rate of each neuron proportional to the analog value. Instead of ADC/DAC conversion introduced in “ Co-design with CMOS Peripheral Circuits ”, the analog current/voltage in the memristor array can be converted by integrate and fire (IF) neurons ( Milo et al., 2016 ; Yang et al., 2019 ). However, this does not fully exploit the benefits and capabilities of an SNN. In the human brain, by encoding information using spike timing, an extremely sparse and energy-efficient representation can be achieved ( VanRullen et al. 2005 ). Conversion of the analog current/voltage in the memristor array into spatial-temporal spike representation in the digital domain using leaky integrate and fire (LIF) neurons with temporal dynamics ( Fang et al., 2019 ) provides the possibility of constructing spatiotemporal spiking neural network. Instead of error backpropagation, one commonly utilized mechanism for learning in SNNs is the STDP of synapses ( Bi and Poo, 1998 ). The STDP learning rule has its biological root originated from the Hebbian learning rule, where “neurons that fire together, wire together” ( Hebb 1949 ). Memristive synapses capable of STDP and triplet-based learning have been widely reported ( Wang et al., 2015 , 2020b ). The weight updates depend on the timing of the presynaptic and postsynaptic spikes: the synapse weight is potentiated if the presynaptic spike precedes the postsynaptic spike, and depressed otherwise. The general realization of this STDP property in memristive devices is based on the engineered shapes of the presynaptic spike signal and the postsynaptic spike signal and their overlap in time ( Linares-Barranco and Serrano-Gotarredona 2009 ; Stoliar et al., 2019 ). It can also be realized without spike overlapping by utilizing the internal dynamics of volatile diffusive memristors, which faithfully emulate what happens in biological synapses ( Wang et al., 2017 ) (see also section “ Exploiting the Volatile Memristive Switching ”). Several neuromorphic systems based on the STDP weight update mechanism have been reported for pattern recognition. A PCM-based one-layer neural network for online pattern learning and recognition has been demonstrated by assuming the alternation of pattern and noise spikes from the pre-neurons and competition between post-neurons ( Ambrogio et al., 2016 ). The essential idea is that the simultaneous pattern spikes in the pre-neurons result in a spike in one of the post-neurons, and potentiation will be induced in their connecting synaptic devices via the STDP rule, while noise spikes following the spike of the post-neuron result in depression of the according synaptic devices. The same methodology can be applied to the neuromorphic system based on RRAM memristive synaptic devices ( Pedretti et al., 2017 ). Based on a similar methodology, the detection of the coincidence of simultaneous spikes representing an image among noise was developed ( Sebastian et al., 2017 ; Prezioso et al., 2018 ) ( Figure 5 A). Figure 5 Construction of Memristive Neuromorphic System Utilizing Brain-Inspired Algorithms Enabled by Memristive Devices (A) Learning and recognition of an image by the detection of the coincidence of simultaneous spikes with the help of PCM synaptic devices capable of the STDP. Reproduced from ( Sebastian et al., 2017 ), copyright © 2016, Springer Nature, CC BY. (B) Spatiotemporal computation considering the precise times of each spike within a spiking pattern. Reproduced from ( Wang et al., 2018b ), CC BY. (C) A Hopfield neural network for associative learning by fully connecting all neurons via bidirectional synapses. Reproduced from ( Milo et al., 2017 ), copyright © IEEE 2017. (D) Coupled nano-oscillators enabling mimicking of neural synchrony for vowel recognition. Reproduced from ( Romera et al., 2018 ), copyright © 2018, Springer Nature. Spatiotemporal spiking patterns can also be learned in a memristive neuromorphic system via a modified STDP learning rule, where the potentiation and depression of memristive synapse can be related to the precise timing of its received spike ( Wang et al., 2018b ) ( Figure 5 B). This can potentially enable direct learning and recognition of spatiotemporal signals in the real world, such as speech, motion, and gesture recognition ( Wang et al., 2019a ). Unsupervised learning based on STDP like learning rule has been demonstrated in a fully memristive neural network integrated with memristive synapse and diffusive memristor-based neurons. Pattern classification has been realized with such fully memristive neural network after unsupervised learning ( Wang et al., 2018 ). Collective-State Computing Hopfield networks realistically describe neurophysiological processes and exhibit associative memory behaviors with the system automatically evolving to attractor states ( Hopfield 1982 ). In the Hopfield network, each neuron receives input from all other neurons, and integrate-and-fire neurons can be employed ( Eryilmaz et al., 2014 ). Thus, when a fixed spiking pattern is presented to the neurons, the synapse can receive overlapping stimuli between self-generated spikes and the input spikes in its two terminals ( Figure 5 C). The synapse weight can thus be updated with Hebbian-like rules, such as STDP ( Milo et al., 2017 ). The learned configuration of synaptic weights forms an attractor state. After the learning, even if only part of the spiking pattern is presented to the neurons, the full spiking pattern can be recalled ( Milo et al., 2018 ), which is the basic concept of associative memory or content-addressable memory. The number of attractors that can be learned in a single synaptic array largely depends on the size of the network and is also affected by the learning rules. The Oja rule is reported to have a larger memory capacity roughly 10 times better than the Hebbian rule ( Wang et al., 2020a ). The cellular neural network only allows local connections between neighboring neural cells ( Chua and Yang 1988 ). In a standard cellular neural network, the neuron cells are arranged in a 2D array and the synapses bridge each cell with its neighboring cells ( Duan et al., 2015 ). Many two-dimensional tasks such as pattern and image analysis can be solved parallelly with such a 2D arrangement of cellular neural networks. Owing to the localized synaptic connections or communication between cells and the fully parallel operations of each cell, it is more suitable to be implemented with the hardware within the neuromorphic regime. Memristive synapses can further reduce the area cost compared with a CMOS only solution ( Dominguez-Castro et al., 1997 ). Simulation results have shown that memristive cellular neural networks execute functions of image processing such as horizontal line detection, edge extraction, and noise removal ( Duan et al., 2015 ). The non-linear I-V relation in memristive devices is incorporated in the analysis and simulation and has not proven to be an issue. However, the impact of other non-idealities of the memristive devices on the system performance needs further investigation. Disordered graphical network maps can be viewed as a special case of cellular neural networks. Theoretical analysis of these networks based on memristive connections shows much richer dynamic behaviors ( Caravelli et al. 2017 ). The oscillation network is another example of collective-state computation. Coupled with a memristive circuit, two self-sustained relaxation oscillators show frequency synchronization and phase locking ( Ignatov et al., 2016 ). This is believed to convey two essential principles of biological computing, namely, synchronization and memory. More recently, the memristive-coupled oscillator network is extended for temporal binding of different attributes of the same object ( Marina Ignatov et al., 2017 ). In another work, the oscillators are implemented by spin-torque memristive devices ( Figure 5 D), while the coupling factors among the oscillators are tuned by the direct current through each oscillator ( Romera et al., 2018 ). Vowel recognition with four coupled spin-torque oscillators was experimentally demonstrated. Outlook The integration of the memristor-based neuromorphic computing systems requires a detailed co-design at various levels, ranging from material optimization to system engineering. At each level, there are various integration methods depending on the approach and goal of the final system. Joint efforts and collaborations from experts in various research fields are needed. This perspective clarifies the goals of the efforts at various integration levels for two approaches to memristive neuromorphic systems: the DL accelerator and the brain-like computation. The implementation of state-of-the-art DL techniques enabled by the material- and device-level optimization and by the array level adaption has been a fruitful exploration in memristive neuromorphic computing. It can be viewed as a model for the co-design of memristive devices and algorithms. This methodology mainly relies on the maturity of DL algorithms and the popularity of these techniques in the AI era. Thanks to this popularity, materials scientists and electrical engineers working on memristive devices have sufficient prior knowledge to explore current machine learning infrastructures and slightly modify the algorithms as needed for real situations encountered in memristive synapses. Prototype systems realizing benchmark cognitive functions, for instance, the image classification for MNIST dataset, CIFAR, have been demonstrated or simulated ( Yao et al., 2020 ; Ambrogio et al., 2018 ). However, demonstrations of large-scale fully integrated memristive neuromorphic solutions for DL acceleration beyond the relatively small tasks (e.g., MNIST, CIFAR), toward more practical applications (e.g., in the scale of ImageNet [ He et al., 2016 ]), are still lacking. Exploiting unique features, including those traditionally viewed as non-idealities, of memristive devices enables a more direct and efficient implementation of brain-inspired algorithms, resulting in artificial synapses, dendrites, and neurons closely resembling their biological counterparts, as well as some basic functionalities in biological systems. However, these brain-inspired algorithms do not directly result in practical computational capabilities. Compared with memristive DL accelerators, memristive brain-like computations are limited to a smaller scale or toy applications so far. Besides the technical issues that need to be addressed, memristive SNNs are relatively underdeveloped mainly because of the lack of a clear understanding of biological information representations and processes that occur in the brain. SNN and collective-state computation are two possible frameworks, both resembling essential features of biological computations, to utilize the brain-inspired algorithms for brain-like computation. The artificial synapses, neurons, and dendrites that can faithfully emulate their biological counterparts may eventually provide building blocks for bio-realistic artificial neural networks. Such neural networks not only serve as computation tools that can generate natural intelligence but also act as faithful biological emulators to verify neuroscience principles. Compared with the biological tissues that essentially compose a “Blackbox” for neuroscience experiments, such an electronic testbed could be considered a “Whitebox” where every node in the neural network can be monitored, measured, and understood. In this way, memristor-based brain-like neural networks will not only benefit from, but also be beneficial for, the understanding of how biological neural networks naturally process information. Co-design between memristive hardware and neural network algorithms is critical for developing such brain-like neural networks."
} | 18,916 |
34825136 | PMC8603217 | pmc | 15 | {
"abstract": "Summary The accumulation of ice will reduce the performance of the base material and lead to all kinds of damage, even a threat to people's life safety. Recent increasing studies suggest that superhydrophobic surfaces (SHSs) originating from nature can remove impacting and condensing droplets from the surface before freezing to subzero temperatures, and it can be seen that hydrophobic/SH coating has good freezing cold resistance. But such anti-icing performances and developments in practical applications are restricted by various factors. In this paper, the mechanism and process of surface icing phenomenon are introduced, as well as how to prevent surface icing on SHS. The development of SH materials in the aspect of anti-icing in recent years is described, and the existing problems in the aspect of anti-icing are analyzed, hoping to provide new research ideas and methods for the research of anti-icing materials.",
"conclusion": "Conclusions It is found that the SHS can increase the free energy barrier of ice core and reduce the heat transfer between the droplet and the surface because of its lower surface energy and smaller droplet contact area. Therefore, surface icing can be controlled to some extent by rational use of SHS with micro/nano or hierarchical surface structure and low surface energy chemical composition. In this review, the research progress of SH materials for anti-icing in recent years has been introduced from the beginning. Next, the mechanism of solid surface icing and two classical nucleation theories are also outlined, which provide theoretical support for adjusting ice formation of solid surface and designing passive anti-icing surface. Compared with the active de-icing method, the SHS can realize passive anti-icing by timely cleaning of water droplets on the surface before freezing, delaying the freezing time, and reducing the adhesion force of ice on the surface. Before water droplets are frozen on the surface, the SHS has a certain self-cleaning function for droplets colliding and condensing on the surface due to the large WCA and the small SA. However, in the high humidity environment, the water droplets on the SHS are always present and cannot be completely removed. In addition, when water droplets freeze on the surface, the SHS can delay the freezing time to prevent ice due to its special micro-nano structure. The micro-nano or layered structure of the SHS can capture the air to form an air bag, thus reducing the heat transfer between the water droplets and the substrate to prolong the freezing time. In addition, the nucleation rate of surface ice nuclei can be reduced by adding nucleation inhibitors to the SHS. When ice inevitably forms on the surface, it has been reported that SHSs can reduce the adhesion of ice to the surface, whereas it has also been suspected that SHSs can increase the anchoring effect of ice on the surface. Further research is needed to confirm this. From a practical point of view, SH materials with good mechanical stability are desirable in anti-icing. For example, grafting the polymer epoxy resin on the SHS not only improves the mechanical stability of the material ( Zhang et al., 2019b ) but also has excellent anti-icing performance. It is well known that the surface morphology of SHS will directly affect its performance, as well as the nucleation and crystallization of ice on the surface. On this basis, ordered rough structures, appropriate micro/nano microstructure, and SHSs with high mechanical strength can be constructed on the surface. At present, multi-functional SHSs with self-repair, self-migration, self-healing, and other characteristics can self-repair and even regenerate after wear, which greatly improves the service life of SH materials and brings hope that SH materials can be widely used from the laboratory to practical life. However, in the current research progress, part of the preparation method is expensive, complex process and is not suitable for large-scale production and application, so there is a great space for development in structural optimization and material selection. From my personal point of view, future research by researchers should focus on preparing surfaces with methods that are easy to operate, low-cost, environmentally friendly, and durable. Finding new solutions to ice is also crucial. Finally, I always believe that there will be more breakthrough work done by researchers in the near future. The practical application of SH materials in daily life and industrial fields will be further improved.",
"introduction": "Introduction In the nature, it is well known that surface icing and frosting are very common phenomena. It sometimes presents us with a kind of visual beauty, such as the unique phenomenon of sprinkling water becoming ice around Mohe River in northeast China. The sight is that when hot water is sprayed quickly into the air it forms a beautiful arc. In fact, it is because the temperature is so low that when pouring hot water into the air, a lot of water vapor condenses directly into ice crystals when it gets cold. In addition, the direct condensation of water vapor is also related to the local air particles and sufficient contact with the air. When the hot water poured out is sufficiently dispersed and fully exposed to the air, it will turn into a fog-like form. And these are interspersed with tiny ice crystals. Thus, the magical scene of water into ice appears. In addition, the ice may cause damage to road traffic, power transmission, building roofs, aircraft wings, wind turbines, ships, and other equipment surfaces, which impedes the operation of equipment, reduces the use efficiency of equipment and facilities, and even causes huge safety hazards ( Azimi Yancheshme et al., 2020 ; Guo et al., 2020 ; Latthe et al., 2019 ; Wei et al., 2019 ; Yin et al., 2020 ).Every year, a large number of safety accidents are caused by surface icing, which not only threatens people's life safety but also causes huge economic losses. For instance, the road problems caused by rain and snow lead to greatly weakened wheel and road friction, which makes the vehicle out of control and results in traffic accidents. The ice on the pavement also makes it easy for pedestrians to slip and fall ( Latthe et al., 2019 ). In addition, in 2008, there was a sudden snowfall in the South (in fact, freezing rain meets cooler objects and freezes quickly), the frost accumulation increased the weight of the wire. Many pylons were destroyed, which caused an economic loss of more than 16 billion dollars ( Azimi Yancheshme et al., 2020 ; Guo et al., 2020 ; Latthe et al., 2019 ; Pan et al., 2020 ; Wei et al., 2019 ; Yin et al., 2020 ). And when the ice is formed on the equipment surface, it can not only increase the energy loss but also cause damage to the surface of the equipment. The accumulation of ice on the wings of the plane can slow down the plane’s flight speed and even lead to seriously reduced upward force. The increased weight of the aircraft will block the moving parts, resulting in the plane to crash ( Latthe et al., 2019 ). Due to the ice on surface and increasing load and stress as well as vibration and turbulence, it inhibits the best performance of wind turbines and results in wind power equipment shutdown ( Azimi Yancheshme et al., 2020 ). Large amounts of snow, ice, and supercooled water freezing on the surface of the hull change the center of gravity of vessels and reduce its stability, thus heightening the possibility of capsizing in Arctic shipping lanes ( Azimi Yancheshme et al., 2020 ). For another example, in the low temperature environment, microelectronic materials (such as polymer sealant and chip connection film layer) may condense water vapor absorbed from the surrounding humid environment into dew or even frost during transportation and storage, which will cause damage to the electronic equipment ( Azimi Yancheshme et al., 2020 ). Icing usually affects the surface roughness of wind turbine blades and thus changes their aerodynamic performance. Moreover, it increases the total weight and load of the blades, causing a large amount of power loss and even the stopping of wind turbines due to the intervention of ice. In addition, uneven ice formation can lead to unstable blades, which can cause excessive vibration and damage to wind turbines, leading to safety problems and so on ( Wei et al., 2019 ). Therefore, in the past few decades, scientists have conducted many studies to improve the deicing function performance on the surface, which can be roughly divided into two kinds of main strategy against freezing, active and passive methods. Active method is mainly through external energy de-icing. There are many such methods and here is a simple introduction of several. For example, the principle of electric heating anti-icing is through applying an electric current by the electric heating element, melting the ice, and reducing the binding force between the ice and the surface ( Ibrahim et al., 2019 ; Pan et al., 2020 ). Especially, heating the wire is an effective method to remove the icing on the transmission line. However, there is a large energy loss and cost, and this method harnesses the current, which may cause electromagnetic interference to the equipment ( Lv et al., 2014 ). Hot air anti-icing is commonly used in the anti-icing system of aircraft. The principle is that electric energy is converted into heat energy by heating resistance, and the heated hot air flow is sent to the designated position to melt the ice attached on its surface. But this method is energy intensive and can cause damage to some materials on the plane. In serious cases, the equipment on the plane may malfunction, causing the plane to crash. Liquid anti-icing, is also known as chemical solution anti-icing. This method is to spray antifreeze liquid such as ethylene glycol, calcium chloride, and urea on the surface of the material ( Talalay et al., 2019 ), which can reduce the freezing point of water or make ice melt easily. In winter, salt is often applied to remove the ice and snow on the road. Because of high solubility in sodium chloride (NaCl), 75%–90% of added salt is reported to enter roadside environment through runoff or splash. Ground water may be contaminated by salt solution percolation. In addition, roadside vegetations, as well as water creatures living alongside the roads, are seriously affected ( Lv et al., 2014 ). Furthermore, certain chemical solutions are corrosive to metals ( Ramakrishna, 2005 ). The ice on the surface of a material is mechanically broken and removed, which is defined as mechanical de-icing. It is mainly driven by directly hitting the ice or using other means such as pneumatic or electric power to drive the machine. This method is usually used for removing ice from equipments that are easily approached, such as overhead transmission lines and power networks. This method often requires that people get direct contact with the lines and high towers. Hence, there is a big security risk. During de-icing, mechanical forces put extra pressure on the network and in some cases can lead to failures ( Lv et al., 2014 ). However, these methods are complicated in design and have the disadvantages such as high energy consumption, high application cost, and time consuming. Therefore, it is of great significance to study effective anti-icing methods. Passive methods, as a strategy to improve active methods, refer to the physical and chemical methods based on surface modification ( Cho et al., 2015 ). With advances in nanotechnology and bionics, SH materials are considered a suitable anti-icing material due to their special surface structure. It is hoped that they can better replace traditional active methods in ideal application conditions without consuming any other energy. Typically, researchers place droplets on a solid surface with a water contact angle (WCA) greater than 150° and a sliding angle (SA) less than 5°, known as an SHS. The initial inspiration of SHS comes from the lotus leaf effect. German botanists Barthlott and Neinhuis ( Barthlott and Neinhuis, 1997 ) revealed the surface structure of lotus leaves and found that the self-cleaning effect of lotus leaves is due to the micro-nano structure of its surface. There are micron papillae on the surface of the lotus leaf, and the papillae are covered with a thin layer of nano-waxy crystals, which can greatly improve the CA of water droplets on the surface of the lotus leaf and make the water droplets fall easily. Wang et al. (2009a) believed that another important reason for lotus leaf's SH effect was the nanoscale structure on the surface of mastoid process and waxy crystal. Some people have also observed the microscopic characteristics of the surface of the lotus leaf and found that the surface of the lotus leaf has a random distribution of nearly hemispherical papillae with the size of 5–10 μm and about 150 nm dendritic mastoid, as shown in the Figure 1 A. And the surface of the lotus leaf showed almost completely suspended water droplets ( Zorba et al., 2008 ). In general, the \"lotus effect\" refers to the self-cleaning ability of the surface of the lotus leaf. When the water droplets fall on the lotus leaf, because the SA of the water on SHS is very small, the water droplets cannot stay on the surface of the lotus leaf. The spherical droplets mingled with the dust attach on the surface of the lotus leaf and tumble down, leaving a clean surface of the lotus leaf. In nature, many plants and animals have SHS similar to lotus leaf, such as rose petals ( Bhushan and Her, 2010 ; Chen et al., 2018 ), clover ( Liu et al., 2019a ), taro leaves ( Chen et al., 2017 ; Yi et al., 2019 ), water striders ( Gao and Jiang, 2004 ; Lu et al., 2018 ), butterfly wings ( Zheng et al., 2007 ), the penguin body feather ( Wang et al., 2016 ), and so on ( Figure 1 ). Inspired by different plants and insects, researchers have developed many SHSs. Currently, SHSs are used not only for self-cleaning ( Jeevahan et al., 2018 ), oil/water separation ( Lv et al., 2018 ), water collection ( Tian and Wang, 2018 ), and anti-corrosion ( Sebastian et al., 2018 ) but also for anti-icing and other applications ( Si et al., 2018 ). It is found that the SH characteristics of solid surface mainly depend on its low surface energy and rough microstructure. Hence, the method to obtain the SH characteristics is to modify and fabricate rough micro surface structure with low surface energy materials on solid surface. In order to construct this SH structure, the researchers used a number of preparation methods, which can be broadly classified as top-down or bottom-up. The usual top-down approach refers to the preparation of highly controlled micro-nanostructures by engraving or machining, with the help of tools and lasers such as photolithography ( Fromel et al., 2020 ), laser treatment ( Li et al., 2019b ), plasma treatment ( Ryu et al., 2017 ), anodic oxidation ( Saji, 2020 ), etc. And the bottom-up approach refers to a material addition process or self-assembly process in which complex surface is formed by adding smaller building blocks of materials by nano- or microfabrication such as sol-gel technology, electrostatic spinning, coating, deposition, and so on ( Jeevahan et al., 2018 ; Sun and Guo, 2019 ; Zhang et al., 2020 ). In conclusion, a suitable roughness on the hydrophobic surface could be created using these techniques and using chemical modification with hydrophobic materials ( Pan et al., 2019 ; Wang et al., 2020b ), such as stearic acid, fluoroalkyl silane, lauric acid, polydimethylsiloxane, Teflon, and so on, in order to reduce the surface energies and produce an SHS ( Liu et al., 2017 ; Wang and Guo, 2019 ). For example, Qing et al. ( Qing et al., 2019 ) used an ingenious two-step method to fill the concave microstructure of sandpaper matrix with a layer of fluorinated inorganic/organic film. The combination of the micro-nano structure and the low surface energy given by fluorinated nanoparticles resulted in FTPSS (FAS-TiO 2 /PDMS SH sandpaper surface) with CA of 160.6° and SA of 3.9°. FTPSS retains its hydrophobic properties under several chemical and mechanical stresses such as finger wiping, water pressure, hot water, tape peeling, and sandpaper abrasion. In addition, surface wettability can be controlled effectively by changing mesh size of sandpaper template. Moreover, the sandpaper showed good resistance to snow and ice after being worn for 50 times and was able to prevent snow and ice from adhering to the substrate ( Qing et al., 2019 ). Feng et al. ( Feng et al., 2018 ) first prepared a layered SH aluminum alloy surface by one-step impregnation process and then grafted long hydrophobic alkyl chain stearic acid (STA) on the surface, so that a large amount of air was trapped on the surface. The SH aluminum alloy obtained by this method not only has rough surface but also has two kinds of compound structures in micron and nanometre scale. Figure 1 SH plants and animals in nature (A) Nature-inspired nanostructures for special nonwetting states. The lotus leaf surface and its microscopic features. Optical image of a droplet statically sitting on the surface of lotus leaf. Image reprinted with permission from Zorba et al. (2008) . (B) Water droplet beading on clover; (inset of (B)) a WCA of clover is 152°, showing a good SH property. (C) Water droplet beading on green Bristlegrass; (inset of (C)) a water CA of green Bristlegrass is 153°, showing a good SH property. Image reprinted with permission from Liu et al. (2019a) . (D) ESEM images of a Chinese red rose petal with rich surface textures in microscale and shape of a water droplet on the petal's surface, indicating its hydrophobicity with a CA of 149.8°. Image reprinted with permission from Chen et al. (2018) . (E) An iridescent blue butterfly. Image reprinted with permission from Zheng et al. (2007) . (F) Water surface dimples pressed by six SH legs of a water strider standing on the water surface. Image reprinted with permission from Lu et al. (2018) . (G) Photograph of the Penguin body feather and WCA. Image reprinted with permission from Wang et al. (2016) . (H) The optical image of tendril peanuts in nature; a water droplet floating on the obverse surfaces of tendril peanut leaves. Image reprinted with permission from Gou and Guo (2018) . (I) The photo of a droplet floating on a lotus leaf and SEM images of the lotus leaf surface taken from the literature. Image reprinted with permission from Yi et al. (2019) . When water droplets are applied to the surface, the contact area between water droplets and air is measured to account for about 92.0% of the total area. Compared with the blank aluminum alloy surface, the freezing time of the SH aluminum alloy surface can be delayed by 5–9 min, and the freezing temperature can be reduced to 2–4°C. It can be seen that the SHS has a certain anti-icing property ( Feng et al., 2018 ). In order to understand the suitability of SH materials in anti-icing, it is necessary to understand the formation process of ice and the corresponding methods to inhibit the formation and growth of ice, which is of great significance for the research of new anti-icing methods on solid surface and the improvement of the application of SH materials in anti-icing. The main content of this paper is the latest progress in the research of SH material for anti-icing. The formation process of ice on the material surface and the common strategies of SH material for anti-icing are summarized. Firstly, the mechanism of surface icing and the existence of critical ice core are summarized and discussed. Then it introduces the main characteristics of SH material for anti-icing, which is mainly divided into three parts. The first part is about being able to clean up surface water droplets in time before freezing, especially the water droplets that collide on the surface and those that condense on the surface. The second part is about controlling the formation of ice cores on the surface after freezing, thus prolonging the freezing time. The third part is about reducing ice adhesion and facilitating the removal of ice. Then the practical application of some SH materials in anti-icing is briefly introduced, and some existing problems, such as the mechanical stability of SH materials, are introduced, and then some existing solutions are introduced. Finally, the conclusions and prospects of developing new anti-icing methods on solid surfaces and improving the application of SH materials in anti-icing are described. Theoretical basis Static contact angle Young's equation In 1805, Thomas Young proposed that when the solid, liquid, and gas phases reach equilibrium on an absolutely smooth and homogeneous ideal surface ( Figure 2 A), the θ 0 of the drop is related to the interfacial energies acting between the solid–liquid ( γ S L ) , solid–vapour ( γ S V ) , and liquid–vapour ( γ L V ) interfaces ( Chu and Seeger, 2014 ): (Equation 1) c o s θ 0 = γ S L − γ S V γ L V Figure 2 Schematic illustration of theoretical wetting models Schematic illustration of a droplet placed onto a flat substrate (A) and rough substrates (B) and (C). Depending on the roughness of the surface, the droplet is either in the so-called Wenzel regime (B) or in the Cassie–Baxter (C) regime. Image reprinted with permission from Chu and Seeger (2014) . Young's equation only represents an ideal surface, and it only applies to the surface with uniform chemical composition and absolute smoothness. However, in reality, solid surface has uneven chemical composition and certain roughness. Therefore, two different models, the so-called Wenzel state ( Wenzel, 2002 ) ( Figure 2 B) and Cassie–Baxter state ( Cassie and Baxter, 1944 ) ( Figure 2 C), were developed to explain the wetting behavior on a rough surface ( Chu and Seeger, 2014 ). The Wenzel model In 1936, Wenzel proposed the Wenzel model and modified the Young model by introducing roughness because the actual solid surface was not uniform and had some roughness. Wenzel model assumes that when liquid comes in contact with the rough solid surface, it can completely fill the grooves on the rough surface (as shown in Figure 2 B). The contact between liquid and the surface of the surrounding body is in the mode of complete contact. Wenzel model can be expressed as follows: (Equation 2) c o s θ 1 = r γ S L − γ S V γ L V = r c o s θ 0 where θ 1 is the apparent CA of the rough surface, and r is defined as the roughness factor, which is the ratio of the actual area of the solid surface to the projected area. According to Wenzel model, for a specific surface, surface roughness has a magnifying effect on the wettability of solid surface: for the hydrophilic surface, the surface becomes more hydrophilic with the increase of surface roughness; for hydrophobic surfaces, surface roughness increases, and it leads to an increase in hydrophobicity. However, some hydrophilic materials can also produce SHSs through special treatment, which can not be explained by Wenzel theory, and Wenzel model is not applicable when the solid surface is composed of different chemical substances, indicating that Wenzel theory also has certain limitations ( Chu and Seeger, 2014 ; Yu et al., 2015 ). The Cassie-Baxter model Cassie and Baxter envisioned the solid surface as a composite surface and assumed that when liquid came into contact with the rough surface, droplets of liquid could encase air in the grooves of the composite surface, forming a solid-liquid-gas three-phase interface. In this case, the contact area of the water droplet consists of two parts, namely, the contact area between the water droplet and the solid surface, and the contact area between the water droplet and the gas trapped in the groove. In this way, Cassie model can be expressed as (Equation 3) c o s θ 2 = f S L c o s θ S L + f L V c o s θ L V where θ 2 is the apparent CA in the Cassie-Baxter state, f S L and f L V are respectively solid-liquid contact area fraction and gas-liquid contact area fraction ( f S L + f L V = 1 ) . θ S L and θ L V represent the CA of solid-liquid and gas-liquid, respectively. When the liquid only touches the top of the solid convexity and the air pocket is stuck below the liquid, a polymorphic plane is formed between the solid and the air, as shown in Figure 1 C. In such a polymorphic surface the air-dependent part of the surface can be considered to be in a nonhumid state. θ S L = θ 0 and assuming that the CA between the air and the liquid, θ L V = 180 ° , Equation 3 can be converted to (Equation 4) c o s θ 2 = f S L ( c o s θ 0 + 1 ) − 1 According to Equation 4 , when the air in the groove is completely filled with liquid, f S L = 1 , Cassie's equation is transformed into Wenzel's equation. At the same time, an SHS with large CA can be obtained by minimizing the contact fraction of solids ( Li et al., 2021 )."
} | 6,263 |
38796464 | PMC11127998 | pmc | 16 | {
"abstract": "By mimicking the neurons and synapses of the human brain and employing spiking neural networks on neuromorphic chips, neuromorphic computing offers a promising energy-efficient machine intelligence. How to borrow high-level brain dynamic mechanisms to help neuromorphic computing achieve energy advantages is a fundamental issue. This work presents an application-oriented algorithm-software-hardware co-designed neuromorphic system for this issue. First, we design and fabricate an asynchronous chip called “Speck”, a sensing-computing neuromorphic system on chip. With the low processor resting power of 0.42mW, Speck can satisfy the hardware requirements of dynamic computing: no-input consumes no energy. Second, we uncover the “dynamic imbalance” in spiking neural networks and develop an attention-based framework for achieving the algorithmic requirements of dynamic computing: varied inputs consume energy with large variance. Together, we demonstrate a neuromorphic system with real-time power as low as 0.70mW. This work exhibits the promising potentials of neuromorphic computing with its asynchronous event-driven, sparse, and dynamic nature.",
"introduction": "Introduction Resource and energy constraints are the major restrictions to deploying traditional AI methods, especially in real-world edge platforms. A promising solution with an attractive low-power feature is neuromorphic computing, which is partially inspired by the human brain that runs even more complex and larger neural networks with a total energy need of just 20 W 1 – 4 . By abstracting the computations in the human brain at the neuron and synapse level, existing neuromorphic platforms, such as the classic BrainScales 5 , SpiNNaker 6 , Neurogrid 7 , TrueNorth 8 , and the most recent Darwin 9 , Loihi 10 , Tianjic 11 , have demonstrated impressive energy efficiency via spike-based communication and computing. However, whether this level of abstraction 2 , 12 , 13 is the most suitable approach for emulating the efficient computation of the brain, and the role that high-level stereo brain mechanisms can play in neuromorphic chips, are challenges that must be addressed at this stage”. An important function of the human brain is the ability to dynamically allocate its resources according to the required demand, which is what we call “dynamic computing” due to the attention mechanism 14 , 15 . Salient stimuli tend to receive greater attention, primarily manifested in the heightened spiking activity of brain regions or neurons associated with the stimulus 16 . Incorporating the high-level dynamic computing nature of the human brain into machine intelligence is very challenging. Specifically, dynamic computing encompasses two connotations: energy consumption is minimal when there is no input, while it significantly varies with input changes. With these understandings as the anchor, we present an application-oriented algorithm-software-hardware co-designed neuromorphic system to investigate the dynamic and sparse computing of spike-based machine intelligence in our newly designed and fabricated neuromorphic chip. To achieve the hardware requirement of no-input consumes no running energy, we design and fabricate the “Speck” (Fig. S1) with the size of 6.1 mm × 4.9 mm, a spike-based and fully asynchronous neuromorphic chip with low processor resting power (only 0.42 mW ). The fully asynchronous architecture of Speck, which renders computing capacity solely dependent on input data, constitutes the key factor behind its persistent “always-on” profile. In this paradigm, the neuromorphic chip no longer needs the global or local clock signal, which efficiently prevents the redundant power consumed by clock empty flips. In other words, the asynchronous design can be understood as the most extreme form of fine granular clock gating for every component in the processing pipeline while being instantly available, requiring no wake-up procedures 17 . Meanwhile, by integrating the DVS 18 – 20 as the “eye” of the chip, Speck becomes the sensing-computing neuromorphic System on Chip (SoC). The DVS asynchronously and sparsely generates a stream of events (binary spikes with addresses) when the brightness of the visual scene changes. The processor in Speck only operates when receiving incoming events, leveraging its hardware circuit design to enable asynchronous event-driven distributed convolution processing of spike trains. Remarkably, the entire system processes a single spike with an ultra-low latency of only 3.36 μs. This collaborative philosophy between neuromorphic hardware and applications perfectly encapsulates the essence of dynamic computing. It empowers Speck with distinct advantages in scenarios with stringent power and latency requirements, such as mobile devices and the Internet of Things. In the context of spike-based computing, it is commonly believed that computation is triggered exclusively by input spikes. With each input, only a subset of the network becomes activated, resulting in the activation of multiple sets of spiking neurons. Therefore, it is natural to believe that varied inputs consume different energy in SNNs. However, we uncover a phenomenon called “dynamic imbalance” that commonly exists but has been ignored for a long time in SNNs, i.e., although spiking neurons are selectively and sparsely activated, spiking networks respond similarly to different inputs. Specifically, we observe that spiking firing rates in vanilla SNNs at each timestep are very similar, which indicates that the scales of the activated sub-networks are similar for diverse input. It implies the connotations of dynamic computing referring to “varied inputs consume energy with large variance” usually does not hold in SNNs. Consequently, the dynamic computing advantage that neuromorphic system naturally have is undermined. To address this issue, we design an attention-based dynamic framework, which can assist SNNs in regulating spiking responses according to the importance of input discriminatively. To efficiently deploy algorithms/models for various dynamic vision applications, Speck provides a complete software toolchain, including data management, model simulation, host management, etc. This enables us to demonstrate the attractive features of the proposed neuromorphic system in accuracy, energy cost, and latency. To this end, dynamic SNNs are evaluated on four demanding event-based action recognition benchmarks. Extensive experiments show that the attention nature of the brain, data-dependent dynamic processing currently underappreciated in SNNs, can confer sparser firing and better performance to SNNs concurrently. By deploying the dynamic SNNs to Speck, we demonstrate a high-accuracy neuromorphic system with real-time power as low as 0.70 mW and ultra-low latency of less than 0.1 ms on a single sample in public datasets. The practice in this work demonstrates the power of the neuromorphic chip in dynamic computing, expands a creative path for the development of neuromorphic computing, and pushes neuromorphic computing a big step toward real-world applications.",
"discussion": "Discussion Although dynamic computing possesses bio-plausibility and fascinating properties such as accuracy, adaptiveness, and computational efficiency, these advantages currently exist only in theory. To truly demonstrate the power of sophisticated dynamic computing with machine intelligence, top-level design of the entire AI system is indispensable. In this work, we have shown an application-oriented algorithm-software-hardware co-designed neuromorphic system that naturally and subtly embodies the unique advantages of dynamic computing regarding energy consumption, output latency, and task accuracy. We have presented a sensing-computing asynchronous neuromorphic SoC like an eye-brain integrated system to realize attention-based dynamic computing. At the algorithmic level, we have revealed that brain-inspired spiking communication makes SNNs inherently capable of dynamic computing. Still dynamic imbalance caused by another fundamental assumption of SNNs, spatio-temporal invariance, undermines this gift. Inspired by the attention-based dynamic response mechanism in the human brain, we proposed dynamic SNNs, which combine the attention-based dynamic framework with vanilla SNNs to improve the ability to focus on important information so that varied inputs consume energy with large variance. Experimental results show that dynamic SNNs can simultaneously achieve the two main considerations for realizing machine intelligence - effectiveness and efficiency 28 . We were pleasantly surprised to find that dynamic SNNs correspond well to the attention mechanisms in the human brain both structurally and functionally 15 . Attention stereoscopically regulates the firing of spikes in the brain’s neural circuits, brain regions, and neurons, and the dynamic framework assists the vanilla SNNs in comprehensively optimizing the spike firing of networks, layers, and neurons. Consequently, dynamic SNNs can kill two birds with one stone by focusing on important information while suppressing noise spikes, significantly reducing network energy consumption while improving performance. At the hardware level, we have demonstrated Speck, which bringing to reality the theoretical advantages of dynamic SNNs at the algorithmic level. The most intriguing feature of Speck is the low resting power (no-input consumes no running energy) brought about by the fully asynchronous design, i.e., always-on, making it particularly competitive in resource-constrained edge computing scenarios. This is also the basic hardware requirement for dynamic computing. We have demonstrated that the energy gain from the sophisticated dynamic algorithm design is completely negligible once the resting power is too high. Moreover, present-day neuromorphic computing frequently separates the design of applications, algorithms, and chips. The needs of hardware and applications are rarely considered when designing neuromorphic algorithms and vice versa. By contrast, Speck incorporates a fully asynchronous spike-based neuromorphic chip with a DVS camera, creating the perfect blend of hardware and applications well suited for dynamic computing. Calculations in Speck are only triggered when DVS generates an event. Comprehensively, based on our top-level design of dynamic algorithms, chip architecture, and real-world application requirements, we have demonstrated mW-level power and ms-level latency solution in typical dynamic visual scenarios. This tapping into the potential of neuromorphic computing will undoubtedly advance the field. In target edge computing environments, the overhead energy is strictly constrained, especially for a small system working in self-powered mode for a long time. Speck is a neuromorphic chip with sensing-computing-integrated functionality, which consumes quite low-power consumption via asynchronous digital design. Such high energy efficiency and low production cost are difficult to promise modeling flexibility and computing precision. Fortunately, it is acceptable in our target scenarios where energy efficiency matters more than the task difficulty and behavior accuracy. We believe Speck can cover a broad range of neuromorphic-vision-specific edge computing tasks distinct from cloud computing while improving modeling flexibility and computing precision under the energy constraint remains an interesting and valuable direction. For example, enriching the supported network types and introducing mixed-precision computing might be possible solutions in future work. At the software level, Speck provides a complete software toolchain to enable neuromorphic computing to be effectively and efficiently deployed in various applications based on dynamic vision. Specifically, the complete software toolchain provided by Speck, including data management, model simulation, host management, etc., can promote the rapid deployment of neuromorphic computing. We look forward to these engineering efforts to promote the advantages of neuromorphic computing in more applications. Finally, incorporating an attention mechanism to SNNs in neuromorphic hardware can be seen as a first step towards porting more sophisticated high-level neural mechanisms in the human brain 15 , 34 into such hardware. As well known, neuromorphic hardware is non-von-Neumann architecture hardware whose structure and function are inspired by brains. Some unique fundamental operational characteristics, including highly parallel operation, collocated processing and memory, inherent scalability, and event-driven computation, stem from their choice to incorporate neurons and synapses to serve as the primary computational units. Although the vast majority of neuromorphic computing works have been based on the model design and hardware implementation of spiking neurons, it is unclear whether they are the only aspects of the biological brain important for performing computations. The practice in this work confirms that the attention mechanism is also very important for computing. Neuromorphic computing should consider the response of neuron granularity and perform overall control from a higher abstraction level, like the human brain. Even more exciting, these high-level abstractions of brain mechanisms may be functionally and structurally well-suited for implementation in brain-inspired neuromorphic computing. In particular, neuromorphic computing may contribute in answering one of computational neuroscience and machine learning’s important open questions: how can diverse high-level neural mechanisms generated during the evolution of the brain be imitated and incorporated into computers to enable machine intelligence to function similarly to the brain?"
} | 3,452 |
39823335 | PMC11740947 | pmc | 17 | {
"abstract": "Certain coral individuals exhibit enhanced resistance to thermal bleaching, yet the specific microbial assemblages and their roles in these phenotypes remain unclear. We compared the microbial communities of thermal bleaching–resistant (TBR) and thermal bleaching–sensitive (TBS) corals using metabarcoding and metagenomics. Our multidomain approach revealed stable distinct microbial compositions between thermal phenotypes. Notably, TBR corals were inherently enriched with microbial eukaryotes, particularly Symbiodiniaceae, linked to photosynthesis, and the biosynthesis of antibiotic and antitumor compounds and glycosylphosphatidylinositol-anchor proteins, crucial for cell wall regulation and metabolite exchange. In contrast, TBS corals were dominated by bacterial metabolic genes related to nitrogen, amino acid, and lipid metabolism. The inherent microbiome differences between TBR and TBS corals, already observed before thermal stress, point to distinct holobiont phenotypes associated to thermal bleaching resistance, offering insights into mechanisms underlying coral response to climate-induced stress.",
"introduction": "INTRODUCTION Corals are foundational species that support reef ecosystems, which provide habitat for more than a third of marine life ( 1 ). These ecologically and economically important ecosystems are threatened by local and/or global stressors ( 2 ), such as pollution ( 3 ) and ocean warming ( 4 ). Coral bleaching (i.e., the disruption of the relationship between the host and the endosymbiont photosynthetic algae from the family Symbiodiniaceae), mainly caused by ocean warming, has been considered one of the main drivers of the massive die-offs in different regions in the past decades ( 5 , 6 ). Despite their general sensitivity to thermal stress, some coral populations ( 7 , 8 ) and individuals ( 8 – 10 ) demonstrate differential susceptibility to thermal bleaching. Although there is no consensus in the literature on the use of the terms thermal “resistance,” “tolerance,” or “sensitivity” for corals or other marine organisms, here, we consider corals that retain their pigmentation during a bleaching event as bleaching resistant corals, as recently suggested by Matsuda et al. ( 10 ). Resistant phenotypes can occur through adaptation or acclimatization of the different members of the holobiont. By definition, adaptation consists of a process that occurs over generation(s), helping individuals and populations to permanently increase their fitness ( 11 ). Acclimatization, in contrast, refers to physiological plasticity that allows a single individual to be temporally and reversibly more tolerant or resistant to a specific condition ( 12 ). These processes can be linked to the coral host, the photosynthetic endosymbiotic algae of the Symbiodiniaceae family, and the other members of the coral microbiome, each of them potentially playing a role in resistance mechanisms ( 13 ). Some studies attributed coral resistance to host adaptation via natural selection of a heat-resistant state due to exposure to, for example, naturally warmer temperature regimes ( 7 , 14 ). This process can be induced by exposure to subbleaching temperatures ( 15 , 16 ) and/or to associations with beneficial microbial assemblages [either Symbiodiniaceae ( 17 , 18 ) or bacterial communities ( 19 , 20 )]. For example, specific associations between corals and some members of the Symbiodiniaceae family, such as those of the genus Durusdinium , have been demonstrated to contribute to differences in heat tolerance ( 21 ). Other members of the microbial community may be an even more plastic and dynamic part of the holobiont and could contribute to physiological improvements on a much more flexible or shorter timescale than genetic adaptations of the host ( 22 ) or changes in the Symbiodiniaceae populations ( 23 ). The role of microbial groups other than the photosynthetic algae, such as bacteria or viruses, has been explored in coral fitness and thermal stress resistance ( 19 , 24 , 25 ). For example, some specific bacteria can quickly respond to environmental impacts ( 22 , 26 ) and contribute to the holobiont’s increased resistance to heat stress ( 27 , 28 ), although the ability to change microbial association (i.e., microbial flexibility) differs between host species ( 22 , 29 ). Despite the interconnections between all members of the holobiont and their potential links to coral resistance, most studies have focused on either host-Symbiodiniaceae ( 23 ) or host-bacteria ( 19 ) associations, with some recent insights into the microbiome associated with cultures of free-living ( 30 ) and in hospite Symbiodiniaceae ( 31 ), or the role of the microbiome in modulating the host’s epigenome ( 32 ). Specific holistic and multidomain assemblages and potential mechanisms associated with corals exhibiting differential thermal bleaching susceptibility and its response to thermal stress have not yet been fully explored. Given the rapid pace of current environmental changes, elucidating the mechanisms and taxa underlying coral health ( 25 , 33 ) and bleaching resistance in coral holobionts is crucial for developing strategies to boost coral resilience and reduce mortality under future climate change scenarios ( 34 – 37 ). Here, we examine two phenotypes of the same coral species ( Mussismilia hispida ) exhibiting different levels of resistance to bleaching that were categorized into thermal bleaching resistant (TBR) and thermal bleaching sensitive (TBS) based on their differential responses to heat stress during long-term experiments. Our results reveal that each phenotype inherently harbors specific microbial assemblages and each group contributes proportionally differently to the holobiont’s metabolic traits. Furthermore, TBR corals exhibited a higher abundance of microbial eukaryotes, predominantly Symbiodiniaceae, influencing their metabolic profiles, particularly photosynthetic and membrane anchoring proteins, as well as the biosynthesis of antibiotic and antitumor compounds. In contrast, bacteria was the main group contributing to metabolic genes associated with TBS corals, including nitrogen, amino acid, and lipid metabolism. While shifts in the microbial structure were observed in response to heat stress over time within each phenotype, no overall changes were detected in the functional profiles of either TBR or TBS, suggesting that inherent stable differences may contribute to distinct thermal resistance. These findings indicate a correlation between the coral-Symbiodiniaceae-microbiome assembly, providing key insights to inform strategies aiming to counteract coral mortality in the face of climate change.",
"discussion": "DISCUSSION Distinct bleaching responses In this study, two different phenotypes of the coral M. hispida demonstrated different bleaching responses during a long-term heat stress experiment. Despite the lack of consensus in the literature regarding the use of the term thermal resistance for corals or other marine organisms, we followed the definition provided by Matsuda and colleagues ( 10 ), which defines bleaching resistant corals as those that maintain their pigmentation during a bleaching event. We used the Coral Health Chart as a color reference to assess bleaching, defining bleaching as the decrease of two or more color units ( 38 ). This approach enabled us to categorize the different phenotypes of M. hispida as either thermally-resistant (TBR) or thermally-sensitive (TBS) corals. Ten days of exposure to an acute heat stress caused bleaching and mortality in TBS corals. Significant decreases in the TBS corals’ F v /F m rates indicate temperature-related damage to the photosystem II (PSII) electron transport of the Symbiodiniaceae, which is consistent with the loss/expulsion of Symbiodiniaceae and visual signs of bleaching ( 39 ). Although TBR corals exhibited no visible signs of bleaching, even when exposed to higher temperatures than TBS corals, some fluctuations in the F v /F m rates during heat stress were observed. Despite this drop in the F v /F m rates in T1, TBR corals restored their F v /F m rates during the recovery period (T2). Inherent differences in the host-Symbiodiniaceae-microbiome assemblages and their functional contributions Beyond examples such as physiological acclimatization ( 8 ), host adaptation ( 8 , 40 ), and the assisted migration of heat-tolerant alleles ( 9 , 41 ), the association of corals with specific groups of Symbiodiniaceae ( 10 ) or other microbial assemblages ( 19 ) represents additional mechanisms that may contribute to thermal bleaching resistance. TBR and TBS corals inherently harbor significant taxonomic differences even under non-stress ambient conditions (i.e., T0). First, TBR corals were dominated by the Symbiodinium A4 type, followed by a small portion of Cladocopium C21-C3vv-C3-C3vw-C50br, whereas TBS corals mainly hosted Cladocopium C3-C3vv-C3ww-C3-C50br-C3vu followed by Symbiodinium A4/A1-A4br-A4bq-A1mi-A1b. Both Symbiodinium A4 and Cladocopium C3 are widespread and generalist ITS2 types ( 42 ) that are commonly found in corals of the genus Mussismilia ( 43 ). Previous studies have demonstrated that Cladocopium C3 is more frequently associated with mild temperature and high light conditions, whereas Symbiodinium A4 is more commonly present in shallow waters and under higher irradiance ( 44 ). In addition, the TBR-dominant Symbiodinium A4 has been reported to facilitate positive growth rates at high temperature in experiments in hospite (i.e., in association with Porites divaricatea ) ( 45 ). Despite differences in the dominant Symbiodiniaceae types between TBR and TBS, both phenotypes maintained a stable algal community composition, even when heat stress was applied. Although the inherent differences between TBR and TBS corals may stem from specific responses to past temperature stress and/or environmental factors from their original site, the long-term rearing of these corals under the same aquarium conditions suggests adaptation to the current environment ( 46 ). This enables us to explore alternative mechanisms contributing to coral bleaching resistance, providing insights beyond site-specific factors. TBR corals harbored a higher relative abundance of eukaryotes (predominantly Symbiodiniaceae) compared to TBS corals. The higher relative abundance of Symbiodiniaceae and other dinoflagellates, such as Dinophysiaceae, in TBR corals is also reflected in the enrichment of photosynthetic genes such as psbA , psbD , and petB , which encode the D1, D2, and cytochrome b6 proteins—core components of the photosynthetic apparatus in algae and other photosynthetic organisms ( 47 , 48 ). This observation aligns with the dominance of Symbiodinium A4 ITS2 type in TBR corals, which has been previously reported to enhance growth rates at high temperatures ( 45 ). Furthermore, disparities in microbial structure and composition mirrored and seem to influence the distinct functional profiles observed in each coral phenotype. For example, microeukaryotes contributed significantly to genes associated with catabolic pathways in TBR corals, while bacteria were the main contributors to metabolic genes associated with TBS corals. The biosynthesis of GPI-anchor proteins, specifically GPI phospholipase D and GPI-anchor transamidase subunit K, was significantly enriched in TBR corals. These proteins play crucial roles in cell wall assembly, hardening, and softening ( 49 , 50 ). In symbiotic dinoflagellates, the cell wall tends to be thinner to facilitate nutrient exchange and communication ( 51 , 52 ), potentially enhancing the metabolic exchange between Symbiodiniaceae and other members of the microbiome. This closer interaction could, in turn, foster specific microbial assemblages that support the resistance of TBR corals. Moreover, GPI-anchor proteins have been proposed as alternative phosphate sources under mildly acidic pH or phosphate-limited conditions ( 53 ), although the exact role of these proteins in cnidarian-dinoflagellate symbiosis under thermal stress requires further investigation. In addition, TBR corals showed enrichment in oleandomycin and maduropeptin biosynthesis. While both compounds exhibit substantial antibiotic properties ( 54 ), maduropeptin also demonstrates potent antitumor activity by targeting rapidly dividing cells ( 55 ). These bioactive molecules could, for example, be involved in controlling pathogens and/or fast-growing organisms within the microbial community of TBR corals, potentially contributing to their overall resistance to thermal stress ( 56 ). The balance between eukaryotic and bacterial players may contribute to TBR corals being proportionally enriched in metabolic pathways related to the metabolism of carbohydrates and proteins, which could shape their microbial associations. Similar to free-living phytoplankton and other microbial eukaryotes, Symbiodiniaceae have been hypothesized to exude metabolites creating an enriched zone around themselves ( 57 , 58 ), which attracts and supports the growth of other microorganisms ( 59 , 60 ). Known as the “phycosphere,” this physical interface might selectively promote associations with other microbial eukaryotes, bacteria, archaea, and viruses ( 61 , 62 ). Our results point to a potential differential phycosphere effect in TBR and TBS corals ( 62 ), which could selectively enrich particular bacterial groups associated with in hospite Symbiodiniaceae ( 31 ). Consequently, this could influence the corals’ phenotypic responses and levels of thermal resilience. In addition, all microbial groups, including microbial eukaryotes, viruses, archaea, and bacteria, seem to differentially associate with each phenotype. For example, compared to TBS, TBR corals present an enrichment of the dinoflagellate Dinophysiaceae, which belongs to the same class as their Symbiodinaceae counterparts (Dinophyceae and Alveolata) ( 63 ). The annotation of several genes of Dinophyceae, including percentage where available, suggests their affiliation within the Symbiodiniaceae family. However, due to the resolution limitations inherent to gene-based taxonomy, particularly in metagenomic approaches, these sequences could not be classified at a more specific taxonomic level within the family. This is a common challenge in metagenomic studies, where the available genetic markers often lack the precision needed for higher-resolution classifications. Nevertheless, the enrichment of either Dinophyceae or Symbiodiniaceae may potentially contribute to their overall fitness and/or to attracting other microbes due to their photosynthetic capacity. Viruses such as Leviviridae and Hepadnaviridae were also inherently enriched in TBR corals. Bacteriophages play a crucial role in shaping bacterial populations within the coral microbiome by selectively targeting specific bacterial species. For instance, Leviviridae, previously detected in soil ecosystems, have been shown to regulate populations of Gammaproteobacteria and Alphaproteobacteria ( 64 – 66 ), which could similarly contribute to microbial regulation in corals. While Hepadnaviridae are primarily associated with vertebrates, they have been occasionally identified in corals ( 67 ). However, their low abundance in metatranscriptomic data suggests that these viruses are not highly active, possibly leading to an underestimation of their influence ( 68 ). Archaea are also known to play crucial roles in various extreme ecosystems, including contaminated and methanogenic soil ( 69 ) and hydrothermal vents ( 70 ). Although not commonly reported in corals, Caldisphaeraceae, Methanospirillaceae, and Nitrososphaeraceae are thermophilic archaea typically found in harsh environments and were also inherently enriched in TBR corals. In TBR corals, these archaea may contribute to nutrient cycling and ammonia oxidation ( 71 ), as they do in other ecosystems, potentially supporting the resilience of these corals under thermal stress conditions. While metagenomic taxonomic profiling provides a broader overview of the microbiome, 16 S rRNA gene sequencing offers greater taxonomic resolution for identifying specific bacterial groups. Although groups such as Rhizobiales, Clostridia, Paramaledivibacter , and Cytophagales were not detected as differentially abundant in the metagenomic analysis comparing coral phenotypes, ASVs from these groups were enriched in TBR corals. Both Clostridia and Rhizobiales have been associated with coral tissue loss and lesion progression in previous studies ( 72 ). However, discerning between detrimental and beneficial organisms within the complex coral holobionts is particularly challenging ( 73 – 75 ). For example, Rhizobiales are known to form symbiotic relationships with plants, contributing to nitrogen fixation, methane oxidation, and microsymbiosis ( 76 , 77 ). Similarly, Young et al. ( 78 ) reported a strong correlation between the expression of healing genes and diseased corals, suggesting that microbial responses to stress may not directly indicate disease progression but may instead reflect a protective or regulatory role performed by the microbiome. Spirochaeta and Sphingorhabdus were also inherently enriched in TBR corals. Members of the Spirochaetaceae family are non-pathogenic, free-living anaerobes known for their ability to fix nitrogen ( 79 ) and degrade organic carbon ( 80 ), and have been previously identified in Acropora palmata ( 81 ) and various color morphs of Corallium rubrum ( 82 ). Cultures of the thermotolerant Symbiodinium pilosum were found to be dominated by Sphingorhabdus [Díaz-Almeyda et al. ( 83 )]. S. pilosum exhibits high thermal tolerance, showing no decline in growth rate or photochemical efficiency even at 32°C ( 83 ). Although Sphingorhabdus has been detected and isolated from gorgonian corals ( 84 , 85 ), its broader role remains underexplored. Further research could uncover its role in supporting S. pilosum and the holobiont’s thermotolerance, offering insights into coral resilience mechanisms, suggesting their potential as markers for tracking thermally resistant corals or as candidates for future studies on beneficial microorganisms. The enrichment of taxa previously associated with diseased corals together with bacteria with known beneficial traits in TBR suggests a dynamic biological regulation by the host and/or other biological forces within the holobiont. In TBS corals, specific bacterial taxa such as Celerinatantimonadaceae, Pasteuriaceae, members of the Rhodobacteraceae family (uncategorized ASVs and Roseovarius ), Rickettsiales, Waddlia , and Agaribacterium were inherently enriched. While some of these taxa, such as Celerinatantimonadaceae, Pasteuriaceae, and Agaribacterium , are not well documented in corals or other marine organisms, others are commonly associated with disease. Members belonging to the Rhodobacteraceae family, for example, are commonly found in development and progression of coral disease and sewage ( 86 – 88 ). Rickettsiales are Gram-negative bacteria known to cause diseases in invertebrates ( 89 , 90 ). Waddlia , and other species classified within the Chlamydia-like group have demonstrated to be strong cytopathic-diseased vectors in fish cell lines ( 91 ). In addition, TBS corals showed an enrichment of bacterial metabolic genes, similar to previous findings where diseased coral tissues were enriched with bacterial sequences, compared to healthy tissues ( 92 ). Metabolic redundancy in corals under stress Although some shifts in the microbiome structure are observed throughout time as a response to the applied heat stress, no overall changes were detected in the functional profile of either TBR or TBS corals. This may suggest some level of metabolic redundancy, in which taxonomic changes are observed, but these shifts are, somehow, regulated in a manner that functions remain the same ( 93 , 94 ). This is one strategy some holobionts may use to maintain its homeostasis. On the basis of our observation, the “winner holobiont” seems to depend on the baseline and inherent holobiont assemblage before any stress is applied, indicating that these specific assemblages underscore determined functions that may support coral holobionts during heat stress. In our study, we found specific Symbiodiniaceae-microbiome assemblages inherently associated with distinct TBR and TBS phenotypes of M. hispida . In addition, the proportion of microbial eukaryotes and bacteria harbored by either TBR or TBS corals seems to directly influence the metabolic profile of these phenotypes, especially their metabolism of proteins, carbohydrates, and energy. More specifically, TBR corals exhibited an inherent enrichment of microbial eukaryotes, especially Symbiodiniaceae, which was mirrored by the enrichment of key functions such as photosynthesis, membrane anchoring, and the production of antibiotic and antitumor proteins. The biosynthesis of GPI-anchor proteins, which are essential for processes like cell wall assembly and regulation of its hardening and softening, was also notably enriched in TBR corals and may contribute to the exchange of metabolites between Symbiodiniaceae and other members of the holobiont. Conversely, TBS corals showed a predominance of bacterial metabolic genes, particularly those involved in nitrogen cycling, amino acid synthesis, and lipid metabolism. Although the observed differences between TBR and TBS can be driven by several factors, such as past environmental conditions, we hypothesize that these distinct and consistent multidomain microbiome assemblages comprise distinct holobiont phenotypes ( Fig. 5 ). It is important to highlight that the differences between TBR and TBS corals reported in this study are inherent, e.g., observed at the beginning of the experiment when all corals were under the same environmental conditions. This baseline distinction suggests that the observed microbial differences are associated to the different phenotypes and not merely (or only) a result of thermal stress and other environmental conditions, providing a profound insight into the correlation between corals, their thermal phenotypes, and their associated microbiomes. These insights lay the ground for further investigations that can support rehabilitation approaches, such as the selection of coral probiotics aimed at mitigating coral mortality in the future climate scenario. Fig. 5. Microbial assemblages and the metabolic profiles associated with TBR and TBS coral phenotypes. Taxonomy and metabolic functions associated with TBR and TBS phenotypes. Taxa recovered from metagenomes are reported at the family level, whereas taxa recovered from 16 S rRNA gene sequencing are reported at the lowest level found. The metabolic functions enriched by the different domains are not necessarily connected to the taxa enriched and reported in the figure. The thick blue arrows indicate the microbial groups contributing a considerably higher portion of genes involved in the holobiont’s metabolism. The morphology of microorganisms in this figure is merely illustrative and represents the general morphology of one representative organism from the mentioned family. *Recovered from metagenomes. **Recovered from amplicon sequencing (16 S rRNA gene for bacteria and ITS2 for Symbiodiniaceae)."
} | 5,864 |
40213024 | PMC11983683 | pmc | 18 | {
"abstract": "ABSTRACT Advanced structural materials are often required to exhibit a combination of light weight, high strength and superior toughness. Biomimetic strategies hold promise for achieving these seemingly conflicting mechanical properties simultaneously. However, current biomimetic structural materials lack active fire-warning and passive flame-retardant functionalities, which poses risks for their application in fire-prone scenarios. Herein, we present a nacre-mimetic alumina–cyanate resin composite (NAC) that has a combination of mechanical robustness with thermochromic and flame-retardant properties. Through controlled atomic doping, chromium atoms are incorporated into alumina microplatelets, forming solid-solution assembly units that exhibit reversible thermochromism and a solid-solution-strengthening effect. The bioinspired ‘brick-and-mortar’ structure endows the NAC with high strength (∼290.1 MPa) and fracture toughness (∼11.1 MPa m 1/2 ). Coupled with a machine-learning-based image-recognition system, the NAC leverages its thermochromic properties to deliver a rapid fire warning within 9 s at 250°C, which is significantly faster than traditional electronic fire alarms. Its layered structure effectively impedes oxygen flow, achieving an oxygen-limiting index of 50%, and thus ensuring excellent flame-retardant performance. This design delays the combustion peak and reduces the heat-release value, thereby enhancing the flame-retardant performance. This work demonstrates the effective integration of a structural and functional design for active early fire warning and passive flame retardancy, paving the way for structural materials in advanced fire-warning systems in challenging environments.",
"conclusion": "CONCLUSION In summary, the prepared NAC combines strong mechanical robustness with active early fire-warning and passive flame-retardant functions. During the sintering process, Cr 3+ migrates into Al 2 O 3 MPs, forming solid-solution assembly units. These units impart thermochromic properties to the NAC, enabling it to undergo rapid color changes in response to elevated temperatures. The bioinspired multiscale BM structure, enhanced by inorganic mineral bridges, synergistically improves the strength and toughness of the NAC. The solid-solution strengthening effect increases the interfacial modulus of Al 2 O 3 MPs, reducing transcrystalline rupture and enhancing fracture toughness. By leveraging the thermochromic property, we utilize the trained K-means clustering model to process color variations digitally, enabling accurate and sensitive fire detection. Additionally, the layered structure provides effective flame retardance and reduces the generation of fire-related gases, addressing the later stages of combustion. This bioinspired design strategy provides a platform for creating structure–function integrated materials that are suitable for application in complex and dynamic environments.",
"introduction": "INTRODUCTION Advanced structural materials are extensively utilized in fields such as aerospace, construction and automotive manufacturing, in which they are often required to exhibit a combination of light weight, high strength and superior toughness [ 1–4 ]. Mimicking of the complex hierarchical structures of natural biomaterials has been shown to achieve these seemingly conflicting mechanical properties [ 5 , 6 ]. For instance, replicating the highly ordered organic–inorganic ‘brick-and-mortar’ (BM) microstructure and the reinforcing effect of mineral bridges that is found in nacre can result in structural materials with enhanced strength and toughness [ 5 , 7–10 ]. As the demand for multifunctional applications continues to grow, these materials must also possess additional functionalities to adapt to complex environmental changes [ 11 , 12 ]. In certain fire-prone scenarios, for example, the high temperatures that are generated by uncontrolled fires can significantly degrade the mechanical performance of structural materials, posing risks of equipment failure or structural collapse [ 13 ]. However, the inherent limited thermal resistance of the organic components in biomimetic composite materials constrains their overall high-temperature performance. This necessitates the addition of fire-alarm and flame-retardant systems, which may compromise reliability due to increased complexity [ 13–15 ]. Therefore, it is of great practical significance to develop structural materials with both active early fire-warning and passive flame-retardant properties. It is crucial to select active fire-warning and passive flame-retardant mechanisms that can be integrated with the intrinsic properties of structural materials. Recently, thermochromic materials combined with image-recognition systems have demonstrated promising potential in early fire-warning applications [ 16 ]. By leveraging the thermochromic response that is triggered by high-temperature heat sources, these systems can generate immediate alarm signals, facilitating timely detection and response to fire hazards [ 16–18 ]. Compared with traditional fire-warning systems, such as smoke alarms, this approach addresses issues of low sensitivity, delayed response times and high false-alarm rates [ 14 , 19–21 ]. Additionally, these systems do not rely on external electrical circuits, thereby reducing the impact of complex environments on their performance [ 14 , 16 ]. For passive flame-retardant mechanisms, the aforementioned nacre-inspired BM microstructure has been demonstrated to possess inherent fire-prevention and flame-retardant capabilities [ 22–24 ]. The primary flame-retardant mechanism is attributed to the barrier effect of the lamellae, which impedes heat transfer, the diffusion of pyrolysis products and the mixing of oxygen [ 25 , 26 ]. However, at this stage, the incorporation of thermochromic functionality into structural materials presents a significant challenge. Herein, we propose a strategy that combines an atomic-doping design with a biomimetic structural design to prepare nacre-mimetic alumina–cyanate resin composites (NACs) with mechanical robustness, thermochromic properties and flame-retardant functionality. Through a controlled solid-solution reaction, chromium-doped alumina microplatelets (Cr-doped Al 2 O 3 MPs) are synthesized to serve as ‘bricks’ in the nacre-inspired BM structure. These microassembly units exhibit unique reversible thermochromic properties. Simultaneously, this atomic-doping strategy facilitates solid-solution strengthening [ 27–30 ], which synergizes with the BM structural design, endowing the NAC with light weight (∼2.3 g cm –3 ), high flexural strength (∼290.1 MPa) and high fracture toughness (∼11.1 MPa m 1/2 ). Moreover, the NAC maintains 62.8% flexural mechanical strength (∼182.4 MPa) at 250°C, significantly surpassing that of layered ceramic scaffolds (∼34.8 MPa) due to the infiltration of high-temperature-resistant cyanate resin (CE). Through image-recognition that is based on the trained K-means model [ 31 , 32 ], the response time of the NAC is 9 s at 250°C for the early fire warning. In addition, the highly ordered BM structure effectively obstructs oxygen conduction, achieving an oxygen-limiting index of 50%, and thus imparting excellent flame-retardant properties to the NAC. This mechanically robust bioinspired composite integrates active early fire-warning and passive flame-retardant strategies, making it a promising smart structural material for fire-warning systems in various harsh environments.",
"discussion": "RESULTS AND DISCUSSION Design and fabrication of NACs The integration of microstructured building blocks with thermochromic properties into a nacre-inspired BM structure is a fundamental approach for constructing NACs. We propose a strategy that combines an atomic-doping design with a biomimetic structural design to realize this vision. Cr 3+ exhibits different electronic transition phenomena in its d orbitals at varying temperatures [ 33–35 ]. The thermochromic properties of 2D microassembly units can be achieved through controlled Cr 3+ doping. Additionally, the solid-solution strengthening effect enhances the mechanical strength of these Al 2 O 3 MPs. By utilizing these thermochromic Cr-doped Al 2 O 3 MPs as ‘bricks’ and the high-temperature-resistant polymer as the ‘mortar’, we aim to construct a nacre-inspired BM structure that simultaneously achieves excellent mechanical performance and flame retardancy in NACs. Specifically, we prepared NACs through a bottom-up assembly process, as illustrated in Fig. 1a . Nacre-inspired films were produced by using evaporation-induced self-assembly. This process employed Al 2 O 3 MPs, silicon dioxide nanoparticles (SiO 2 NPs) and chromium oxide nanoparticles (Cr 2 O 3 NPs) as inorganic components, with bacterial cellulose nanofibers (BCNFs) serving as the organic component (Fig. 1b and Fig. S1 ). The BCNFs network trapped Al 2 O 3 MPs and NPs, forming homogeneous films ( Fig. S2 ). The green nanocomposite films were laminated and sintered to construct layered ceramic scaffolds ( Fig. S3 ). The sintered SiO 2 NPs acted as mineral bridges connecting Al 2 O 3 MPs, while Cr 2 O 3 NPs reacting with Al 2 O 3 MPs to form Cr-doped Al 2 O 3 solid solution at a high temperature of 1500°C. Then, the ceramic scaffolds underwent surface treatment with N- [3-(trimethoxysily)propyl]ethylenediamine (Z6020) ( Fig. S4 ). After infiltration with CE, we obtained a densified, multiscale-structured bulk composite (Fig. 1c and Fig. S5a ). The orientation degree of Cr-doped Al 2 O 3 MPs in the bioinspired bulk composite was quantified as 89.9% by using 2D small-angle X-ray scattering analysis ( Fig. S5b ), indicating a high degree of orientation in the NACs. As shown in Fig. 1d , the distribution of individual elements in the bulk composite reveals that the elemental Si is located between the elemental Al, confirming the successful formation of mineral bridges between Al 2 O 3 MPs. Additionally, the overlap of the elemental Cr with the elemental Al preliminarily proves the formation of the solid solution (Fig. 1d ). Figure 1. Fabrication and microstructure characterization of the NAC. (a) Schematic illustration of multiscale structure design and fabrication of the biomimetic bulk composite. (b) Cross-sectional scanning electron microscope (SEM) image of the layered nanocomposite film. The upper-left insert shows a photograph of the nanocomposite films. (c) Photograph of the NAC. (d) SEM elemental maps of the NAC. Thermochromic mechanism of NACs The thermochromic properties of NACs are attributed to the atomic-doping design of the microassembly units. To characterize the atomic structure of Cr-doped Al 2 O 3 MPs, we mixed Al 2 O 3 MPs and Cr 2 O 3 NPs at a weight ratio of 1:0.1, consistently with that used in NACs, and then annealed them at 1500°C. Notably, the MPs powder turned pinkish after annealing ( Fig. S6 ). Furthermore, we cut the Cr-doped Al 2 O 3 MP by using a focused ion beam (FIB) and observed Cr diffusion into the MP (Fig. 2a and Fig. S7 ). High-resolution transmission electron microscopy (HRTEM) images revealed that the lattice of the Cr-doped Al 2 O 3 MP was slightly larger than that of the undoped Al 2 O 3 MP ( Fig. S8 ). To better observe the Cr monoatomic state, we determined that the crystal plane that was parallel to the surface of the single-crystal Al 2 O 3 MP was the (006) crystal plane ( Fig. S9 ). The transmission electron microscopy (TEM) sample was then prepared from vertical (006) crystal planes (Fig. 2b ), parallel to the c -axis of the crystal, by using the FIB technique [ 36 ]. We employed aberration-corrected (AC) high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) with energy-dispersive X-ray spectroscopy (EDS) to reveal the precise location of the Cr atoms in the Cr-Al 2 O 3 solid solution, in which two high-density bright spots corresponded to Cr single atoms (Fig. 2c and Fig. S10 ). The integrated pixel intensity profiles further confirmed the substitution of Al single atoms with Cr single-atom dopants (Fig. 2d and Fig. S11 ). Furthermore, the shift of the Cr 2 p peak in the X-ray photoelectron spectroscopy (XPS) spectra verified the different chemical environments for Cr atoms before and after doping (Fig. 2e ). To determine specific substitution sites, we analysed the high-resolution XPS spectra of Al 2 p and O 1 s . The Al 2 p peaks of the doped sample shifted towards lower binding energies (binding energy difference ∼0.3 eV) (Fig. 2f ), suggesting that the Cr substitution for Al strengthened the average ionicity between the Al and the O in the doped sample [ 37 ]. Additionally, the binding energy of O also decreased by ∼0.3 eV compared with the undoped sample, attributed to the donor-doping effect [ 38 , 39 ] induced by the Cr, which donates extra electrons to the system ( Fig. S12 ). Therefore, the incorporation of Cr atoms significantly altered the chemical environments within the Al 2 O 3 crystal. Figure 2. Thermochromic mechanism of NACs. (a) TEM elemental maps of the cross-sectional Cr-doped A 2 O 3 MP. Scale bars, 200 nm. (b) Illustration of the Cr-doped Al 2 O 3 MP cut by FIB parallel to the c -axis of the crystal shows the (006) crystal plane. (c) Atomic-resolution HAADF-STEM image and the corresponding EDS mapping of the Cr-doped Al 2 O 3 MP, taken with the c -axis of the crystal parallel to the electron beam. Scale bars, 0.2 nm. (d) Intensity profiles of elemental Cr along the white line in (c). (e and f) High-resolution XPS spectra of (e) Cr 2 p and (f) Al 2 p . (g) Photographs of the thermochromic NAC from 25°C to 250°C. Scale bars, 1 cm. (h) In situ heated HRXRD profiles of Cr-doped Al 2 O 3 MPs powder. (i) Lattice parameters of the Cr–Al 2 O 3 solid solution at 25°C and 250°C. (j) In situ heated Raman spectra and (k) UV–vis absorption spectra of the Cr-doped layered ceramic scaffold. (l) Calculated band gap of the Cr-doped layered ceramic scaffold from Tauc's relationship as a function of the temperatures of the Cr-doped layered ceramic scaffold. By using Cr-doped Al 2 O 3 MPs as the assembly building blocks, we successfully endowed layered ceramic scaffolds and biomimetic bulk composites with thermochromic properties (Fig. 2g and Fig. S13 ). X-ray diffraction (XRD) spectra revealed peaks of a new phase near the Al 2 O 3 crystal peaks, attributed to a Cr-doped Al 2 O 3 solid-solution crystal [ 40 ] ( Fig. S14a–c ). Meanwhile, the addition of SiO 2 to the ceramic scaffolds had no effect on this crystal phase change ( Fig. S14d ). To explore the thermochromic mechanism, we utilized in situ heating high-resolution XRD (HRXRD) to obtain the diffraction peaks shift data of the Cr-doped Al 2 O 3 MPs powder. The diffraction peaks shifted to lower angles (Fig. 2h ), indicating lattice expansion with increasing temperature. The change in the lattice parameters at 25°C and 250°C were calculated by using the Rietveld refinement method [ 41 , 42 ] to validate this observation (Fig. 2i and Fig. S15 ). The colors of the transition-metal complexes are derived from electronic transitions between d orbitals that are split by the ligand field [ 35 ]. This splitting of the d orbitals is also the origin of the red color in the Cr-doped Al 2 O 3 [ 35 ]. At high temperatures, lattice expansion reduces the constraint of the O 2 – anions on the Cr 3+ valence orbitals, allowing the material to regain its original green color based on the ligand field theory of transition-metal complexes [ 35 , 43 ]. Furthermore, the temperature-dependent evolution of the bond length and the lattice dimension was investigated by using in situ Raman spectroscopy from 25°C to 300°C upon heating. The Cr–O stretching vibration exhibited an intensity peak at 810 cm –1 ( Fig. S16a and b ). As the temperature increased, the Cr–O stretching vibration intensity peak redshifted (Fig 2j and Fig. S16c ). This was attributed to Cr–O bond elongation, further demonstrating the temperature-dependent evolution of the bond length and lattice dimensions for color change [ 34 , 44 ]. In addition, ultraviolet-visible (UV–vis) absorption spectroscopy was carried out at different temperatures to quantitatively characterize the thermochromic properties. The absorption spectra (Fig. 2k and Fig. S17a ) revealed two broad bands in the visible range, at 370–430 and 530–580 nm. These bands are associated with the d–d electronic transitions of Cr 3+ : 4 A 2 g → 4 T 1 g (370–430 nm) and 4 A 2 g → 4 T 2 g (530–580 nm) ( Fig. S17b–d ) [ 45 ]. The ligand field theory for Cr 3+ in an octahedral environment predicts the existence of three absorption bands [ 45 ]. The energies of the first two electronic spins allowed the transitions 4 A 2 g → 4 T 1g (F) and 4 A 2 g → 4 T 2g (F), which corresponded to visible light [ 43 , 45 ], whereas the third spin allowed the transition from 4 A 2 g to 4 T 1g (P), corresponding to ultraviolet light, which does not affect the color [ 43 , 45 ]. The ligand field that was created by the oxide ions allows adjustment of the positions of these absorption bands, resulting in different colors at varying temperatures (Fig. 2k ). Moreover, the optical band gap of the layered ceramic scaffolds was determined by extrapolating the linear region of the absorption edge to the energy–axis intercept ( Fig. S18 ). As the temperature increased from 25°C to 300°C, the optical band gap reduced from 4.22 to 4.09 eV (Fig. 2l ). This narrowing of the band gap was likely due to deformation of the Cr–O polyhedron and consistent with the observed color change from pinkish to gray as the temperature rose [ 34 , 44 , 46 ]. Mechanical properties of NACs The hierarchical architectures of nacre-mimetic materials play a dominating role in their exceptional mechanical properties. To evaluate the effectiveness of the multiscale structure, we systematically studied the mechanical properties of NACs. We selected a weight ratio of Cr 2 O 3 NPs to Al 2 O 3 MPs of 1:0.1, considering the significant color change and high mechanical strength at this ratio ( Figs S19 and S20 ). Subsequently, by altering the content of the SiO 2 NPs, we achieved an optimal flexural strength of ∼290.1 MPa for the NAC (Fig. 3a and Fig. S21 ). This is mainly due to the optimization of the layered orientation degree of the bioinspired composite at the microscopic level ( Fig. S22 ). Moreover, the number of mineral bridges connecting the Cr-doped Al 2 O 3 MPs increased with the weight ratio of the SiO 2 NPs, which also led to enhanced bending strength of the layered ceramic scaffolds ( Fig. S23a ). Besides, the inorganic content of the NACs increased ( Fig. S23b ). These factors also contributed to the increased modulus of the bioinspired bulk composites (Fig. 3a ). Furthermore, comparison of the flexural strength of the layered ceramic scaffolds without SiO 2 and Cr 2 O 3 NPs incorporation confirmed that sintered SiO 2 , which connects Cr-doped Al 2 O 3 MPs, significantly contributes to the mechanical support of layered ceramic scaffolds ( Fig. S24 ). In addition, the optimal NAC exhibited advantages in mechanical strength and specific strength compared with engineering Al 2 O 3 ceramics and synthetic CE ( Fig. S25 ). Figure 3. Macromechanics of biomimetic bulk structural materials. (a) Flexural strength and Young's modulus of NACs. (b) Fracture toughness for crack initiation ( K Ic ) and stable crack propagation ( K Jc ) of NACs, CE and Al 2 O 3 ceramics. (c) Ashby diagram of specific strength and specific toughness for NACs compared with a series of engineering materials. (d) DIC maps of the NAC during the crack propagation. (e) Long-range crack deflection. (f) Relative displacement of Al 2 O 3 MPs derived from shear loading. (g) Al 2 O 3 MPs pull-out with crack growth. The fracture toughness K Ic , indicating resistance to crack initiation, was ∼3.7 MPa m 1/2 for the optimal NAC, which was slightly lower than that of engineering Al 2 O 3 ceramics (∼4.5 MPa m 1/2 ) but higher than that of CE (∼2.19 MPa m 1/2 ) (Fig. 3b ). The maximum fracture toughness K Jc of the NAC (∼11.1 MPa m 1/2 ) was measured to be approximately three times higher than the initial resistance value, far exceeding those of engineering Al 2 O 3 ceramics (∼4.7 MPa m 1/2 ) and CE (∼2.21 MPa m 1/2 ) (Fig. 3b ). These results illustrate that the high strength and toughness of the bioinspired composite are attributed to its hierarchical structure. Figure 3c and Table S1 show that the specific strength and specific toughness of the hierarchical NAC surpass those of various engineering materials [ 47–49 ]. Multiple extrinsic toughening mechanisms effectively resist crack growth, primarily operating in the crack wake [ 5 ]. These mechanisms were further investigated via fracture mechanics analysis. The fracture resistance and deformation mechanisms were evaluated by using the notched NAC and confirmed by 2D digital image correlation (DIC). A significant deflection phenomenon was observed from crack initiation to expansion, with the strain level at failure reaching as high as 13.87% (Fig. 3d and e ). Besides, X-ray tomography imaging showed that specimen cracks propagated across different planes ( Fig. S26 ). Those results proved that the fracture propagation process of the NAC was stable and long-lasting, and exhibited characteristics of quasi-plastic fracture [ 50 , 51 ]. In contrast, CE and engineering Al 2 O 3 ceramics did not exhibit significant crack deflection due to their isotropic and homogeneous structure ( Figs S27 and S28 ). Furthermore, the failure strain of the Al 2 O 3 ceramic was only 1.59%, indicating that it was a brittle fracture ( Fig. S28 ). In addition, the state of crack growth was observed by using a combination of a high digital camera and an optical microscope. The structural design of the NAC can effectively prevent crack growth and prolong the path and duration of crack growth ( Movies S1–S3 ). Higher-magnification SEM images illustrate the relative displacement of individual MPs and their pull-out from the organic phase due to local tensile stress and interfacial shear stresses (Fig. 3f and g , and Fig. S29 ), which are important micromechanisms for energy absorption to improve toughness [ 52 , 53 ]. These results confirm that the multiscale microstructure contributes to the load redistribution and toughness enhancement of NACs. To further explore the microscopic mechanics of the NAC, we conducted nanoindentation tests to measure the modulus and hardness of its components. The modulus of Cr-doped Al 2 O 3 MPs, CE and SiO 2 mineral bridges were measured as ∼248.0, ∼10.2 and ∼90.0 GPa, while their hardness was tested as ∼34.07, ∼0.50 and ∼12.72 GPa, respectively (Fig. 4a and Fig. S30 ). The stacked microstructure, composed of alternating soft and hard phases, similarly to natural nacre, is a key factor in achieving the light weight, high strength and high toughness of the NAC (Fig. 4b and c ) [ 5 , 54 , 55 ]. To better detect the interface mechanics of Cr-doped Al 2 O 3 MPs, we conducted atomic force microscopy (AFM) to obtain modulus mapping (Fig. 4d ). From the magnification modulus mapping, we found that the modulus of the Cr-doped Al 2 O 3 MPs interface was higher than that of the Al 2 O 3 MPs themselves, due to the solid-solution strengthening effect (Fig. 4e and Fig. S31 ). Moreover, this effect facilitates the reduction of transgranular fractures in brittle inorganic MPs, thereby enhancing toughness [ 56 , 57 ]. In contrast, in the composites without doped Cr 2 O 3 NPs, the modulus of the Al 2 O 3 MPs was the highest, causing them to rupture more easily compared with Cr-doped MPs during crack propagation (Fig. 4f and g ). Furthermore, the fracture toughness of the NACs was higher than that of undoped ones (Fig. 4h ). Simultaneously, the extensive rising crack-resistance curves ( R -curves) exhibited behavior that was similar to those of natural nacre [ 5 ], indicating enhanced resistance to fracture during crack propagation (Fig. 4i ). These results demonstrate that the atomic-doping design not only imparts thermochromic properties to NACs, but also enhances their fracture toughness. Figure 4. Micromechanics of NACs related to microstructure. (a) Microstructure modulus and hardness of NACs. (b and c) Maps of the (b) hardness and (c) modulus of the NAC. (d) Elastic modulus map of the NAC based on AFM measurements. (e) Magnifying elastic modulus map in (d). (f) Elastic modulus map of the NAC without doped Cr 2 O 3 NPs. (g) Magnifying elastic modulus map in (f). (h) Comparison of facture toughness of NACs with doped and undoped Cr 2 O 3 NPs. (i) R -curves showing the resistance to fracture in terms of the stress intensity, K Jc , as a function of the crack extension, Δ a, for NACs with doped and undoped Cr 2 O 3 NPs. NACs for the early fire-warning system Ideal thermochromic materials for early fire warning should exhibit sensitive thermochromic responses prior to flame emergence and effective flame-retardant properties thereafter (Fig. 5a ). The ignition temperatures of common combustibles in a fire range from 250°C to 500°C [ 13 ]. Therefore, it is crucial for NACs to ensure a thermochromic response and mechanical stability below the ignition temperatures of various combustible polymeric materials. As shown in Fig. S32a , the dimensions of the NAC and ceramic scaffold showed minimal variation with temperature whereas CE exhibited significant changes as the temperature increased. Due to the anisotropy of the layered structure, the out-of-plane thermal conductivity improved significantly following resin infiltration ( Fig. S32b ). This enhancement facilitates heat dissipation, thereby accelerating the speed of color change. Furthermore, we conducted dynamic mechanical analysis (DMA) to explore the storage modulus, loss modulus and mechanical loss angle (tanδ) as the temperature increased at a frequency of 1 Hz. After CE was infiltrated into the layered ceramics to form NACs, both the storage energy modulus and loss modulus decreased as the temperature increased, with the peak tanδ value occurring at 276°C ( Fig. S33 ). This indicates that temperatures of ≤250°C, which are exactly under the ignition temperatures of common combustibles for fire warnings, can be considered the safe operating range for NACs. The ambient bending strength of NACs decreased as the temperature increased (Fig. 5b ). They can be employed as engineering materials with a flexural strength of 182.4 MPa at 250°C (Fig. 5b ). However, the bending strength decreased significantly to 84.9 MPa at 300°C, posing threats to practical use (Fig. 5b ). Meanwhile, the color change of the NAC corresponded to the trend observed in crystal heating color changes at temperatures of <250°C but deviated from this trend at 300°C (Fig. 5c and Figs S34 and S35 ). There is a distinct color difference in the NAC at 25°C and 250°C (Fig. 5d ), making it suitable for early fire-warning signals. Figure 5. Application of NACs for the early fire-warning system. (a) Schematic illustration of NACs for the active early fire warning and passive fire retardance. (b) Ambient flexural strength of NACs at different temperatures. (c) Reflectance spectra of the NAC under different temperatures. (d) International commission on illumination (CIE) color coordinates map of NACs at 25°C and 250°C. (e) Color information processing based on K-means model. (f) 2D H–S color histograms of the NAC at different temperatures. The color bar values represent pixel frequency. (g) Clustering of images based on H–S features. (h) Comparison of response time of the NAC with the thermochromic material reported in reference (PMS refers to phthalocyanines precursor molecular sensor) [ 16 ]. (i–k) Comparison of flame retardance of the NAC and CE in terms of LOI (i), heat-release rate (j) and smoke-production rate (k). To describe the color change quantitatively, the HSV (Hue, Saturation, Value) color model is employed to transform the information into a recognizable format by the computer. By extracting the HSV features of image colors at different temperatures, the trained K-means clustering model is used to process and classify color information (Fig. 5e ). Figure 5f and Figs S36 and S37 show the 2D color histograms that were extracted at different temperatures, corresponding to the H and S values. Furthermore, the H–S features of the NACs were analysed by clustering at different temperatures (Fig. 5g ). Temperatures of >250°C were defined as hazardous and the input of an image with colors that corresponded to these temperatures triggered a fire warning (Supplementary Movie S4 ). Moreover, we put NACs on the heat surface to simulate the high temperature that may occur during the precombustion phase of a fire-response process. The early-warning temperature thresholds are associated with specific colors, corresponding to defined temperature ranges. The sensitivity was then evaluated by measuring the time required for NACs to change color. By utilizing the trained image-recognition algorithm, the response time at various temperatures is determined based on the categorized HSV color features. Thus, the response time of the NAC is 11, 10 and 9 s at 150°C, 200°C and 250°C, respectively, demonstrating its high sensitivity at warning temperatures (Fig. 5h ) compared with another thermochromic material that was reported in the literature [ 16 ]. Then, the accuracy of the image-recognition program was verified by using a thermal infrared camera ( Figs S38–S40 ). Therefore, thermochromic NACs combined with intelligent image-recognition technology provide rapid and accurate monitoring for early fire-warning detection. Besides active fire warning, NACs possess passive fire-prevention and flame-retardant functions under high-temperature ablation during the later stage of fire. The limiting oxygen index (LOI) test shows that the LOI value of NACs is 50%, compared with 26% for CE, attributed to the barrier effect of their layered structure (Fig. 5i ). Furthermore, we utilize cone calorimetry testing to study the combustion behavior of the NAC in a real fire scenario. After testing, the NAC maintained its morphology integrity whereas the CE was fully carbonized ( Fig. S41a ). The peak heat-release rate of CE is ≤493 kW m –2 whereas that of the NAC decreases to 65 kW m –2 (Fig. 5j ). Analysis of the heat-release rate and total heat-release curves indicates that delayed combustion peaks and decreased heat-release values enhance flame retardation (Fig. 5j and Fig. S41b ). Simultaneously, the NAC releases significantly less smoke than CE, to enhance fire safety (Fig. 5k and Fig. S41c ). Thus, highly ordered inorganic MPs in NACs not only provide excellent mechanical robustness, but also achieve superior flame retardance."
} | 7,767 |
36991829 | PMC10058286 | pmc | 19 | {
"abstract": "Memristors mimic synaptic functions in advanced electronics and image sensors, thereby enabling brain-inspired neuromorphic computing to overcome the limitations of the von Neumann architecture. As computing operations based on von Neumann hardware rely on continuous memory transport between processing units and memory, fundamental limitations arise in terms of power consumption and integration density. In biological synapses, chemical stimulation induces information transfer from the pre- to the post-neuron. The memristor operates as resistive random-access memory (RRAM) and is incorporated into the hardware for neuromorphic computing. Hardware composed of synaptic memristor arrays is expected to lead to further breakthroughs owing to their biomimetic in-memory processing capabilities, low power consumption, and amenability to integration; these aspects satisfy the upcoming demands of artificial intelligence for higher computational loads. Among the tremendous efforts toward achieving human-brain-like electronics, layered 2D materials have demonstrated significant potential owing to their outstanding electronic and physical properties, facile integration with other materials, and low-power computing. This review discusses the memristive characteristics of various 2D materials (heterostructures, defect-engineered materials, and alloy materials) used in neuromorphic computing for image segregation or pattern recognition. Neuromorphic computing, the most powerful artificial networks for complicated image processing and recognition, represent a breakthrough in artificial intelligence owing to their enhanced performance and lower power consumption compared with von Neumann architectures. A hardware-implemented CNN with weight control based on synaptic memristor arrays is expected to be a promising candidate for future electronics in society, offering a solution based on non-von Neumann hardware. This emerging paradigm changes the computing algorithm using entirely hardware-connected edge computing and deep neural networks.",
"conclusion": "8. Conclusions Hardware-implemented neuromorphic computing based on memristor arrays was reviewed herein, including the operation mechanisms of single memristors as well as those of crossbar arrays for intelligent applications, such as pattern recognition, image processing, and AI chips. Moore’s law faces the challenge of attaining a decreasing physical scale in semiconductor processing while enhancing device performance. Emerging human-brain-inspired neuro-morphic computing aims to address the memory bottleneck associated with von Neumann architectures, which hinders memory storage and processing in big data and AI areas. Memristor-based RRAM is emerging as a promising candidate for overcoming the memory bottleneck, as it allows for high-density integration and energy-efficient memory processing using neuromorphic computing. Beyond the two-terminal memristor array, to satisfy multi-bit data storage, heterosynaptic plasticity is desired for tunable synaptic function, similar to that operating in the human brain. Although the memristor-based one-resistor (1R) RRAM structure is simple and capable of high-density integration, its synaptic modulation performance is poor. Therefore, one-transistor–one-resistor (1T1R) arrays should be introduced for complex AI computing and data processing tasks. Diffusive metal electrodes, such as Ag, Cu, and Ti, degrade memristive switching owing to residue filament formation in the channel. Even when external-substance-induced memristive switching occurs, the degraded endurance of RRAM critically limits practical applications. Thus, in addition to cation filament formation, vacancy migration or tunneling-based synaptic devices, which generate memristive switching inside the channel, also display significant potential for reliable neuromorphic computing. To mimic the human nerve system with memory applications, it is essential to understand the specific functionalities and characteristics of the nervous system, such as those of the soma, dendrites, and nodes of Ranvier. Using these components, advanced neuromorphic computing based on realistic bio-inspired artificial neural systems can be developed. Furthermore, integration with optic functionality broadens the potential applicability to include optogenetic and photo-induced memory. The fabrication of 2D-material-based large-scale RRAM arrays remains challenging owing to limited synthetic methods and low-yield processing. To demonstrate high-performance neuromorphic computing, fundamental advanced synthetic approaches that deliver high-uniformity in wafer-scale should be considered. The abovementioned strategies, namely, vacancy-induced migration, realistic nervous system imitation, and optogenetic integration, improve the functionality and endurance of present neuromorphic computing. Memristor-based neuromorphic computing offers significant potential and functionality for in-memory processing and edge computing and should therefore be investigated further for future AI technologies.",
"introduction": "1. Introduction According to Moore’s law, the performance of semiconductors will double every 24 months, and this has been maintained through the development of state-of-the-art foundry and chipmaker technologies [ 1 , 2 , 3 , 4 , 5 ]. However, the development of few-atom-scale semiconductor processes to achieve low-power operation with fast information processing has several limitations, including those imposed by Moore’s law [ 6 , 7 , 8 , 9 , 10 , 11 ]. Limitations of conventional computing technology include memory bottlenecks and high-cost energy processing (data processing between memory and processor). Moreover, the emerging artificial intelligence (AI) techniques require parallel information processing, big data analysis, and integrated systems entailing in-memory and on-chip computing [ 12 , 13 , 14 , 15 , 16 ]. However, conventional von Neumann architecture suffers memory bottlenecks as a result of continual data processing between the memory and processor, resulting in low-efficiency energy and low-speed memory processing [ 17 , 18 , 19 ]. Neuromorphic computing has been developed to overcome the memory bottleneck associated with von Neumann architecture [ 20 , 21 , 22 , 23 , 24 ]. Biological synapse-mimetic devices exhibit human-brain-like operations and perform information processing using electrical or optical spikes [ 25 , 26 , 27 ]. Various operational mechanisms, such as transistors, tunneling devices, and memristors, can be used for neuromorphic computing [ 28 , 29 , 30 ]. High-density integration with two-terminal memristors is emerging as a suitable approach for the fabrication of future devices characterized by low-power/low-thermal budgets and in-memory and on-chip computing [ 31 , 32 , 33 ]. A memristor combines the concepts of memory and resistors and exhibits resistive switching (RS) under an electrical bias resulting from the movement of anions and cations in materials. RS is generated by the formation of a conductive filament in the memristor [ 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 ]. When RS occurs from the high-resistance-state (HRS) to the low-resistance-state (LRS) owing to channel formation, the operation is called “SET”, whereas in reversed cases, it is termed “RESET” [ 43 , 44 , 45 , 46 ]. The RS operation mechanism can be classified depending on the materials, including transition metal dichalcogenides, transition metal oxides, boron nitride, silicon, or a layered combination of more than two materials [ 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 ]. A synapse array comprising memristors can provide synaptic characteristics and combine with a deep neural network for advanced data processing in neuromorphic computing. This review focuses on recent research regarding hardware-implemented neuromorphic computing using memristor arrays, covering aspects ranging from operational mechanisms to intelligent applications. A neurobiological synapse-mimetic memristor array providing RS under electrical or optical spikes is reviewed, including the material candidates, synapse operation and characterization, and neuromorphic computing processes for practical applications, such as image sensors, pattern recognition, and image or pattern processing. Critically, AI-embedded chips comprising memristor arrays overcome the memory bottleneck in conventional von Neumann architectures. A resistive random-access memory (RRAM) operation can provide low-power intelligent data processing by mimicking neurobiological synapses and will play a crucial role in future AI and memory computing applications. Scheme 1 compares the biological neuron system in the human brain with a memristor-based synapse array for human-brain-mimicking neuromorphic computing. In biological neurons, active potentials are created from the pre-synaptic to the post-synaptic neurons via Ca 2+ channels in the synaptic cleft. The post-synaptic neuron absorbs the ion into the ion channel receptor and generates a neural signal to transfer it to the next neuron. This process occurs in the nervous system of the human brain, which is capable of cognitive thinking and object detection using optical nerves. Compared to biological neurons, memristors operate by forming conductive filaments under electrical spikes and can mimic synapses by acting as a large-scale crossbar array. By integrating memristor arrays with artificial neural networks (ANNs), hardware-embedded human-brain-mimicking neuromorphic computing can serve as an efficient platform for emerging technologies, such as those implemented in image processing, pattern recognition, the Internet of Things (IoT), and other AI tasks. Below, the technologies for memristor-based AI are categorized according to devices and basic operations, synaptic behaviors and synapse arrays, and convolutional neural networks (CNNs) or optic-integrated image sensors."
} | 2,485 |
31409030 | PMC6723837 | pmc | 20 | {
"abstract": "The profound mutualistic symbiosis between corals and their endosymbiotic counterparts, Symbiodiniaceae algae, has been threatened by the increase in seawater temperatures, leading to breakdown of the symbiotic relationship—coral bleaching. To characterize the heat-stress response of the holobiont, we generated vital apo-symbiotic Euphyllia paradivisa corals that lacked the endosymbiotic algae. Using RNA sequencing, we analyzed the gene expression of these apo-symbionts vs. symbiotic ones, to test the effect of the algal presence on the tolerance of the coral. We utilized literature-derived lists of “symbiosis differentially expressed genes” and “coral heat-stress genes” in order to compare between the treatments. The symbiotic and apo-symbiotic samples were segregated into two separate groups with several different enriched gene ontologies. Our findings suggest that the presence of endosymbionts has a greater negative impact on the host than the environmental temperature conditions experienced by the holobiont. The peak of the stress reaction was identified as 28 °C, with the highest number of differentially expressed genes. We suggest that the algal symbionts increase coral holobiont susceptibility to elevated temperatures. Currently, we can only speculate whether coral species, such as E. paradivisa , with the plasticity to also flourish as apo-symbionts, may have a greater chance to withstand the upcoming global climate change challenge.",
"conclusion": "5. Conclusions Overall, our results suggest that the presence of algal symbionts can impair coral holobiont susceptibility to elevated temperatures. At the present stage, we can only speculate whether coral species with the plasticity to also flourish as apo-symbionts, such as E. paradivisa , may have a greater chance to withstand the upcoming global climate change challenge. Future studies should consider looking deeply into the effects of global climate change on the holobionts of symbiotic and apo-symbiotic corals.",
"discussion": "4. Discussion In this study, we analyzed the gene expression response following thermal stress exploring the effect of an algal symbiont presence on the transcriptome of the coral. The novelty of our case study approach lay in utilizing symbiotic and apo-symbiotic E. paradivisa corals as a model of a relatively tolerant coral [ 23 ]. Until now, studies of symbiotic and apo-symbiotic cnidarians have been designed to reveal symbiosis specificity and maintenance, and have not been utilized for a better and necessary understanding of corals under the threat of global climate change. One of our key findings was that the presence of algal symbionts greatly influences the coral host’s gene expression pattern, indeed, much more than the environmental temperature conditions that the holobiont is subjected to. This was demonstrated unambiguously in the PCA and hierarchical clustering ( Figure 2 ), which indicated two distinct clusters: symbiotic E. paradivisa and apo-symbiotic E. paradivisa samples. Moreover, the analysis of specific “symbiosis genes” and “heat-stress genes” supported this general trend; the differential expression of at least three times more genes could be correlated to the morph state than were related to the temperature treatment. Thus, the nature of the gene expression response to thermal stress would likely be determined firstly by the coral symbiotic state and presence, and not by the stringency of the heat. The identified 139 differentially expressed genes that were apparent in all treatments (common genes), constituted a general response to the heat, and were not restricted to specific temperature or morph. Interestingly, they behaved differently in symbiotic and apo-symbiotic corals (mostly down- and up-regulated in symbiotic and apo-symbiotic, respectively). This result further strengthened our finding of symbiotic and apo-symbiotic differential gene expression response to thermal stress. In addition to the common genes, there were differences in gene expression between different temperatures and morphs. In the symbiotic corals, the milder 28 °C temperature triggered the strongest gene expression response (samples 28S). The number of differentially expressed genes in the 28S samples was almost three times higher than in the apo-symbiotic preparations ( Figure 3 A,B). However, the response was in the form of acute shut down of genes, which may be the result of a cellular shock in the symbiotic E. paradivisa . Increasing the temperature to 31 °C appeared to moderate the gene expression response to the heat, in terms of number of genes and respective GOs, which was mostly diverted into up-regulation of genes. This result might indicate a severe cellular state, where the mechanisms to cope with the stress by activating stress genes are overwhelmed and irresponsive in a way. The observation of a higher number of DEGs at 28 °C than at 31 °C was also documented by RNA microarray analysis for the Red Sea coral Stylophora pistillata [ 16 ]. Another example where the heat-stress response (DEGs numbers) was not directly correlated with the time/temperature of coral exposure was described by Seneca and Palumbi (2015). These researchers reported that the transcriptome response of Acropora hyacinthus colonies was more significant after 5 hours of exposure to heated conditions than after 15 h [ 38 ]. The heat map of hierarchical clustering ( Figure 2 B) also supported this DEG behavior, as the 31S and 25S samples were more comparable to each other than either one was to the 28S sample. This pattern identified the peak of the stress reaction in the symbiotic coral as 28 °C, with remission at 31 °C. The peak temperature seems to be a critical threshold for this mesophotic coral, as in its natural habitat in the Red Sea, maximum summer temperatures do not exceed 28 °C [ 39 ]. Considering this behavior, it is not surprising that the 28S samples represented those most enriched for GOs ( Figure 5 A). Not only did the 28S samples contain up to six times more categories of GOs than any of the other treatments, but there were also more genes in each enriched category (thus, more significant). More specifically, there were two heat-stress-related GOs that were apparent only in the 28S samples; “cellular response to stress” and “protein ubiquitination.” Down-regulation of cell-cycle-associated genes in the 28S samples could indicate that the cells were in the process of cell cycle arrest or even cell degradation. Though all the treatments probably caused cellular oxidative stress (GO: cell redox homeostasis, oxidation-reduction), the symbiotic coral apparently suffered the strongest stress at 28 °C, as evidenced by the finding that the cellular response to stress was escalated to protein degradation (GO: protein ubiquitination) and to DNA damage (GO: DNA damage). Our results imply that not only did the presence of the photosymbiont influence the coral’s thermal resilience, but that it had a negative effect, as can be seen in both the “heat-stress genes” analysis ( Figure 4 B) and in the GO model ( Figure 5 ). Analysis of GOs demonstrated down-regulation of oxidative phosphorylation in both the 28S and 31S samples ( Figure 5 B). This is a common outcome of environmental stress, which disrupts ATP production in the mitochondria [ 12 ]. We postulate that the increased susceptibility of symbiotic E. paradivisa to oxidative stress may be largely attributed to the ROS-producing algal symbionts within their gastrodermal tissues. Following thermal stress, the chloroplasts of the photosymbiont may represent the prime source for ROS [ 13 ]. In fact, coral bleaching has been suggested as a survival technique, where the host expels the compromised ROS-producing photosymbionts and the symbiosis breaks down [ 40 , 41 ]. We have shown that in the massive Porites sp., the expression of stress genes was postponed after bleaching. Thus, it was suggested that expelling the algal symbionts enabled mitigation of the oxidative damage so that the associated stress genes could be activated at a later stage [ 42 ]. It is possible to postulate that a coral that can be vitally sustained as an apo-symbiont may be more resilient to thermal stress in that state. Some coral species have the capability as apo-symbionts to switch from acquiring fixed carbon primarily photoautotrophically to relying on alternative sources of fixed carbon: heterotrophic feeding and use of stored energy reserves [ 43 ]. Our experiments have shown that E. paradivisa has a heterotrophic plasticity required for long term durations without algal symbionts, and thus raise the question of whether the E. paradivisa –dinoflagellate relationship is more of a facultative nature, rather than obligatory. However, further research is needed in order to prove or disprove these hypotheses. Our results are also in agreement with the hypothesis proposed by Caroselli et al. [ 44 , 45 ] that non-symbiotic (asymbiotic) scleractinians may be more tolerant to temperature increase than symbiotic varieties. They measured variations in biometric parameters (polyp size, linear extension, calcification rate, skeletal mass, and skeletal density) and population density of symbiotic and asymbiotic corals along a latitudinal sea surface temperature (SST) gradient on western Italian coasts (Mediterranean Sea). While the asymbiotic coral seemed unaffected by SST and the biometric parameters did not change between the sites, the parameters of the symbiotic coral were inversely correlated to SST. The higher tolerance of the asymbiotic coral was attributed to the absence of photosymbionts, and thus freedom from inhibition of host physiological processes by the heat-stressed algal cells [ 44 , 45 ]. Several GOs were differentially enriched in symbiotic versus apo-symbiotic corals. Cellular carbohydrate biosynthetic processes were found to be enriched in the mild temperature apo-symbiotic samples ( Figure 5 ), which was probably related to metabolic changes in the apo-symbiotic host resulting from the lack of photosynthates regularly transferred from the phototrophic symbiont in symbiotic corals [ 46 ]. However, the pathways resulting in the breakdown of carbohydrate derivatives were enriched in the symbiotic morphs (31S), which benefited from the photosynthetic products translocated from the algal endosymbiont. Symbiosis-related GOs appeared only in symbiotic E. paradivisa samples, but, for some reason, these were temperature-dependent and were not consistent in the 28S and 31S samples ( Figure 5 ). Enrichment of the vacuole organization GO could be explained by the fact that algal symbiont is maintained distinct from the host cytoplasm within a ‘symbiosome,’ which is a host-derived vacuole [ 47 ]. Another set of GOs that were enriched only in the symbiotic samples and absent from the apo-symbiotic were those associated with reproduction and meiosis. These results agree with studies showing that bleaching profoundly affects the reproductive characteristics of corals (i.e., reduced numbers of eggs and testes, egg morphometry, and sperm motility) in both the short- and long-term [ 48 , 49 , 50 ]. Studies on gene expression responses of corals to environmental stress have largely focused on shallow-water corals [ 12 ]. Nevertheless, understanding these processes in mesophotic corals is crucial as well, given the forecasted degradation of shallow-water reef habitats [ 4 ]. Mesophotic coral reef ecosystems have been hypothesized to serve as a refuge for shallow and mid-depth species from various environmental disturbances (‘deep reef refuge’ hypothesis) [ 51 , 52 ]. However, to the best of our knowledge, there has been no experiment that has compared the tolerance of shallow-water and mesophotic corals to elevated temperatures under a controlled system. The scenario where mesophotic corals can thrive in the reef without their algal symbionts, possibly calcify their skeletons at a slow rate [ 53 , 54 ], be less susceptible to elevated temperatures, and regain the symbionts when conditions allow it, should be further explored in future studies."
} | 3,036 |
31468725 | PMC6922519 | pmc | 21 | {
"abstract": "Summary Microbial conversion offers a promising strategy for overcoming the intrinsic heterogeneity of the plant biopolymer, lignin. Soil microbes that natively harbour aromatic‐catabolic pathways are natural choices for chassis strains, and Pseudomonas putida \n KT 2440 has emerged as a viable whole‐cell biocatalyst for funnelling lignin‐derived compounds to value‐added products, including its native carbon storage product, medium‐chain‐length polyhydroxyalkanoates ( mcl ‐ PHA ). In this work, a series of metabolic engineering targets to improve mcl ‐ PHA production are combined in the P. putida chromosome and evaluated in strains growing in a model aromatic compound, p ‐coumaric acid, and in lignin streams. Specifically, the PHA depolymerase gene phaZ was knocked out, and the genes involved in β‐oxidation ( fad BA 1 and fad BA 2 ) were deleted. Additionally, to increase carbon flux into mcl ‐ PHA biosynthesis, phaG, alkK, phaC1 and phaC2 were overexpressed. The best performing strain – which contains all the genetic modifications detailed above – demonstrated a 53% and 200% increase in mcl ‐ PHA titre (g l −1 ) and a 20% and 100% increase in yield (g mcl ‐ PHA per g cell dry weight) from p ‐coumaric acid and lignin, respectively, compared with the wild type strain. Overall, these results present a promising strain to be employed in further process development for enhancing mcl ‐ PHA production from aromatic compounds and lignin.",
"introduction": "Introduction Lignocellulosic biomass offers a source of renewable carbon that can reduce reliance on fossil fuels, reduce greenhouse gas emissions and build a foundation for a sustainable bioeconomy. Recent studies have demonstrated that the co‐production of hydrocarbon fuels from sugars and chemicals from lignin streams increases the value proposition for biorefinery processes (Davis et al ., 2013 ; Ragauskas et al ., 2014 ; Corona et al ., 2018 ). By leveraging natural host metabolic capabilities and applying genetic engineering techniques, carbon from complex and heterogeneous substrates, such as lignin, can be funnelled into single products (Linger et al ., 2014 ; Beckham et al ., 2016 ). Of particular relevance to this work, the production of oleochemicals by native and engineered microbes has gained increased attention in the last decade due to the demand for more sustainable fuels and consumer and industrial products (Pfleger et al ., 2015 ). An example of these oleochemicals is medium‐chain‐length polyhydroxyalkanoates ( mcl ‐PHAs). These polymers can be used in the production of biodegradable plastics, medical devices and chemical and material precursors (Philip et al ., 2007 ; Linger et al ., 2014 ; Mozejko‐Ciesielska and Kiewisz, 2016 ; Prieto et al ., 2016 ; Chen and Jiang, 2018 ). The individual monomers comprising mcl ‐PHAs can range from C6 to C14 in chain length. mcl ‐PHAs are biosynthetic polyesters that can be produced from a wide range of carbon sources in many bacteria. Some of the most well studied mcl ‐PHAs producers are fluorescent Pseudomonads (Madison and Huisman, 1999 ; Prieto et al ., 2007 ). Among these Pseudomonads, Pseudomonas putida KT2440 naturally produces mcl ‐PHAs as a carbon storage compound in scenarios of carbon excess and nutrient limitation (de Eugenio et al ., 2010 ). This bacterium is genetically tractable and metabolically diverse, with many advantageous features for biorefinery processes (Nikel and de Lorenzo, 2014 ). P. putida can produce mcl ‐PHAs from multiple carbon sources such as fatty acids, which directly undergo β‐oxidation, or from sugars and aromatic compounds, which are subjected to fatty acid de novo biosynthesis (Prieto et al ., 2016 ) (Fig. 1 ). Specifically, metabolic intermediates in the fatty acid biosynthetic pathway are converted to (R)‐3‐hydroxyacyl‐ACP, which is hydrolyzed to a free hydroxy‐fatty acid and subjected to CoA ligation, via PhaG and AlkK, respectively, to produce (R)‐3‐hydroxyacyl‐CoA. Similarly, the β‐oxidation intermediate 2‐trans‐enoyl‐CoA can be hydrated to generate (R)‐3‐hydroxyacyl‐CoA by PhaJ (Tsuge et al ., 2000 ). These CoA monomers can then be polymerized into mcl ‐PHAs via the PHA synthases, PhaC1 and PhaC2 (Ren et al ., 2009 ). The reverse process, PHA degradation, can also occur in scenarios of complete carbon depletion or sudden nutrient increases and is conducted via the PHA depolymerase, PhaZ (de Eugenio et al ., 2010 ). Figure 1 The mcl ‐ PHA production pathway in P. putida \n KT 2440 via fatty acid biosynthesis and competing fatty acid β‐oxidation pathway. Red boxes indicate genes targeted for deletion, and green boxes indicate genes targeted for overexpression. AccA‐D, acetyl‐CoA carboxylase; FabD, malonyl CoA‐ ACP transacylase; FabH, 3‐ketoacyl‐ ACP synthase; FabG, 3‐ketoacyl‐ ACP reductase; FabA and FabZ, 3‐hydroxyacyl‐ ACP dehydratase; FabI and FabV, enoyl‐ ACP reductase; FabB and FabF, 3‐oxoacyl‐ ACP synthase; PhaG, hydroxyacyl‐ ACP acyl‐transferase; AlkK, acyl‐CoA‐synthase; PhaC1 and PhaC2, PHA polymerases; PhaZ, PHA depolymerase; PhaJ, R‐specific enoyl‐CoA hydratase; FadB, enoyl‐CoA hydratase / 3‐hydroxyacyl‐CoA dehydrogenase; FadA, 3‐ketoacyl‐CoA thiolase; FadE, acyl‐CoA dehydrogenase; FadD, long‐chain acyl‐CoA synthetase. Metabolic engineering has been applied to improve mcl ‐PHA production through various routes (Chen and Jiang, 2018 ) such as (i) shutting down competing pathways (β‐oxidation), (ii) overexpressing the PHA synthesis operon (via plasmid or chromosomal integration) with different ribosome binding sites (RBS) and/or promoters, (iii) enhancing NADH or NADPH supply for PHA synthesis, (iv) engineering cell morphology to increase cell size and (v) eliminating the ability to consume PHAs. For instance, to decrease the flux of PHA pathway intermediates to the fatty acid β‐oxidation pathway, the genes fadA and fadB were deleted in P. putida which resulted in a 2.5‐fold increase in mcl ‐PHA production (wt.% basis) when grown on nitrogen‐rich medium supplemented with heptanoate and octanoate (Wang et al ., 2011 ). In route (ii) above, the overexpression of phaG in the PHA synthesis operon of Pseudomonas jessenii resulted in a fourfold increase in mcl ‐PHA accumulation (wt.% basis) from phenylacetic acid (Tobin et al ., 2007 ). The effect of overexpressing other genes in combination with phaG on mcl‐ PHA production was also tested in E. coli . The expression of phaG and phaC1 (STQK) resulted in minimal mcl ‐PHA production from glycerol (0.9 mg l − 1 ), while the expression of phaG , phaC1 (STQK) and alkK increased mcl ‐PHA accumulation to 25 mg l − 1 when grown in the same conditions (Wang et al ., 2012 ). Another strategy to enhance PHA accumulation is to eliminate the degradation of the polymer through the deletion of the PHA depolymerase gene, phaZ . This deletion was evaluated in P. putida and resulted in a 1.9‐fold increase in mcl‐ PHA titre (g l − 1 ) and 1.3‐fold increase on PHA production (wt.% basis) when grown on octanoate under nitrogen‐limited conditions (Cai et al ., 2009 ) and in a 47% PHA increase (wt.% basis) when grown on glycerol as a sole carbon source (Poblete‐Castro et al ., 2014 ). De Eugenio et al . (de Eugenio et al ., 2010 ) also reported a similar result in a comparable background strains after 48 h of incubation utilizing octanoate as carbon source while Cai et al . (Cai et al ., 2009 ) showed a significant improvement in the knockout strain after 5 days of cultivation. Most of the work reported to date on metabolic engineering in Pseudomonads and other organisms to improve mcl ‐PHA production utilizes carbohydrates and oil sources as substrates, while only a few studies describe the use of aromatic compounds or lignin as a carbon source (Table 1 ). Furthermore, many reported strains employ plasmid‐based approaches, which limits applicability during scale‐up due to the need for antibiotic use to retain plasmids. In this study, we sought to engineer a base strain of P. putida to produce mcl‐ PHAs from aromatic compounds and lignin using genomic integration of DNA throughout. We demonstrate improved mcl ‐PHA production from p ‐coumaric acid ( p ‐CA) and a lignin stream that contains this aromatic compound as a major carbon source (Rodriguez et al ., 2017 ). The combination of overexpressing mcl‐ PHA synthesis genes and deleting mcl ‐PHA degradation genes increased p ‐CA and lignin conversion into mcl ‐PHA. Overall, this work presents a robust base strain whose background can be utilized for further process development and/or engineering to produce new compounds that utilize similar biosynthetic pathways. Table 1 Literature describing mcl ‐PHA production from lignin‐derived aromatic compounds and lignin streams by native and engineered bacteria Strain Substrate Antibiotic Cultivation mode Cultivation time (h) CDW (mg ml −1 ) \n mcl ‐PHA (mg l −1 ) \n % mcl ‐PHA yield (g per g CDW) References Native strains \n P. putida JCM13063 Vanillic acid – Batch, flask 72 210 Traces c \n < 1 Tomizawa et al . ( 2014 ) \n P. putida GPO1 p‐Coumaric acid – Batch, flask 72 270 Traces c \n < 1 Tomizawa et al . ( 2014 ) \n P. putida KT2440 p‐Coumaric acid Batch, flask 72 378 160 41 This study \n P. putida KT2440 p‐Coumaric acid – Batch, flask 48 470 160 34 Linger et al . ( 2014 ) \n P. putida KT2440 Ferulic acid – Batch, flask 48 436 170 39 Linger et al . ( 2014 ) \n P. putida KT2440 Lignin‐containing stream (corn stover) d \n Batch, flask 78 399 35 8.8 This study \n P. putida KT2440 Lignin‐containing stream (corn stover) d \n – Bioreactor, FB 48 787 252 32 Linger et al . ( 2014 ) Engineered strains \n P. putida A514 DVJ4C1 \n Kraft lignin T, G Batch, flask 40 \n b \n 70 \n b \n Lin et al . ( 2016 ) \n P. putida A514 ΔphaJ4/phaC1 \n Vanillic acid T, G Batch, flask \n b \n \n b \n \n b \n 73.5 Lin et al . ( 2016 ) \n P. putida \n Δxyl_alkKphaGC1 \n Vanillic acid T Batch, flask 50 715 246 34 Wang et al . ( 2018 ) \n P. putida KT2440 a \n Lignin‐containing stream (corn stover) d \n \n a \n \n a \n \n a \n 5300 1000 17.6 Liu et al . ( 2017 ) \n P. putida AG2162 p‐Coumaric acid – FB, flask 72 483 241 50 This study \n P. putida AG2162 p‐Coumaric acid – FB, flask, HCD 85 1758 953 54.2 This study \n P. putida AG2162 Lignin‐containing stream (corn stover) d \n – Flask, batch 78 654 116 17.7 This study CDW, cell dry weight; FB, fed‐batch; G, gentamicin; HCD, high‐cell density; T, tetracycline. \n a . The strain is not specified. In the materials and methods section, the authors specify the use of a native strain (in batch mode) while in their results authors stress the use of an engineered strain (in fed‐batch mode). \n b . Not reported. \n c . Not clear if authors analyzed mcl ‐PHAs or only polyhydroxybutyrate [P(3HB)]. \n d .The origin and preparation of these lignin streams is different in each case. John Wiley & Sons, Ltd",
"discussion": "Results and discussion Genetic modifications of P. putida to improve mcl‐PHA accumulation To improve mcl ‐PHA accumulation, P. putida KT2440 was genetically engineered to (i) eliminate mcl ‐PHA depolymerization, (ii) decrease flux of mcl ‐PHA pathway intermediates to fatty acid degradation and (iii) increase carbon flux from fatty acid chain elongation to mcl ‐PHA production (Fig. 1 ). To eliminate mcl ‐PHA depolymerization, the gene encoding the PHA depolymerase ( phaZ ; PP_5004) was deleted, resulting in strain AG2102 (Table 2 ). To decrease 3‐hydroxyacyl‐CoA flux towards fatty acid β‐oxidation, two previously identified chromosomal copies of the enoyl‐CoA hydratase/3‐hydroxyacyl‐CoA dehydrogenase ( fadB ) and 3‐ketoacyl‐CoA thiolase ( fadA ) genes (Liu et al ., 2007 ) were deleted. The fadBA1 genes are encoded at loci PP_2136‐2137 in a putative two‐gene operon. The second fadBA genes are clustered within loci PP_2214‐2217, where FadB is encoded at two separate coding regions – PP_2214 (3‐hydroxyacyl‐CoA dehydrogenase) and PP_2217 (enoyl‐CoA hydratase) – while FadA is encoded at PP_2215. This putative operon also encodes an acyl‐CoA dehydrogenase ( fadE ; PP_2216). These two gene clusters, PP_2136‐2137 and PP_2214‐2217, were deleted in strain AG2102, resulting in strain AG2228. Finally, to increase carbon flux from 3‐hydroxyacyl‐ACP to mcl ‐PHAs, an additional, codon‐optimized copy of the hydroxyacyl‐ACP thiolase ( phaG ; PP_1408), the hydroxyacyl‐CoA synthase ( alkK ; PP_0763), and the two PHA polymerases ( phaC1 and phaC2 ; PP_5003 and PP_5005), were integrated into the chromosome of AG2228 and overexpressed using the constitutive P \n tac promoter. These genes were inserted into the chromosome by replacing an acetaldehyde dehydrogenase ( aldB ; PP_0545) that is presumably not involved in aromatic catabolism, resulting in strain AG2162. Table 2 \n P. putida KT2440 genotypes and strain designations. Plasmids and strains used in this work were constructed using standard protocols as described in the Appendix S1 and as reported before (De Boer et al ., 1983 ; Johnson and Beckham, 2015 ; Kvitko and Collmer 2011 , Marx, 2008 ). Strain Genotype AG2102 \n P. putida KT2440 Δ phaZ \n AG2228 \n P. putida KT2440 Δ phaZ ΔfadBA1 ΔfadBAE2 \n AG2162 \n P. putida KT2440 Δ phaZ ΔfadBA1 ΔfadBAE2 ΔaldB::P \n tac \n ‐phaG‐alkK‐phaC1‐phaC2 \n John Wiley & Sons, Ltd Strain evaluation for mcl‐PHA production from the model aromatic compound p‐CA To determine if mcl ‐PHA accumulation was affected by the genetic modifications, these strains were grown in nitrogen‐limited medium containing 2 g l − 1 \n p‐ CA (12.2 mM) and 0.13 g l − 1 (NH 4 ) 2 SO 4 (1 mM). AG2228 and AG2162 presented longer growth lags than KT2440 and AG2102 (Fig. 2 A). Despite these initial growth profiles, p ‐CA maximum utilization rates were higher in the former strains (i.e. 0.15 ± 0.00 g l − 1 h − 1 in AG2162 and 0.10 ± 0.03 g l − 1 h − 1 in KT2440) and p ‐CA was nearly depleted at a similar time (48 h) in all the strains (Fig. 2 B). mcl ‐PHA titres (mg l − 1 ) at the sample collection time (72 h) only increased significantly in the engineered strain AG2162 (242.0 ± 9.8 mg l − 1 ) when compared to the wild type (157.8 ± 10.2 mg l − 1 ) (Fig. 2 C). In all cases, 3‐hydroxydecanoate (C10) and 3‐hydroxyoctanoate (C8) were the major mcl ‐PHA components produced, while 3‐hydroxydodecanoate (C12) and 3‐hydroxytetradecanoate (C14) were present in minor abundance, as expected (Fig. 2 C). The proportions of the four constituents were similar in the tested strains as well (as a percentage of total mcl ‐PHAs produced, 22‐26% C8, 66% C10, 7‐10% C12 and 1‐2% C14). The PHA yields (g mcl ‐PHA per g CDW) in AG2228 and AG2162 also exhibit significant increases compared with the wild type. Particularly, the yield increased from 41.9 ± 2.8% in wild type to 47.3 ± 1.2 and 49.8 ± 3.5% in AG2228 and AG2162, respectively. Figure 2 Production of mcl ‐ PHA s from p ‐ CA . A. Optical density at 600 nm ( OD \n 600 ) as a function of time and cell dry weight ( CDW ) at the end time point. B. p ‐ CA consumption profiles. C. mcl ‐ PHA titres and composition (bars), detected via depolymerization and derivatization to hydroxyacyl methyl esters ( HAME s), and mcl ‐ PHA yields (g mcl ‐ PHA per g CDW ) at 72 h (black circles) in four different strains. Results present the average of biological triplicates and error bars show the standard deviation. A statistical analysis (t‐test) was also performed for mcl‐ \n PHA titres and yields between the wild type and the engineered strains. *Significant difference at 95% confidence (see yields). **Significant difference at 99% confidence (see titres). Batch shake flask cultivations were performed in nitrogen‐limited modified M9 minimal medium ( pH 7.2) containing 0.13 g l − 1 (= 1 mM ) ( NH \n 4 ) 2 \n SO \n 4 and 2 g l − 1 \n p ‐ CA in triplicate. Cells were then washed in M9 (without carbon and nitrogen source) and the flasks were inoculated to an initial OD \n 600 of ~ 0.1 and incubated for 72 h. Samples for mcl ‐ PHA analysis were washed twice with distilled water and lyophilized for cell dry weight ( CDW ) measurements and PHA extraction. For mcl ‐ PHA production and composition analysis, samples were derivatized in BF \n 3 ‐methanol and quantified by gas chromatography‐mass spectroscopy ( GC ‐ MS ) as described in Appendix S1 . Although the deletion of phaZ did not improve titres or yields, this genetic background will be still advantageous in further process development, which requires longer cultivations subjected to carbon‐starvation. While the further deletion of fadBA1 and fadBAE2 in AG2228 led to a statistically significant increase in yields when compared to the wild type, titres did not improve. A similar result was previously reported in a different P. putida strain when only deleting fadA and fadB , but utilizing fatty acids as a carbon source (Wang et al ., 2011 ). The overexpression of phaG alone was previously reported not to affect P. putida mcl ‐PHA production from phenylacetic acid (Tobin et al ., 2007 ). Separately, overexpression of phaC1 combined with phaJ4 was sufficient to increase mcl ‐PHA accumulation from vanillic acid in a plasmid‐bearing P. putida strain (Lin et al ., 2016 ) (Table 1 ) while in E. coli, overexpressing the genes encoding PhaC and AlkK was necessary to enhance mcl ‐PHA accumulation from glycerol (Wang et al ., 2012 ). Even though the present study has not evaluated single overexpressed genes, we demonstrate that the selected gene combination (gene knockouts and gene overexpression integrated into the genome) significantly improves carbon flux from p ‐CA into mcl ‐PHA biosynthesis in P. putida . Evaluation of mcl‐PHA production by AG2162 under different culture conditions Carbon (C)‐to‐nitrogen (N) ratio (de Eugenio et al ., 2010 ) and cell density (Davis et al ., 2015 ) are known to affect mcl‐ PHA accumulation in P. putida . Thus, to obtain higher mcl ‐PHA titres and yields than those obtained in the previous experiment (Fig. 2 C), we evaluated AG2162 for mcl ‐PHA production at different C:N ratios and concentrations in a high‐cell density, fed‐batch, shake flask experiment. The experimental setup consisted of a batch phase containing either 4 or 8 g l − 1 \n p ‐CA as a carbon source with (NH 4 ) 2 SO 4 at different concentrations yet still nitrogen‐limited and a fed‐batch phase where the feeding contained only p ‐CA as a carbon source without any supplementary nitrogen (see Fig. 3 legend). Figure 3 Production of mcl‐ \n PHA by AG 2162 in fed‐batch mode at different C ( p ‐ CA , g l − 1 ):N (( NH \n 4 ) 2 \n SO \n 4 , mM ) ratios and concentrations in the batch phase, (1) 4:0, (2) 4:1, (3) 8:2, (4) 8:4, and fed‐batch phase (1) 2.5: 0, (2) 2.5:0, (3) 5:0, (4) 5:0. (A) Consumption of p ‐ CA and CDW , (B) mcl ‐ PHA yields, and (C) mcl ‐ PHA titres. The ‘inocula’ case corresponds to the seed culture data before inoculation. Results show the average of two biological replicates. Error bars present the absolute difference from the biological duplicate. These experiments were conducted in shake flasks. AG 2162 was precultured from glycerol stocks in modified M9 medium containing 2 g l − 1 \n p ‐ CA and non‐limiting nitrogen (10 mM ( NH \n 4 ) 2 \n SO \n 4 ) for 24 h. The preculture was then washed twice in M9 medium (without carbon or nitrogen), and inoculated at an OD \n 600 of 4 in modified M9 medium containing different carbon ( p ‐ CA ):nitrogen (( NH \n 4 ) 2 \n SO \n 4 ) ratios in the combinations mentioned above. When p ‐ CA was depleted (42 h), a pulse of 2.5 or 5 g l − 1 \n p ‐ CA was also applied to the combination (1,2) or (3,4) respectively. Flasks were incubated at 30°C and 300 rpm for 85 h and samples were taken at 42 and 85 h to evaluate CDW and PHA production. \n p ‐CA was depleted at the end of the batch phase and its concentration was < 0.75 g l − 1 at the end of the fed‐batch phase in all the cultivation conditions (Fig. 3 A). Maximum p ‐CA utilization rates (calculated after the feeding pulse) decreased at higher C:N ratios. Specifically, when nitrogen was absent from the media during the batch phase (case 1), p ‐CA utilization rate was 0.06 ± 0.00 g l − 1 h − 1 and, at the lowest C:N (case 4), the rate was 0.16 ± 0.01 g l − 1 h − 1 . Utilization rates were the same (0.11 ± 0.01 g l − 1 h − 1 ) in cases 3 and 4, which correspond to the same C:N ratio but at different initial substrate concentrations. Regarding CDW (Fig. 3 A), the highest values were observed in case 4, at the lowest C:N ratio. The mcl ‐PHA yields were similar at the end of the fed‐batch phase in all cases (between 48% and 54% with errors up to 3.2%) (Fig. 3 B), and case 4 presented the greatest CDW. Therefore, mcl‐ PHA titres were also higher in the latter case, up to 953 ± 44 mg l − 1 (Fig. 3 B). The average yields at the end of the batch phase increased at the highest C:N ratio (case 1, without nitrogen added). However, since nitrogen starvation limits cell growth, the titres were ultimately similar to those found under other culture conditions (cases 3 and 4). These results suggest that the C:N ratios evaluated in this study do not have a critical effect on mcl ‐PHA yields if produced in fed‐batch mode and high‐cell density cultivations. However, that ratio is critical to enhance cell biomass and thus titres and productivity. Production of mcl‐PHAs from a soluble and process‐relevant lignin‐rich stream As demonstrated above, wild type and engineered P. putida strains are able to convert the lignin‐derived product p ‐CA to mcl ‐PHAs effectively. Thus, to finalize this study, we also tested the ability of both strains to convert a heterogeneous lignin stream that contains p ‐CA, ferulic acid and high molecular weight lignin as major carbon sources, to mcl ‐PHAs. This lignin comes from the solid fraction generated after enzymatic hydrolysis of pretreated corn stover (Chen et al ., 2016 ). Then, it is further washed with water (to remove sugars) and solubilized via base‐catalyzed depolymerization (Rodriguez et al ., 2017 ; Salvachúa et al ., 2018 ). Lignin solubilization was approximately 53% (lignin content in soluble stream/lignin content in initial solid stream) and contained ~ 4 g l − 1 of p ‐CA and 0.1‐0.2 g l − 1 ferulic acid (as major aromatic compounds) from an initial total lignin content of approximately 22 g l − 1 . Wild type P. putida and strain AG2162 were grown in the lignin liquor (75% v/v containing 1 mM (NH 4 ) 2 SO 4 ) and reached stationary phase between 24 and 48 h, likely due to the total consumption of readily accessible carbon sources and/or nitrogen (Fig. 4 A). Strain AG2162 increased the mcl ‐PHA yield by ~ 100% compared with the wild type (17.7 ± 0.2 vs. 8.9 ± 0.8% respectively) and titre by 3.3‐fold (116 ± 35 vs. 35 ± 5 mg l − 1 respectively) (Fig. 4 B) which demonstrates the robustness of AG2162 and the increased carbon flux into mcl ‐PHA biosynthesis even in complex lignin streams. The main hydroxyacid species accumulated in both strains was again 3‐hydroxydecanoate. However, unlike the proportions observed in Fig. 2 C, 3‐hydroxyoctanoate was lower in these lignin cultures, representing 10% and 18% of the hydroxyacids in AG2162 and KT2440, respectively (Fig. 4 B), instead of 22‐26%. We also analyzed the lignin molecular weight profile by gel permeation chromatography (GPC) at the end of the cultivations (78 h). Low molecular weight lignin (indicated as monomeric aromatic compounds in Fig. 4 C) disappeared after the bacterial treatments, which aligns with the total p ‐CA depletion shown in Fig. 4 A. As observed in previous work (Salvachúa et al ., 2015 ), both strains also decreased the high molecular weight lignin content, although that decrease is more evident in AG2162 cultivations (Fig. 4 C) which suggests the conversion of oligomeric lignin. To confirm if high molecular weight lignin is metabolized by strain AG2162 to a higher extent than KT2440, we also analyzed the lignin content at the end of the bacterial treatments. Lignin utilization was 23.5 ± 1.7% and 18.3 ± 1.5% in AG2162 and KT2440, respectively, which verifies the GPC observations. Overall, these results corroborate that AG2162 is a robust and improved mcl ‐PHA production strain compared with KT2440 from both pure aromatic compounds, such as p ‐CA, and a process‐relevant lignin stream. Figure 4 Performance of wild type and AG 2162 P. putida strains in a process‐relevant soluble lignin stream. A. Bacterial density and p ‐ CA utilization over time, (B) mcl ‐ PHA titres (mg l − 1 ) and composition (bars), detected via depolymerization and derivatization to hydroxyacyl methyl esters ( HAME s), and mcl ‐ PHA yields (g mcl ‐ PHA per g CDW ) at 78 h (black circles) and (C) GPC lignin profiles after the bacterial treatments and in non‐inoculated lignin controls. Results show the average of two biological replicates with error bars representing the absolute difference. The cultivation conditions were the same as those presented in the legend of Fig. 2 except that in this case, the cultivations were performed in 250 ml baffled flasks containing 50 ml of medium (modified M9 plus 75% sterile soluble lignin stream). Non‐inoculated lignin cultures were used as a control. For lignin content, a compositional analysis was performed in freeze‐dried lignin supernatants according to the procedure in NREL LAP / TP ‐510‐42618 (Sluiter et al ., 2006 ). For molecular weight, gel permeation chromatography ( GPC ) analysis was also conducted on freeze‐dried samples (30 mg) as described before (Salvachúa et al ., 2016 ). The analysis of p ‐ CA in non‐lignin containing media was analyzed by high performance liquid chromatography ( HPLC ) on an Agilent 1100 series equipped with a Phenomenex Rezex RFQ ‐Fast Fruit H + column and cation H + guard cartridge at 85°C, using 0.01 N sulphuric acid as a mobile phase at a flow rate of 1.0 ml min −1 and a diode array detector scanning at 325 nm. The analysis of aromatic compounds in lignin cultures was conducted as previously described (Salvachúa et al ., 2018 ). Culture conditions (Davis et al ., 2015 ), carbon sources (Cai et al ., 2009 ) and volume ratios (v media :v flask ) (Poblete‐Castro et al ., 2014 ) are critical parameters in mcl ‐PHA production. Considering the number of variables, quantitative comparison of mcl ‐PHA production studies on an equivalent basis is challenging. In fact, comparisons become more complicated when using a heterogeneous substrate as lignin since, in many cases, lignin streams contain carbon sources other than aromatic compounds (e.g. acetic acid and sugars) that can lead to the production of mcl ‐PHAs (Linger et al ., 2014 ), or contain very different lignin concentrations [e.g. 10 g l − 1 (Salvachúa et al ., 2015 ) to 30 g l − 1 of lignin (Rodriguez et al ., 2017 )]. In addition, considering the increased titres obtained in fed‐batch mode from p ‐CA (Fig. 3 ) as well as the mcl ‐PHA titres (1 g l − 1 ) achieved from a lignin stream in fed‐batch mode in a recent publication (Liu et al ., 2017 ) (Table 1 ), it is likely that mcl ‐PHA titres from the current lignin stream could be further improved by using a different feeding strategy. However, in this study we did not pursue optimizing titres from lignin because the main limitation currently faced in valorizing lignin is its low content of bioavailable monomeric aromatic species and carbon to the production hosts (Beckham et al ., 2016 ). Nevertheless, it is worth highlighting that the lignin stream utilized in this study contains up to 15% of monomeric species (mainly p ‐CA) (Rodriguez et al ., 2017 ; Salvachúa et al ., 2018 ), which is already a reasonable concentration to be upgraded. Overall, while there is extensive space to improve the conversion of aromatic compounds and lignin to mcl ‐PHAs through process development in bioreactors, our results suggest that strains developed here can be a reasonable starting platform to efficiently convert lignin‐derived aromatic compounds into different value‐added molecules that are derived from fatty acid biosynthesis (e.g. fatty alcohols, ketones, chemically‐functionalized mcl ‐PHAs). Furthermore, the AG2162 background can also be utilized for further pathway engineering to increase mcl‐PHA titre, rate and yield by increasing flux into fatty acid biosynthesis in future work."
} | 7,160 |
30410432 | PMC6209684 | pmc | 22 | {
"abstract": "Spiking neural networks (SNNs) are inspired by information processing in biology, where sparse and asynchronous binary signals are communicated and processed in a massively parallel fashion. SNNs on neuromorphic hardware exhibit favorable properties such as low power consumption, fast inference, and event-driven information processing. This makes them interesting candidates for the efficient implementation of deep neural networks, the method of choice for many machine learning tasks. In this review, we address the opportunities that deep spiking networks offer and investigate in detail the challenges associated with training SNNs in a way that makes them competitive with conventional deep learning, but simultaneously allows for efficient mapping to hardware. A wide range of training methods for SNNs is presented, ranging from the conversion of conventional deep networks into SNNs, constrained training before conversion, spiking variants of backpropagation, and biologically motivated variants of STDP. The goal of our review is to define a categorization of SNN training methods, and summarize their advantages and drawbacks. We further discuss relationships between SNNs and binary networks, which are becoming popular for efficient digital hardware implementation. Neuromorphic hardware platforms have great potential to enable deep spiking networks in real-world applications. We compare the suitability of various neuromorphic systems that have been developed over the past years, and investigate potential use cases. Neuromorphic approaches and conventional machine learning should not be considered simply two solutions to the same classes of problems, instead it is possible to identify and exploit their task-specific advantages. Deep SNNs offer great opportunities to work with new types of event-based sensors, exploit temporal codes and local on-chip learning, and we have so far just scratched the surface of realizing these advantages in practical applications.",
"introduction": "1. Introduction Training and inference with deep neural networks (DNNs), commonly known as deep learning (LeCun et al., 2015 ; Schmidhuber, 2015 ; Goodfellow et al., 2016 ), has contributed to many of the spectacular success stories of artificial intelligence (AI) in recent years (Goodfellow et al., 2014 ; Amodei et al., 2016 ; He et al., 2016 ; Silver et al., 2016 ). Models of cortical hierarchies from neuroscience have strongly inspired the architectural principles behind DNNs (Fukushima, 1988 ; Riesenhuber and Poggio, 1999 ), but at the implementation level, only marginal similarities between brain-like computation and analog neural networks (ANNs) as used in AI applications can be recognized. One obvious difference is that neurons in ANNs are mostly non-linear but continuous function approximators that operate on a common clock cycle, whereas biological neurons compute with asynchronous spikes that signal the occurrence of some characteristic event by digital and temporally precise action potentials. In recent years, researchers from the domains of machine learning, computational neuroscience, neuromorphic engineering, and embedded systems design have tried to bridge the gap between the big success of DNNs in AI applications and the promise of spiking neural networks (SNNs) (Maass, 1997 ; Ponulak and Kasinski, 2011 ; Grüning and Bohte, 2014 ). This promise of SNNs results from their favorable properties exhibited in real neural circuits like brains, such as analog computation, low power consumption, fast inference, event-driven processing, online learning, and massive parallelism. Furthermore, event-based vision and audio sensors (Lichtsteiner et al., 2008 ; Posch et al., 2014 ; Liu et al., 2015 ) have reached an increasingly mature level, and deep SNNs are one of the most promising concepts for processing such inputs efficiently (Tavanaei et al., 2018 ). This line of research has coincided with an increased interest in efficient hardware implementations for conventional DNNs, since the massive hunger for computational resources has turned out to be a major obstacle as deep learning makes its way toward real-world applications such as automated driving, robotics, or the internet of things (IoT). Concepts such as so-called binary networks , which allow in-memory computations, share a binary and potentially sparse communication scheme with SNNs. However, such networks are typically executed in a synchronized manner, which is different from the event-driven (asynchronous) mode of execution in SNNs. Consequently, a fruitful interdisciplinary exchange of ideas to build neuromorphic systems for these concepts is taking place. In this review, we provide an overview of several key ideas behind deep SNNs, and discuss challenges and limitations of SNNs compared to their ANN counterparts, as well as opportunities for future applications, in particular in conjunction with novel computing models and hardware currently being developed. This article is structured as follows: Section 2 discusses the preparation of input and output in order to perform inference with deep SNNs. In section 3, we give an overview of how deep SNNs can be trained, how this is connected to training conventional DNNs, and how to possibly learn on spike level. Section 4 discusses efficient implementations of deep SNNs on neuromorphic hardware and their limitations, as well as highlights similarities to hardware-efficient solutions for conventional DNNs. In section 5, we present possible use cases of deep SNNs, and argue that their strengths are complementary to those of conventional DNNs. Finally, section 6 provides a discussion of the state-of-the-art, and gives an outlook on promising research directions. 1.1. What is a deep spiking neural network? Neural networks are typically called deep in case they have at least two hidden layers computing non-linear transformations of the input. In this article, we consider only feed-forward networks, which compute a mapping from input to output (for an example see Figure 1A ), and do not address recurrent neural networks. Our definition includes multi-layer fully-connected networks, convolutional neural networks (CNNs; LeCun and Bengio, 1995 ), deep belief networks (DBNs; Hinton et al., 2006 ), deep autoencoders, and many more. Figure 1 Comparison of deep spiking neural networks (SNNs) to conventional deep neural networks (DNNs). (A) Example of a deep network with two hidden layers. Here, exemplarily a fully-connected network is shown. Neurons are depicted with circles, connections with lines. (B) Time-stepped layer-by-layer computation of activations in a conventional DNN with step duration Δ T . The activation values of neurons (rectangles) are exemplarily visualized with different gray values. The output of the network, e.g. categories in the case of a classification task, is only available after all layers are completely processed. (C) Like (B) , but with binarized activations. (D) The activity of a deep SNN showing a fast and asynchronous propagation of spikes through the layers of the network. (E) The membrane potential of the neuron highlighted in green in (D) . When the membrane potential (green) crosses the threshold (black dashed line) a spike is emitted and the membrane potential is reset. (F) The first spike in the output layer (red arrow in D ) rapidly estimates the category (assuming a classification task) of the input. The accuracy of this estimation increases over time with the occurrence of more spikes (red line and Diehl et al., 2015 ). In contrast, the time-stepped synchronous operation mode of DNNs results in later, but potentially more accurate classifications compared to SNNs (blue dashed line and red arrows in B,C ). Spiking neural networks were originally studied as models of biological information processing (Gerstner and Kistler, 2002 ), in which neurons exchange information via spikes (for an example, see Figure 1D ). It is assumed that all spikes are stereotypical events, and, consequently, the processing of information is reduced to two factors: first, the timing of spikes, e.g., firing frequencies, relative timing of pre- and postsynaptic spikes, and particular firing patterns. Second, the identity of the synapses used, i.e., which neurons are connected, whether the synapse is excitatory or inhibitory, the synaptic strength, and possible short-term plasticity or modulatory effects. Depending on the level of detail of the simulation neurons are either point neurons in which arriving spikes immediately change their (somatic) membrane potentials, or are modeled as multi-compartment models with complex spatial (dendritic) structure, such that dendritic currents can interact before the somatic potential is modified. Different spiking neuron models such as the integrate-and-fire, spike response, or Hodgkin-Huxley model describe the evolution of the membrane potential and spike generation in different levels of detail. Typically, the membrane potential integrates currents from arriving spikes and generates a new spike whenever some threshold is crossed (Figure 1E ). Once a spike is generated, it is sent via the axon to all connected neurons with a small axonal delay and the membrane potential is reset toward a given baseline. The most direct connection between analog and spiking neural networks is made by considering the activation of an analog neuron as the equivalent of the firing rate of a spiking neuron assuming a steady state. Many models of neuronal measurements have used such rate codes to explain computational processes in brains (Hubel and Wiesel, 1959 ; Rieke, 1999 ). However, spiking neuron models can also model more complex processes that depend on the relative timing between spikes (Gütig, 2014 ) or on timing relative to some reference signal, such as network oscillations (Montemurro et al., 2008 ). Temporal codes are of high importance in biology where even a single spike or small temporal variations of single neuron firing may trigger different reactions (Gerstner et al., 1996 ; Stemmler, 1996 ; Rieke, 1999 ; Machens et al., 2003 ), because often decisions have to be made before a reliable estimate of a spike rate can be computed. Besides the biologically motivated definition of SNNs, there is a more pragmatic application-oriented view coming from the field of neuromorphic engineering, where SNNs are often called event-based instead of spiking (Liu et al., 2015 ). Here, an event is a digital packet of information, which is characterized by its origin and destination address, a timestamp, and - in contrast to biologically motivated SNNs—may carry a few bits of payload information. The origin of this view is the address event representation (AER) protocol (Mahowald, 1994 ; Boahen, 2000 ), which is used to connect, e.g., event-based sensors (Lichtsteiner et al., 2008 ) via digital interconnect to neuromorphic chips (Indiveri et al., 2011 ; Amir et al., 2017 ) or digital post-processing hardware (Furber et al., 2014 ). Event-based vision sensors use the payload bits to distinguish visual ON or OFF events, but the payload can also be used to send any other type of relevant information to the postsynaptic targets potentially computing more sophisticated functions than simple integrate-and-fire (Stefanini et al., 2014 ). 1.2. Advantages of deep SNNs A motivation for studying SNNs is that brains exhibit a remarkable cognitive performance in real-world tasks. With ongoing efforts toward improving our understanding of brain-like computation, there are expectations that models staying closer to biology will also come closer to achieving natural intelligence than more abstract models, or at least will have greater computational efficiency. SNNs are ideally suited for processing spatio-temporal event-based information from neuromorphic sensors, which are themselves power efficient. The sensors record temporally precise information from the environment and SNNs can utilize efficient temporal codes in their computations as well (Mostafa, 2018 ). This processing of information is also event-driven meaning that whenever there is little or no information recorded the SNN does not compute much, but when sudden bursts of activity are recorded, the SNN will create more spikes. Under the assumption that typically information from the outside world is sparse, this results in a highly power-efficient way of computing. In addition, using time domain input is additional valuable information compared to frame-driven approaches, where an artificial time step imposed by the sensor is introduced. This can lead to efficient computation of features such as optical flow (Benosman et al., 2014 ) or stereo disparity (Osswald et al., 2017 ), and in combination with learning rules sensitive to spike timing leads to more data-efficient training (Panda et al., 2017 ). In deep SNNs, the asynchronous data-driven mode of computing leads to fast propagation of salient information through multiple layers of the network. To best exploit this effect in practice, SNNs should be run on neuromorphic hardware. In combination with an event-based sensor, this results in pseudo-simultaneous information processing (Farabet et al., 2012 ; Camuñas-Mesa et al., 2014 ), which means that a first approximate output of the final layer is available immediately after recording the first input spikes. This is true even for multi-layer networks, because spikes begin to propagate immediately to higher layers as soon as the lower layer provides sufficient activity (Figure 1D ). It is not necessary to wait for the complete input sequence to finish, which is in contrast to conventional DNNs, where all layers need to be fully updated before the final output can be computed (Figures 1B,C ). The initial output spikes are necessarily based on incomplete information, hence it has been shown that deep SNNs improve their classification performance the longer they are given time to process more spikes of their input (Figure 1F ). SNNs can also be trained specifically to reduce the latency of approximate inference (Neil et al., 2016a ). SNNs are the preferred computational model to exploit highly energy-efficient neuromorphic hardware devices, which support the data-driven processing mode, and keep computations local, thereby avoiding expensive memory access operations. 1.3. Limitations of deep SNNs One of the biggest drawbacks of deep SNNs is that despite recent progress (Rueckauer et al., 2017 ; Sengupta et al., 2018 ) their accuracy on typical benchmarks such as MNIST (Lecun et al., 1998 ), CIFAR (Krizehvsky and Hinton, 2009 ), or ImageNet (Russakovsky et al., 2015 ) do not reach the same levels as their machine learning counterparts. To some extent, this can be attributed to the nature of these benchmarks, which are on conventional frame-based images. Thus, some form of conversion from images into spike trains is required that is typically lossy and inefficient. Another limiting factor is the lack of training algorithms that make specific use of the capabilities of spiking neurons, e.g., efficient time codes. Instead, most approaches use rate-based approximations of conventional DNNs, which means that no accuracy gains can be expected. Deep SNNs might still be useful in such scenarios, because approximate results might be obtained faster and more efficiently than on conventional systems, especially if the SNN is run on neuromorphic hardware. Training algorithms for SNNs are also more difficult to design and analyze, because of the asynchronous and discontinuous way of computing, which makes a direct application of successful backpropagation techniques as used for DNNs difficult. The performance of SNNs on conventional AI benchmarks should only be seen as a proof-of-concept, but not as the ultimate research goal. If spiking networks model biology, then we should expect them to be optimized for the behaviorally most relevant tasks, such as making decisions based on continuous input streams while moving in the real world. Image classification corresponds to the task of classifying a random image suddenly flashed on the retina, without any supporting context. While brains are able to solve such tasks (Thorpe et al., 1996 ), they are certainly not optimized for it. We are currently lacking both good benchmark datasets and evaluation metrics that could measure efficient real-world performance. One fruitful direction is the collection of dynamic vision sensor (DVS) benchmarks (Orchard et al., 2015a ; Serrano-Gotarredona and Linares-Barranco, 2015 ; Hu et al., 2016 ; Liu et al., 2016 ), in particular for relevant use cases such as automated driving (Binas et al., 2017 ; Sironi et al., 2018 ).",
"discussion": "6. Discussion Advances in deep SNNs have helped closing the performance gap to conventional DNNs. However, the promise of low-power inference is not fulfilled yet, since network conversion (section 3.2) and training of constrained networks (section 3.3) result in spike-based networks that encode information mostly in their neurons' mean firing rates, but do not exploit the potential of encoding information in the timing of single spikes. Although these networks achieve a remarkable performance on various benchmark datasets, the average firing rates of their neurons are comparably high for static input images, and hence their energy-efficiency on neuromorphic systems is not significantly better than for conventional DNNs on GPUs (section 4). To reduce firing rates and increase energy-efficiency spike-based training methods (section 3.4) and local learning rules (section 3.5) have become increasingly popular research topics. Their accuracy on machine learning benchmarks is not quite at the level of converted networks, but recent approaches by Lee et al. ( 2016 ) or Jin et al. ( 2018 ) could partly close the gap. Besides, the choice of benchmarks that usually consist of frame-based images converted to spiking representations (section 2) puts spike-based rules at a disadvantage. Finding local learning rules that can achieve the same performance as backpropagation would be a result with great implications beyond machine learning applications, since it could possibly explain how brains can learn their remarkable capabilities with the constraints for information and error signal routing imposed by biology (Bengio et al., 2015a ). We have argued in section 5 that further opportunities for deep SNNs will arise when appropriate benchmark datasets recorded with event-based sensors become available. The rise of deep learning has largely been driven by the availability of large common benchmarks such as ImageNet (Russakovsky et al., 2015 ). Similarly large and challenging neuromorphic datasets are not available, yet, but we see a positive trend and increased awareness of the community. First benchmarks for real-world applications in automated driving (Binas et al., 2017 ; Sironi et al., 2018 ) and robotics (Mueggler et al., 2017 ) have been released, and together with convincing results on problems where conventional systems struggle (Kim et al., 2016 ; Vidal et al., 2018 ), we expect that this will lead to increasing demand for efficient event-based post-processing systems. Fully event-based systems are not only energy-efficient, but could also better exploit the rich temporal dynamics of the real world than frame-based approaches, which artificially introduce time steps through sensing or processing components. For agents interacting with the real world, temporal information on different time-scales plays an important role, because critical situations require short reaction times hardly accessible by frame-based perception. Deep SNNs have the important property of providing good early estimates, which improve when given more processing time (see Figure 1F ). Although mechanisms to provide early estimations are also proposed for conventional DNNs (e.g., Teerapittayanon et al., 2016 ), their implementations are rather artificial and not as seamlessly integrated as in SNNs. Fischer et al. ( 2018 ) proposed a hybrid solution between conventional DNNs and deep SNNs, called streaming rollouts , which are conventional synchronous DNNs that share a dense temporal integration and fast response times with deep SNNs. A future direction of research may be the incorporation of recurrence into deep SNNs improving the storage and integration of temporal information. Recurrent SNNs have shown remarkable performance in sequence recognition (Zhang et al., 2015 ; Panda and Srinivasa, 2018 ) and generation tasks (Rajan et al., 2016 ; Panda and Roy, 2017 ). In these cases, instead of a deep or structured recurrent architecture the networks were configured as liquid state machines (Maass et al., 2002 ), which consist of a reservoir of randomly and recurrently connected neurons, followed by a linear readout. Recent work (e.g., Diehl et al., 2016 ; Bellec et al., 2018 ) have shown how standard recurrent network architectures such as long short-term memory networks (LSTMs, Hochreiter and Schmidhuber, 1997 ) can be ported into the spiking domain, whereas Neil et al. ( 2016b ) have shown a way to process event-based data with otherwise standard recurrent networks. Combining recurrent architectures with the intrinsic short-time memory of spiking neurons appears as a promising route for efficiently solving real-world pattern recognition tasks. As deep SNNs become larger and capable of solving tasks that are more complex, training time is likely to become a bottleneck due to the more complex training methods compared to conventional DNNs, as well as inefficient spiking simulations on conventional computing platforms. It is therefore important to advance neuromorphic hardware systems for large-scale deep SNNs, and not only consider energy-efficient inference, but also training. Efficient training can be either realized via on-chip learning rules like STDP as discussed in section 4.2, by using neuromorphic systems in-the-loop, i.e., computing weight updates on the host computer and then re-configuring the hardware system, or by hybrid solutions. However, contemporary neuromorphic systems share a comparably low bandwidth to the host computer, usually sufficient for spike input and output, but inappropriate for a frequent re-configuration of the device. This is why the development and investigation of hierarchies of learning rules both on algorithmic and hardware level, ranging from in-memory plasticity rules like STDP to global reward signals, would be a valuable topic for future studies. Friedmann et al. ( 2017 ) and Lin et al. ( 2018 ) already proposed architectures that go into this direction, and it will be interesting to see first large-scale experimental results and further developments in the near future. Although such systems may allow for the exploration of networks with a size and complexity not accessible with current hardware systems, their development is time consuming, costly, and will most likely not offer the flexibility to catch up with the latest algorithmic developments. Compared to digital systems, analog systems promise a higher energy-efficiency. However, the training of analog systems requires additional efforts (see section 4.1) and the short- and long-term variations in their parameters and computations, e.g., caused by temperature fluctuations, pose great challenges."
} | 5,875 |
32694978 | PMC7339957 | pmc | 23 | {
"abstract": "In the last few years, spiking neural networks (SNNs) have been demonstrated to perform on par with regular convolutional neural networks. Several works have proposed methods to convert a pre-trained CNN to a Spiking CNN without a significant sacrifice of performance. We demonstrate first that quantization-aware training of CNNs leads to better accuracy in SNNs. One of the benefits of converting CNNs to spiking CNNs is to leverage the sparse computation of SNNs and consequently perform equivalent computation at a lower energy consumption. Here we propose an optimization strategy to train efficient spiking networks with lower energy consumption, while maintaining similar accuracy levels. We demonstrate results on the MNIST-DVS and CIFAR-10 datasets.",
"introduction": "1. Introduction Since the early 2010s, computer vision has been dominated by the introduction of convolutional neural networks (CNNs), which have yielded unprecedented success in previously challenging tasks such as image recognition, image segmentation or object detection, among others. Considering the theory of neural networks was mostly developed decades earlier, one of the main driving factors behind this evolution was the widespread availability of high-performance computing devices and general purpose Graphic Processing Units (GPU). In parallel with the increase in computational requirements (Strubell et al., 2019 ), the last decades have seen a considerable development of portable, miniaturized, battery-powered devices, which pose constraints on the maximum power consumption. Attempts at reducing the power consumption of traditional deep learning models have been made. Typically, these involve optimizing the network architecture, in order to find more compact networks (with fewer layers, or fewer neurons per layer) that perform equally well as larger networks. One approach is energy-aware pruning, where connections are removed according to a criterion based on energy consumption, and accuracy is restored by fine-tuning of the remaining weights (Molchanov et al., 2016 ; Yang et al., 2017 ). Other work looks for more efficient network structures through a full-fledged architecture search (Cai et al., 2018 ). The latter work was one of the winners of the Google “Visual Wake Words Challenge” at CVPR 2019, which sought models with memory usage under 250 kB, model size under 250 kB and per-inference multiply-add count (MAC) under 60 millions. Using spiking neural networks (SNNs) on neuromorphic hardware is an entirely different, and much more radical, approach to the energy consumption problem. In SNNs, like in biological neural networks, neurons communicate with each other through isolated, discrete electrical signals (spikes), as opposed to continuous signals, and work in continuous instead of discrete time. Neuromorphic hardware (Indiveri et al., 2011 ; Esser et al., 2016 ; Furber, 2016 ; Thakur et al., 2018 ) is specifically designed to run such networks with very low power overhead, with electronic circuits that faithfully reproduce the dynamics of the model in real time, rather than simulating it on traditional von Neumann computers. Some of these architectures (including Intel's Loihi, IBM's TrueNorth, and SynSense's DynapCNN) support convolution operations, which are necessary for modern computer vision techniques, by an appropriate weight sharing mechanism. The challenge of using SNNs for machine learning tasks, however, is in their training. Mimicking the learning process used in the brain's spiking networks is not yet feasible, because neither the learning rules, nor the precise fitness functions being optimized are sufficiently well-understood, although this is currently a very active area of research (Marblestone et al., 2016 ; Richards et al., 2019 ). Supervised learning routines for spiking networks have been developed (Bohte et al., 2002 ; Mostafa, 2017 ; Nicola and Clopath, 2017 ; Shrestha and Orchard, 2018 ; Neftci et al., 2019 ), but are slow and challenging to use. For applications which have little or no dependence on temporal aspects, it is more efficient to train an analog network (i.e., a traditional, non-spiking one) with the same structure, and transfer the learned parameters onto the SNN, which can then operate through rate coding. In particular, the conversion of pre-trained CNNs to SNNs has been shown to be a scalable and reliable process, without much loss in performance (Diehl et al., 2015 ; Rueckauer et al., 2017 ; Sengupta et al., 2019 ). But this approach is still challenging, because the naive use of analog CNN weights does not take into account the specific characteristics and requirements of SNNs. In particular, SNNs are more sensitive than analog networks to the magnitude of the input. Naive weight transfer can, therefore, lead to a silent SNN, or, conversely, to one with unnecessarily high firing rates, which have a high energy cost. Here, we propose a hybrid training strategy which maintains the efficiency of training analog CNNs, while accounting for the fact that the network is being trained for eventual use in SNNs. Furthermore, we include the energy cost of the network's computations directly in the loss function during training, in order to minimize it automatically and dynamically. We demonstrate that networks trained with this strategy perform better per Joule of energy utilized. While we demonstrate the benefit of optimizing based on energy consumption, we believe this strategy is extendable to any approach that uses back-propagation to train the network, be it through a spiking network or a non-spiking network. In the following sections, we will detail the training techniques we devised and applied for these purposes. We will test our networks on two standard problems. The first is the MNIST-DVS dataset of Dynamic Vision Sensor recordings. DVSs are event-based sensors, and, as such, the analysis of their recordings is an ideal application of spike-based neural networks. The second is the standard CIFAR-10 object recognition benchmark, which provides a reasonable comparison on computation cost to non-spiking networks. For each of these tasks, we will demonstrate the energy-accuracy trade-off of the networks trained with our methods. We show that significant amounts of energy can be saved with a small loss in performance, and conclude that ours is a viable strategy for training neuromorphic systems with a limited power budget.",
"discussion": "4. Discussion and Conclusion We used two techniques which significantly improve the energy requirements of machine learning models that run on neuromorphic hardware, while maintaining similar performances. The first improvement consisted in optimizing the energy expenditure by directly adding it to the loss function during training. This method encourages smaller activations in all neurons, which is not in itself an issue in analog models, but can lead to discretization errors, due to the lower firing rates, once the weights are transferred to a spiking network. To solve this problem, we introduced the second improvement; quantization-aware training, whereby the network activity is quantized at each layer, i.e., only integer activations are allowed. Discretizing the network's activity would normally reduce all gradients to zero: this can be solved by substituting the true gradient with a surrogate. Applying these two methods together, we achieved an up to 10-fold drop in the number of synaptic operations and the consequent energy consumption in the DVS-MNIST task, with only a minor (1-2%) loss in performance, when comparing to simply transferring the weights from a trained CNN to a spiking network. To demonstrate the scalability of this approach, we also show that, as the network grows bigger to solve a much more complex task of CIFAR-10 image classification, the SynOps are reduced to 42% of the MAC, while losing 1% of accuracy (90.37% at 127M). The accuracy-energy trade-off can be flexibly tuned at training time. We also showed the consequences of using this method on the distribution of network weights and the network's accuracy as a function of time. While training based on static frames is not the optimal approach to leverage all the benefits of spike-based computation, it enables fast training with the use of state-of-the-art deep learning tools. In addition, the hybrid strategy to train SNNs based on a target power metric is unique to SNNs. Conversely, optimizing the energy requirement of an ANN/CNN requires modification of the network architecture itself, which can require large amounts of computational resources (Cai et al., 2018 ). In this work, we demonstrated that we can train an SNN to a target energy level without a need to alter the network hyperparameters. A potential drawback of this approach of (re)training the model as opposed to simply transferring the weights of a pre-trained model is brought to light when attempting to convert very deep networks trained over large datasets such as IMAGENET. Pre-trained deep CNNs trained over large datasets are readily available on the web and can be used to quickly instantiate a spiking CNN. The task becomes much more cumbersome to optimize for power utilization using the method described in this paper, ie. one has to retrain the network over the relevant dataset for optimal performance. However, our method can also be effectively used to fine-tune a pre-trained network, removing the need for training from scratch ( Supplementary Figure 1 ). Furthermore, no large event-based datasets of the magnitude of IMAGENET exist currently, and perhaps when such datasets are generated, the corresponding models optimized for spiking CNNs will also be developed and made readily available. The quantization and SynOp-based optimization used in this paper can potentially be applied, beyond the method illustrated here, in more general contexts such as algorithms based on back-propagation through time to reduce power usage. Such a reduction in power usage can make a large difference when the model is ran on a mobile, battery-powered, neuromorphic device, with potential for a significant impact in the industrial applications."
} | 2,545 |
34723173 | PMC8554302 | pmc | 24 | {
"abstract": "Memristors show great promise in neuromorphic computing owing to their high-density integration, fast computing and low-energy consumption. However, the non-ideal update of synaptic weight in memristor devices, including nonlinearity, asymmetry and device variation, still poses challenges to the in-situ learning of memristors, thereby limiting their broad applications. Although the existing offline learning schemes can avoid this problem by transferring the weight optimization process into cloud, it is difficult to adapt to unseen tasks and uncertain environments. Here, we propose a bi-level meta-learning scheme that can alleviate the non-ideal update problem, and achieve fast adaptation and high accuracy, named Rapid One-step Adaption (ROA). By introducing a special regularization constraint and a dynamic learning rate strategy for in-situ learning, the ROA method effectively combines offline pre-training and online rapid one-step adaption. Furthermore, we implemented it on memristor-based neural networks to solve few-shot learning tasks, proving its superiority over the pure offline and online schemes under noisy conditions. This method can solve in-situ learning in non-ideal memristor networks, providing potential applications of on-chip neuromorphic learning and edge computing.",
"introduction": "Introduction Memristors are considered as leading device candidates for neural network accelerators ( Yang et al., 2013 ; Chen et al., 2015 ; Tsai et al., 2018 ; Zidan et al., 2018 ) due to their ability to physically store synaptic weights in conductance state, which enable in-memory computing. Implementation of neural networks in memristor-based hardware exhibits high density integration, low power consumption and high efficiency ( Cai et al., 2019 ; Yu 2018 ). It can also greatly promote the development of brain-inspired computing systems to achieve human-like intelligence ( Zhang et al., 2016 ). However, memristors possess some non-ideal properties that challenge the hardware implementations. The weight updates on memristors are asymmetric, nonlinear and low precision, significantly degrading the learning accuracy ( Kataeva et al., 2015 ; Agarwal et al., 2016 ; Wang et al., 2020a ). Additionally, the outputs of networks, determined by input currents and conductance of memristors, are also perturbed by the variability of circuits, including input currents, reference voltage, output resistance ( Yang et al., 2013 ; Agarwal et al., 2016 ). Currently, there are mainly two types of co-optimization learning schemes to overcome these challenges ( Hu et al., 2016 ; Zidan et al., 2018 ; Agarwal et al., 2016 ; Chen et al., 2015 ). One type is online learning, which allows training models to be implemented on neuromorphic hardware by using backpropagation ( Yu et al., 2016 ) or biological local learning, such as spike-timing-dependent plasticity (STDP) ( Guo et al., 2017 ). For fast online learning, conductance tuning with less operations on hardware is preferred, which the weights of networks are directly written without verification by reading. However, the nonlinearity and asymmetry of memristors cause the accuracy loss of the neural networks during learning ( Yu et al., 2016 ; Kataeva et al., 2015 ). To mitigate the adverse effects of memristors, various approaches have been reported. Some work initialized the weights at each update step to achieve linear and symmetric weights ( Li et al., 2019 ; Geminiani et al., 2018 ). Retraining of networks and developing highly robust algorithms, such as Neural State Machine, have also been proved to overcome the asymmetric properties of memristors to a certain degree ( Liu et al., 2017 ; Tian et al., 2020 ; Tian et al., 2021 ). The other type is offline learning, which maps the pre-trained network to hardware, and only performs inference in neuromorphic chips ( Hu et al., 2016 ; Shafiee et al., 2016 ; Chi et al., 2016 ). The non-ideality of weight updates can be concealed by iterative programming with a write-verify technique, reading the conductance and rewriting for accuracy. However, for a new task, the entire process must be restarted from scratch through offline learning. For most algorithms, all tunable parameters in the neural network must be re-trained for a new task, resulting in a large number of operations. Therefore, there is usually a trade-off between speed and performance for memristor networks from offline learning to online learning. Different from the current on-chip learning schemes, humans can quickly adapt to the environment by drawing on prior experience or learning to learn. In machine learning, this learning approach is named meta-learning ( Thrun and Pratt 1998 ), which has made significant progress in recent years ( Wang et al., 2020b ; Hospedales et al., 2020 ; Vanschoren, 2018 ). There are many meta-learning techniques, such as optimizee ( Andrychowicz et al., 2016 ; Ravi and Larochelle 2017 ), metric based ( Hu et al., 2020 ; Vinyals et al., 2016 ; Snell et al., 2017 ) and fine-tuning ( Antoniou et al., 2018 ; Li et al., 2017 ; Finn et al., 2017 ; Nichol et al., 2018 ). Particularly, Model-Agnostic Meta-Learning (MAML) ( Finn et al., 2017 ) is a general meta-learning framework that provides a good initial condition of network for fine-tuning on similar tasks, which can simplify the optimization to a few steps for new unseen tasks. MAML can also be applied to fields such as reinforcement learning ( Gupta et al., 2018 ) and continual learning ( Al-Shedivat et al., 2018 ). The studies on the silicon-based neuromorphic chips have proven that meta-learning schemes can significantly accelerate the learning of new tasks and improve their performance ( Bohnstingl et al., 2019 ; Stewart et al., 2020 ). However, optimization methods for memristor-based networks with the meta-learning scheme have yet to be developed. In this work, we propose a meta-learning scheme for memristor-based neural networks that can overcome the non-ideal synapse weights for training and provide improved performance. Our method consists of two phases, including pre-training and task adaptation, as shown in Figure 1 . Firstly, a good initial network for a group of tasks is trained in software and then mapped to hardware by iterative programming with write-verify. Then, a rapid training in one-step adaption is performed for an unseen task with a few samples of the in-situ hardware network. This scheme can free the memristor networks from unnecessary operations, mitigating the problem of performance degradation in online learning. It also has the ability to accomplish new tasks through quick adaptations, which is more powerful than the offline trained networks. Since only one update step is needed, a new task requires significantly less training time, only a few samples and little computation consumption. These merits make our scheme very suitable for situations with limited computing power and limited data, such as edge computing. Our main contributions are as follows: 1. We propose a hybrid learning scheme of offline learning and online learning for meta-learning on memristor-based neural networks. It combines the advantages of offline learning and online learning for the hardware to achieve high accuracy and fast adaption for unseen tasks. 2. We report the Rapid One-step Adaption (ROA) algorithm, which enables memristor-based neural networks with meta-learning capability. It mitigates the effects of non-ideal characteristics of memristor-based neural networks, and achieves superior performance by introducing dynamic learning rate, regularization constraint and one-step adaption. 3. In order to evaluate our model on few-shot tasks, we built a simulator based on the experimental characteristics of memristor, which can better support the acceleration of large-scale network and the quantitative analysis of networks with noise. On this basis, we comprehensively evaluate the proposed model on two typical few-shot learning datasets. Our results reveal a good performance of memristor networks on few-shot learning task with significant improvement of accuracy than the baseline. \n FIGURE 1 The framework of ROA (Rapid One-step Adaption). In the pre-training phase: (1) we train a good initial network across tasks with parameters \n g 0 i j \n ; (2) we load it into the memristor array as \n G 0 i j \n with iterative programming with write-verify. In the task adaptation phase: (3) we map the task into voltage input of the memristor for prediction \n y ^ i \n , then we calculate the loss function with true label \n y i \n ; (4) we calculate the gradient descent \n Δ g i j \n with loss by backpropagation; (5) we convert \n Δ g i j \n into update pulse period \n t i j \n and update the hardware network; (6) we evaluate the network on the testing dataset. The color and its brightness of the matrix indicate different forms of data and its value.",
"discussion": "Discussion and Conclusion In this work, we developed a bi-level meta-learning scheme, ROA, for memristor neural networks. It is a hybrid approach that combines online learning and offline learning, which can effectively alleviate the impact of the non-ideal properties of memristors through one update step. A simulator was built based on the parameters extracted from our memristor devices and evaluated using the Omniglot dataset and MiniImageNet dataset. Our experimental results demonstrate that the ROA method can significantly improve data efficiency and training speed, thereby achieving better performance than multi-step adaption and offline learning under similar conditions. In addition, our method shows a strong robustness to noise, which facilitates the real-world applications of memristor networks. The results suggest that, with the proposed, the memristors are suitable as an accelerator for rapid learning hardware rather than just a hardware inference or in-situ learning with massive updates. Moreover, memristor networks and the ROA method can benefit from each other. The rapid adaption of ROA requires that the weights in the neuromorphic hardware have on-chip plasticity, which can be easily achieved by memristor networks. On the other hand, the one-step adaption allows the hardware network to extricate the weight update from the non-ideal properties of memristors, thereby reducing the accuracy loss in the mapping process. Collectively, memristors are very suitable for accelerators to achieve learning-to-learn capability. Our ROA scheme can improve the performance of memristors, and facilitate broad applications in neuromorphic architecture. The rapid adaptation process could be implemented in a local mode without the support of cloud servers, indicating a low adaptation latency. Hence, the users’ personal data do not need to be uploaded to server, ensuring privacy and security. Furthermore, the flexible learning scheme can benefit hardware neural networks to handle uncertain environments and individual demands."
} | 2,746 |
38873286 | PMC11169843 | pmc | 25 | {
"abstract": "The spiking convolutional neural network (SCNN) is a kind of spiking neural network (SNN) with high accuracy for visual tasks and power efficiency on neuromorphic hardware, which is attractive for edge applications. However, it is challenging to implement SCNNs on resource-constrained edge devices because of the large number of convolutional operations and membrane potential (Vm) storage needed. Previous works have focused on timestep reduction, network pruning, and network quantization to realize SCNN implementation on edge devices. However, they overlooked similarities between spiking feature maps (SFmaps), which contain significant redundancy and cause unnecessary computation and storage. This work proposes a dual-threshold spiking convolutional neural network (DT-SCNN) to decrease the number of operations and memory access by utilizing similarities between SFmaps. The DT-SCNN employs dual firing thresholds to derive two similar SFmaps from one Vm map, reducing the number of convolutional operations and decreasing the volume of Vms and convolutional weights by half. We propose a variant spatio-temporal back propagation (STBP) training method with a two-stage strategy to train DT-SCNNs to decrease the inference timestep to 1. The experimental results show that the dual-thresholds mechanism achieves a 50% reduction in operations and data storage for the convolutional layers compared to conventional SCNNs while achieving not more than a 0.4% accuracy loss on the CIFAR10, MNIST, and Fashion MNIST datasets. Due to the lightweight network and single timestep inference, the DT-SCNN has the least number of operations compared to previous works, paving the way for low-latency and power-efficient edge applications.",
"introduction": "1 Introduction Spiking neural networks (SNNs) are inspired by the brain and use spikes (binary signals) to transmit information between neurons. Neuromorphic hardware only requires processing the spike-based accumulate (ACC) operations, effectively bypassing the need to compute zero input values to attain remarkable power efficiency. Consequently, SNNs exhibit significant energy efficiency when implemented on neuromorphic hardware (Pei et al., 2019 ), making them increasingly appealing for edge applications (Zhang et al., 2020 ; Liu and Zhang, 2022 ). Spiking convolutional neural networks (SCNNs) is a kind of SNN widely used in vision tasks (Cao et al., 2015 ; Kheradpisheh et al., 2018 ) with accuracy similar to convolutional neural networks (Wu et al., 2019 ). It consists of convolutional, pooling, and fully connected layers. The SCNNs extract image features through hierarchical convolutional layers, providing strong image processing capabilities. Each convolutional layer generates many spiking feature maps (SFmaps) from the same number of membrane potentials (Vm). As a kind of SNN, SCNNs also have high energy efficiency in neuromorphic hardware. However, SCNNs must generate several SFmaps to ensure a high processing accuracy, leading to many convolution operations, weights, and Vm storage. This makes deploying SCNNs on edge devices difficult due to the limited computing power, power consumption, and storage capacity. Researchers have made significant efforts to solve this issue. First, some methods are proposed to decrease the timesteps 1 to decrease operations and memory access. SCNNs have achieved high precision with few timesteps (Chowdhury et al., 2021 ), with the spatio-temporal backpropagation (STBP) training method (Zhu and Shi, 2018 ), direct input encoding (Wu et al., 2019 ), and re-training strategy (Chowdhury et al., 2021 ). Second, a series of methods have been proposed to compact SCNNs, such as network pruning (Liu et al., 2022 ; Schaefer et al., 2023 ) to increase the sparsity and low-bit quantization (Kheradpisheh et al., 2022 ; Shymyrbay et al., 2022 ) to reduce the computational precision. However, these studies overlook the similarity between SFmaps, leading to wasteful calculations. Figure 1A displays the SFmaps of the 1st convolutional layer of a typical SCNN. Pairs of similar SFmaps are marked with boxes of the same color. Figure 1B shows the generation process of two similar SFmaps. The input maps are processed through convolution operations to update two Vm maps. Each Vm map generates an SFmap via threshold comparisons. There are minor differences (ΔSFmaps) between two similar SFmaps, but they are obtained through the processes described above, resulting in redundant operations and data volume. The study (Han et al., 2020 ) indicates that similarity in feature maps is vital for achieving high accuracy. Therefore, there are challenges to reduce these redundancies while preserving similar SFmaps. Figure 1 Similarity between SFmaps and their generation process. (A) The SFmaps of a SCNN convolutional layer where those corresponding to the same color boxes have great similarity and (B) generation process of two similar SFmaps. To address this challenge, this work proposes a novel lightweight dual-threshold spiking convolutional neural network (DT-SCNN) model and a variant spatio-temporal back propagation (STBP) training method. We simplify the training process in Chowdhury et al. ( 2021 ) into a two-stage training strategy to train DT-SCNNs with only one timestep. The DT-SCNN uses dual-threshold to obtain two similar spike feature maps from one Vm map, reducing the number of convolutional weights and Vm values by half with minimal impact on the accuracy. As the network model is lightweight and requires only a single timestep, the number of operations and memory access can be significantly reduced, paving the way for low-latency and power-efficient edge visual applications. This work proceeds as follows. Section 2 briefly reviews the general concept of conventional SCNNs and proposes the DT-SCNN model. The training implementation of the DT-SCNN is also introduced. Section 3 analyzes the experimental performance and compares it to other works. Finally, Section 4 discusses and concludes this work.",
"discussion": "4 Discussion This work proposed a lightweight DT-SCNN structure that reduces the number of operations and memory access of SCNNs with minimal accuracy loss. A variant STBP training method with a two-stage strategy reduces the timestep of the DT-SCNNs to 1 to reduce the computing delay and power consumption of SCNNs when deployed on edge devices. Experimental analyzes show that the DT-SCNNs are the best choice for balancing trade-offs between the computational requirements and accuracy for edge applications. This work only investigated the effect of applying two thresholds to convolutional layers. Future work will explore more thresholds or apply multiple thresholds to different types of layers, and expand DT-SCNN to larger networks."
} | 1,698 |
36475796 | PMC9728966 | pmc | 26 | {
"abstract": "Survival of symbiotic reef-building corals under global warming requires rapid acclimation or adaptation. The impact of accumulated heat stress was compared across 1643 symbiont communities before and after the 2016 mass bleaching in three coral species and free-living in the environment across ~900 kilometers of the Great Barrier Reef. Resilient reefs (less aerial bleaching than predicted from high satellite sea temperatures) showed low variation in symbioses. Before 2016, heat-tolerant environmental symbionts were common in ~98% of samples and moderately abundant (9 to 40% in samples). In corals, heat-tolerant symbionts were at low abundances (0 to 7.3%) but only in a minority (13 to 27%) of colonies. Following bleaching, environmental diversity doubled (including heat-tolerant symbionts) and increased in one coral species. Communities were dynamic ( Acropora millepora ) and conserved ( Acropora hyacinthus and Acropora tenuis ), including symbiont community turnover and redistribution. Symbiotic restructuring after bleaching occurs but is a taxon-specific ecological opportunity.",
"introduction": "INTRODUCTION Corals form the structural and biological foundation of tropical reefs—among the most biodiverse ecosystems on the planet. Corals build and maintain reefs through the accretion of skeletons, underpinned by the nutritional symbiosis with photoautotrophic microbes Symbiodiniaceae ( 1 ). Many reef-building corals cannot survive without the transfer of carbon and nitrogen from their algal symbionts ( 2 ). However, this obligate relationship is threatened by rapid global warming where heat wave conditions disrupt the symbiosis and cause corals to lose their symbionts (termed coral bleaching). If heat is severe or prolonged, then coral bleaching can lead to disease or death of the animal host, affecting reefs at global scales ( 3 ). As temperatures continue to rise, extreme episodes of heat stress will increase mass coral bleaching with substantial negative impacts for reef ecosystems. Coral reef biodiversity is already changing in response to warming ( 4 ), with repeated extreme episodes of bleaching leading to the restructuring of corals ( 5 ), fish, and other invertebrate communities ( 6 ). To better predict coral reef futures, there is a need to understand whether climate change will restructure this foundational symbiosis to a more vulnerable or more resilient assemblage ( 7 ). One mechanism of rapid acclimation or adaptation to the environment is through changes to host-associated microbial communities ( 8 , 9 ). Quantifying the propensity for symbionts to change therefore enables the prediction of corals’ evolutionary trajectories during climate change ( 10 ). Corals generally associate with specific symbionts ( 11 – 15 ), with most Pacific coral species hosting more speciose communities compared to other regions ( 13 , 16 ). Coral-symbiont associations are generally stable over time with disruption requiring substantial environmental stress ( 17 ), although specific changes can enhance host survival during short-term or prolonged heat waves ( 18 , 19 ). Although damaging, bleaching can create a high-risk ecological opportunity [sensu ( 20 )] to mitigate stress [adaptive bleaching hypothesis (ABH) sensu ( 21 , 22 )] by partnering with previously undocumented combinations of symbionts better suited to present or future conditions (i.e., those that are heat tolerant). ABH may be a mechanism for rapid acclimation to heat, indicated by host-directed expulsion ( 23 ) and evolutionary selection on symbiont communities ( 24 ) and supported by the availability of physiologically diverse symbionts free-living in the environment ( 25 ). Symbiodiniaceae include hundreds of symbiotic and free-living putative “species” ( 11 – 15 ). The current paradigm is that stress is ameliorated by two mechanisms that restructure the symbiont community: “shuffling” the relative abundance of existing symbionts [predominately from the genera Cladocopium to Durusdinium ( 1 , 26 )] or by “switching” to new symbionts acquired from the environment (potentially from a pool of free-living symbionts in reef sediments). To investigate the impacts of mass bleaching on coral symbioses, we quantified the temporal changes in Symbiodiniaceae communities in the surviving colonies of three abundant coral species and in the environment (reef sediments), across latitudinal and cross-shelf gradients on the Great Barrier Reef (GBR) over a 16-year period (2003–2019; total, n = 1643) and acknowledged that bleaching events were not the only disturbance during this period (e.g., cyclones and crown-of-thorns predation). There are also currently alternative methodologies for describing Symbiodiniaceae communities, each with their merit. Here, we highlight and apply multiple methods [DIV (defining intragenomic variant) and ASV (amplicon sequence variant) approaches; see Materials and Methods] to analyze data that include collections before (“pre”) and after (“post”) the 2016 mass bleaching event from 26 reefs ( Fig. 1 and table S1). Fig. 1. Sampling of corals and the environment along reefs from the northern and central GBR. Samples for genetic analysis of corals’ symbiotic communities included 1644 samples from individual coral colonies of A. hyacinthus , A. tenuis , and A. millepora ( n = 1454) and sediments representing the environmental pool ( n = 189) from 2003 to 2019. Twenty-six reefs were surveyed genetically and varied in their responses to bleaching in 2016 (circles) and 2017 (triangles). The deviance in the maximum degree heating week (mDHW) residual (red to blue) between the accumulated heat stress measured in DHWs and the aerial survey bleaching response is shown per reef surveyed.",
"discussion": "RESULTS AND DISCUSSION Symbiont communities varied among species and environmental gradients The functional diversity within Symbiodiniaceae is extensive ( 1 ) and can be detected across multiple taxonomic levels ( 27 ). Here, we mainly present sequence data using the ASV approach (see Materials and Methods for information regarding Symbiodiniaceae taxonomic methods and further comparisons). Contemporary coral-algal symbioses are globally dominated by stress-sensitive members of the genus Cladocopium ( 1 ), and this pattern was evident before and after bleaching within Acropora spp. ( Fig. 2 and fig. S1). Symbiont communities varied by sample group (i.e., coral host species or sediments), sector (north/central), region (inshore/offshore), and bleaching history (pre- or post-2016) [permutational multivariate analysis of variance (PERMANOVA), P < 0.001; table S2]. Variation was greatest along an inshore to offshore gradient in coral and environment samples [coefficient of determination ( R 2 ) variance explained = 0.1 to 0.39; table S3], consistent with established biogeographical patterns ( 14 , 28 , 29 ). In comparison, variability in symbionts was lower between the northern and central sectors and between pre- and postbleaching time points ( R 2 = 0.02 to 0.09). Bleaching history explained greater variation in Acropora millepora symbiont communities compared to Acropora hyacinthus or Acropora tenuis ( R 2 = 0.09 versus 0.07 and 0.006). We also tested multiple, independent methods on an informative subset of highly variable reefs and time points in the most variable species, A. millepora . This comparison confirmed that both approaches render comparable interpretations, specifically in which the relative proportion of Cladocopium was lower in presamples and replaced by Durusdinium in the postsamples (fig. S2). Fig. 2. Normalized relative abundances (%) of Symbiodiniaceae communities sequenced from coral and environmental samples collected along reefs from the northern and central GBR before and after the 2016 mass bleaching event. Each bar is a separate reef. Barplots depict the variance-normalized relative abundance of the nine Symbiodiniaceae “genera” ( 1 ) (colors from “A to I”) in the three coral species and in the environment categorized by each reefs’ mDHW residual category [colors correspond to H-H (high bleaching–high mDHW), H-L (high bleaching–low mDHW), L-H (low bleaching–high mDHW), and L-L (low bleaching–low mDHW)]. When averaged per paired reef (pre versus post; Figs. 2 and 3 ), only a minority (13.3 to 26.8%) of individual corals sampled contained Durusdinium before the 2016 bleaching ( A. tenuis , 47 of 352 colonies hosted Durusdinium ; A. hyacinthus , 34 of 160; A. millepora , 44 of 164). After bleaching, only A. millepora colonies had increased Durusdinium prevalence (61.4%; fig. S1, A and B). The relative abundances of Durusdinium per reef ranged from <1% before to ~50% after 2016, and changes were highly species-specific and driven by inshore reefs ( Figs. 2 and 3 , fig. S1A, and table S4). One species ( A. millepora ) hosted the highest abundances of stress-tolerant members of the genus Durusdinium ( 19 ), with a notable difference compared to A. hyacinthus and A. tenuis colonies, which exhibited changes only at low relative abundances (<3%), mostly on northern ( A. hyacinthus ) or central, inshore reefs ( A. tenuis ) ( Fig. 4A , tables S4 and S5, and fig. S1A). Fig. 3. Diversity of Durusdinium (“D”) Symbiodiniaceae communities sequenced from coral and environmental samples collected along reefs from the northern and central GBR before and after the 2016 mass bleaching event. Bubble plots depict the diversity of Durusdinium at each reef before and after bleaching with each bar representing a separate reef. Bubbles represent normalized relative abundances, grouped into classes that correspond to lower taxonomic resolution (colors correspond to D1 to D10; NA = unclassified). Boxes in gray denote reefs with paired pre- and post- bleaching samples. Reefs are grouped into the three coral species and in the environment categorized by each reefs’ mDHW residual category (colors correspond to H-H, H-L, L-H, and L-L). Please note that bubble plots are mainly used for depicting diversity and not relative abundances, although dot sizes here are scaled to variance normalized relative abundances. Fig. 4. Changes in Symbiodiniaceae richness in coral and environmental samples collected along reefs from the GBR before and after the 2016 mass bleaching. ( A ) The Chao1 richness of ASVs, from 2003 to 2019, collected from reefs colored by sample type. The number of samples (density, y axis) corresponding to the Choa1 metric for each sample type. ( B ) The means ± SD of Chao1 richness shown as the linear trend of the metric through time. The dashed lines represent the 2002, 2016, and 2017 mass bleaching events. Asterisks (*) signify significant differences calculated with linear mixed effects models (lmer, P < 0.05; exact statistical values in table S7; n values in table S1) in community composition between coral species and environmental samples. The colors correspond to each of the four sample types (gray, A. hyacinthus ; pink, A. millepora ; yellow, A. tenuis ; black, environment). The prevalence of Durusdinium in the environment was consistently high (97.1 to 100%) compared to coral samples (pre, 21 of 21 samples; post, 34 of 35 samples). These values suggest that heat-tolerant Durusdinium was found across almost all environmental gradients and was thus available for coral uptake. The abundance of Durusdinium decreased in the environment after bleaching by >56.6%. This was not explained by sequencing read depth given the greater abundance of reads retrieved after bleaching (mean reads ± SE: pre, 48,982 ± 1493; post, 108,026 ± 1789) but could have been influenced by seasonal fluctuations or water patterns. The increased abundance of Durusdinium in corals (i.e., postbleaching uptake) mostly occurred in A. millepora ( Fig. 3 ) and was limited in A. hyacinthus and A. tenuis . This coral species–specific pattern could have potentially been driven by host genotype or other heritable mechanisms ( 24 , 30 ). Shuffling or switching to specific heat-tolerant symbiont genera may therefore be a specialized mechanism to contend with environmental change and not be as common across the coral phylogenetic tree as previously hoped ( 19 , 31 – 33 ). High variation in changes in symbiont abundances and communities at the reef and host species levels—as shown through sequence variation ( Fig. 2 )—suggests that corals have several potential mechanisms to acclimate to rapid ocean warming. Understanding this variability will be essential to predicting reef vulnerability and recovery potentials. The three coral species varied significantly in symbiont ASV community composition (PERMANOVA; table S1). The most abundant ASVs differed by coral species (in A. hyacinthus : C3k, Cspc, and C29; A. millepora : Cspc and C3k; A. tenuis : Cspc, C1m, and C3k; table S5) and may influence different physiological tolerances across corals. Free-living environmental symbiont communities were highly diverse compared to those in corals, spanning all major Symbiodiniaceae genera ( Fig. 2 , environment) ( 1 ) and may be a critically important reservoir of unknown but functionally important symbiont diversity for corals. Heat-tolerant Symbiodiniaceae (e.g., D1 and D1a) were abundant in the environment, as were the common and shared ASVs among all three Acropora species (C1m, Cspc, and C3k). Leading up to 2016, other disturbances had already affected the GBR ( 34 ), including bleaching. The 2016 cutoff was selected as a feature that best represents responses to greater accumulated thermal stress over this long sampling period as opposed to the specific 2016 heat wave and acknowledges that the 2016 bleaching event represented the most extreme bleaching that had occurred before or after within our sampling time frame from 2003 to 2019 ( 7 ), with previous bleaching occurring outside our sampling window in 1998 and 2002 ( 35 ). Although 2005 to 2015 were typified by relatively low sea surface temperature and low summertime marine heat waves ( 35 ), given how bleaching impacts varied across the GBR in 2016, we also explored a subset of the data (2011–2019) to examine whether earlier warming and bleaching were driving community variability (i.e., in the years directly after the 2002 bleaching). PERMANOVA results using this subset of data (table S10) were highly similar to previous results using the full time-course data (table S2) and underscored our previous findings that symbiont communities varied by sample group (i.e., coral host species or sediments), sector (north/central), region (inshore/offshore), and bleaching history ( P < 0.001; table S10). Symbiont community richness and diversity varied through time Network modeling suggests that corals with symbiont communities characterized by high richness are more susceptible to bleaching ( 36 ). To explore this, we examined patterns in richness (Chao1 index of ASVs) and diversity (number of ASVs) of symbiont communities in corals and in the environment, in which both metrics varied before and after the 2016 bleaching ( Fig. 4 ). There was a general pattern of decline in corals over the sampling times, as well as when examined before and after bleaching, with variability between sampling time points. After accounting for interreef variability, richness significantly decreased after 2016 in A. hyacinthus and A. millepora but not A. tenuis (tables S6 and S7; lmer, P < 2 × 10 −16 , P < 0.0006, and P < 0.38, respectively). In the environmental samples, richness estimates increased in both regions by ~50%, most notably in the north ( P < 2 × 10 −16 ). Mean diversity decreased significantly in all species but not A. millepora (tables S8 and S9). During this time frame, the GBR has experienced several environmental events, including less severe bleaching and cyclones ( 34 ); therefore, these changes in abundance may be a response to greater accumulated thermal stress over this ~16-year period as opposed to the specific 2016 heat wave. Current management action often focuses on protecting coral diversity. However, this potential loss in symbiont diversity is concerning given the previous modeling that suggests protecting generalist symbionts with high heat tolerances, but not coral host diversity, contributes more to coral reef resilience ( 37 ). This highlights the importance of protecting the coral-symbiont relationship. Changes in richness and diversity can also signal larger shifts in ecosystem functioning. As observed on Caribbean reefs ( 38 ), introduced Durusdinium can competitively displace other symbionts after repeated environmental stress, driven by the rapid spread of these tolerant “opportunists” ( 39 ) into novel locations and hosts. We detected this pattern here, where A. millepora exhibited, on average, some of the lowest symbiont richness overall paired with the highest Durusdinium abundance. Also noteworthy were the increases in Durusdinium diversity in the environmental samples after bleaching ( Fig. 2 ). This increase may have been caused by strong selection from extreme heat waves influencing rapid diversification (e.g., adaptive radiation) of free-living symbionts or potentially other disturbances, including the 2011 freshwater influx that may have reduced free-living symbiont diversity in the pre-2016 samples. There were also substantial linkages between the species displaying the largest changes in relative abundances and richness ( A. millepora ) and the environmental communities. In this, 29% of Durusdinium ASVs were found and shared in A. millepora and the environment after bleaching, and two of the 10 most abundant coral symbionts were also shared. The large postbleaching shifts in the sediment symbiont pool ( Figs. 5 and 6 and fig. S1) and the increasing richness and diversity in the short-term suggest similar responses to repeated disturbance as those observed in the Caribbean ( 38 ). The longer-term consequences of these changes remain unknown. Fig. 5. Shuffling of Symbiodiniaceae communities after the 2016 bleaching on the GBR. Symbiont dynamics were explored by quantifying the change in variance normalized relative abundances of symbionts compared to their change in prevalence (presence or absence). Shuffling of Symbiodiniaceae communities, in which ASVs that significantly changed in variance normalized relative abundances ( P adj < 0.05) either before or after 2016 are colored by genus. Specific symbiont taxa are labeled. ASVs recovered from reefs from central (triangle) and northern (circle) locations are indicated. Fig. 6. Switching of Symbiodiniaceae communities after the 2016 bleaching on the GBR. Community turnover describes the gain and loss (switching) of Symbiodiniaceae ASVs per reef that were not reported before 2016 but were detected after 2016 (gain) and were reported before 2016 but not after 2016 (loss), as well as the sum of these ASVs (total) across the sample types. Asterisks (*) signify significant differences (lmer, P < 0.05) in ASV community composition between coral species and environmental samples ( P values are referenced within the text; n = values in table S1). The colors correspond to each of the four sample types (gray, A. hyacinthus ; pink, A. millepora ; yellow, A. tenuis ; black, environmental). Symbiodiniaceae shuffling and switching in response to mass bleaching Symbiont communities extend the host phenotype, increasing acclimation and adaptation potential of the organism. Although generally explored only at the taxonomic resolution of genera ( 33 ), here, we quantify shuffling and switching at the more ecologically relevant level ( 40 , 41 ) of ASVs and then assess the propensity for both mechanisms at the reef level (see Materials and Methods). Overall, symbiont communities were restructured after the 2016 bleaching in both coral and environmental samples, driven by changes in specific symbionts between samples collected before and after the 2016 bleaching ( Figs. 5 and 6 ). The three Acropora species exhibited relatively more shuffling compared to switching (22.8% of A. millepora ASVs, 18% A. hyacinthus , and 36.1% A. tenuis ; figs. S1A, S3, and S4 and table S4). In addition to the classic shuffling response (from Cladocopium to Durusdinium ) in A. millepora , the other two coral species either shuffled or switched taxa within Cladocopium (C3k, C1m, and Cspc; Fig. 5 ). Shuffling occurred in the environmental samples but was relatively less common and included a greater diversity of symbiont taxa ( Fig. 5 and fig. S4), decreases in Durusdinium abundance, and the appearance of new taxa after bleaching ( Fig. 5 ). Symbiont switching occurs when new taxa appear in the established community within the host coral, although evidence for corals to accommodate these changes remains limited ( 42 , 43 ). Here, we apply a standard ecological metric, community turnover ( 4 ), to quantify the proportion of ASVs gained or lost in samples taken before and after bleaching, at the reef level ( Fig. 6 ). Switching was prevalent across many sampled reefs and was significantly different among the three Acropora species and environmental communities even after accounting for between-reef differences (lmer, P ≤ 0.0001 to 0.0033; Fig. 6 ; total percentage of ASVs that changed, 68.1 ± 3% SE). Postbleaching switching of symbiont taxa was twice as high in the environment [91.6 ± 0%; mean community turnover metric (MRS), 108.5 to 354.2] compared to corals (60.9 ± 2.7%; MRS, 35.8 to 193), mostly driven by the appearance of significantly more new taxa after bleaching relative to all three coral species (72.6 ± 3.1% versus 29.5 ± 3.6%; all comparisons, P ≤ 0.0001 to 0.009). The proportion of symbionts lost was only significantly less compared to A. hyacinthus and A. millepora ( P = 0.005 to 0.03). Overall, total turnover in corals was driven by significantly fewer losses of ASVs within A. hyacinthus , especially in offshore, northern reefs, compared to A. tenuis ( P = 0.005 to 0.03). This indicates the relative stability of A. hyacinthus communities. Together, we conclude that symbiosis flexibility to environmental change varies across coral species, driven mostly by symbiont taxa lost, not gained. Symbiont community changes differed depending on bleaching exposure at each reef The mass coral bleaching observed in 2016 was associated with pronounced accumulated heat exposure ( Figs. 1 and 7A ) ( 5 ), with extreme maximum degree heating weeks (mDHWs) compared to previous events from 1985 to 2015 ( 35 ). To better understand the variability in responses across regions on the GBR, we examined the expected versus observed bleaching responses across each reef. We calculated the deviance between the expected bleaching responses given the amount of accumulated heat for each reef and the actual bleaching observed, expressed as the residuals between the observed mDHW and the Bayesian posterior predicated mean mDHW (see Materials and Methods) and classified these responses into four categories ( Fig. 7A ). As expected ( 44 ), greater mDHWs were associated with higher aerial bleaching scores [denoted as high bleaching–high mDHW (H-H)], high bleaching even under relatively low mDHW (H-L), low bleaching and high mDHW [low bleaching–high mDHW (L-H)], or no bleaching [low bleaching–low mDHW (L-L)]. Notably, four reefs diverged from the typical bleaching conditions associated with high temperatures and did not bleach even when exposed to high heat stress (L-H), a phenomenon also observed in some Eastern Tropical Pacific reefs ( 45 ). Symbiont communities differed between the four coral bleaching response categories ( Fig. 7, B to E ), mostly driven by the variability within each group. Differences, regardless of time point, in the dispersion of samples suggest variability in the responses of symbiont communities within each bleaching group (table S11). This observation is supported by significant differences in centroid and dispersion differences (adonis test; table S12). Fig. 7. Symbiont community responses in corals and the environment on the GBR. ( A ) Bleaching experienced by each reef in 2016 and 2017 compared to the accumulated heat stress experienced in mDHWs. The deviance of the relationship between mDHW and bleaching score is represented by the mDHW residual. This residual was used to classify each reef into the following four categories: H-H, H-L, L-H, and L-L. Gray densities indicate the posterior distribution of mDHW for each bleaching category, while black points and bars indicate the mean and 66% and 95% intervals for the posterior. Boxplots include the median values (center lines), upper and lower quartiles (box limits), 1.5× interquartile range (whiskers), and outliers (points). ( B to E ) Symbiodiniaceae communities plotted in ordination space and colored by DHW deviance categories. Nonmetric multidimensional scaling (NMDS) using Bray-Curtis distance of variance normalized ASVs for the four sample sources. Individual points represent symbiont communities categorized as either before 2016 (circle) or after 2016 (triangle) sampling. Asterisks (* and **) signify significant differences ( P < 0.05) in the multivariate dispersion between the four mDHW categories (*) or differences in the combined variability of the centroids and dispersion of the four categories (**). Exact P values in tables S11 and S12; n = values in table S1. None of the A. tenuis communities from H-L reefs were associated with Durusdinium . Communities from H-H reefs were highly variable in the three coral species and the environment, especially compared to “baselines” demonstrated by L-L reefs ( Fig. 7C ), supporting evidence that stress promotes variability ( 46 , 47 ). Symbiont communities were less variable in L-H reefs, especially in A. tenuis , and may be indicative of resilient coral symbiont communities. H-H reefs generally had divergent communities compared to the other categories (notably in A. tenuis ; Fig. 7D ). This suggests that communities in some species such as A. tenuis and A. millepora were distinct before and after bleaching, not only in response to bleaching ( Fig. 7 ). A. millepora communities from bleached reefs (H-H) had distinct communities compared to those that do not bleach (L-L) ( Fig. 7C ). Samples from L-H reefs followed similar patterns to L-L reefs and did have Durusdinium after heat stress ( Figs. 2 and ( 7C ). Last, there was high overlap in the shifts associated with changes in Cladocopium ASVs in coral and environmental samples, suggesting that these symbionts may contribute heavily to bleaching responses. These data demonstrate that symbiont community structure varies under different bleaching histories and outcomes. Here, we provide key insights that the consequences of mass coral bleaching on algal symbiont communities across the GBR are of both loss and gain. The loss of dominant and background generalist symbionts can reduce corals’ ability to resist bleaching ( 37 ), limit recovery and resilience ( 13 , 16 ), and have negative, ecosystem-wide effects on the modeled long-term stability of symbioses ( 13 , 16 , 41 ). Here, some corals, but not all, responded by increasing the abundance of heat-tolerant Durusdinium , likely from their availability in the free-living sediment community. Knowledge of the species-specific ability of corals to shuffle or switch symbionts is essential for managing and conserving these ecosystems under continual ocean warming. Although the future risk to reef building corals from climate change around the world remains, variation in corals’ responses highlights pathways for increased reef resilience."
} | 6,957 |
29192903 | PMC5739009 | pmc | 27 | {
"abstract": "The association between corals and photosynthetic dinoflagellates ( Symbiodinium spp.) is the key to the success of reef ecosystems in highly oligotrophic environments, but it is also their Achilles‘ heel due to its vulnerability to local stressors and the effects of climate change. Research during the last two decades has shaped a view that coral host– Symbiodinium pairings are diverse, but largely exclusive. Deep sequencing has now revealed the existence of a rare diversity of cryptic Symbiodinium assemblages within the coral holobiont, in addition to one or a few abundant algal members. While the contribution of the most abundant resident Symbiodinium species to coral physiology is widely recognized, the significance of the rare and low abundant background Symbiodinium remains a matter of debate. In this study, we assessed how coral– Symbiodinium communities assemble and how rare and abundant components together constitute the Symbiodinium community by analyzing 892 coral samples comprising >110 000 unique Symbiodinium ITS2 marker gene sequences. Using network modeling, we show that host– Symbiodinium communities assemble in non-random ‘clusters‘ of abundant and rare symbionts. Symbiodinium community structure follows the same principles as bacterial communities, for which the functional significance of rare members (the ‘rare bacterial biosphere’) has long been recognized. Importantly, the inclusion of rare Symbiodinium taxa in robustness analyses revealed a significant contribution to the stability of the host–symbiont community overall. As such, it highlights the potential functions rare symbionts may provide to environmental resilience of the coral holobiont.",
"conclusion": "Conclusions Our results reveal a highly structured community assembly in coral host– Symbiodinium associations and suggest functional importance of the Symbiodinium community as the sum of all of its members, not only of the most dominant members. The host– Symbiodinium association is shaped by a few dominant symbionts as well as a rare, diverse community characterized by robust scaling power laws that escaped previous analyses based on reduced sample size and/or sequencing depth. The existence of distinct co-occurrence networks involving rare members of the symbiont pool suggests that the functional role of Symbiodinium in the holobiont is not determined by a single individual taxon, but rather by assemblages of symbionts. Intriguingly, the inclusion of rare microbial members in robustness analyses resulted in a significant stability increase of the host–symbiont community and thus further highlights a potential role that rare symbionts may have in corals threatened by environmental change. Given that host– Symbiodinium association plays a central role in coral holobiont response to perturbation and resilience, future studies should focus on experimental elucidation of the specific functions contributed to the holobiont by rare Symbiodinium members and the synergistic effects arising from their interactions. Data accessibility Sequence data used in this publication can be accessed in the NCBI Sequence Read Archive ( http://www.ncbi.nlm.nih.gov/sra ) under accession number PRJNA306572.",
"introduction": "Introduction Corals and coral reef ecosystems are immediately threatened by global climate change and local anthropogenic impacts ( Hoegh-Guldberg et al. , 2007 ). This is because the obligate association of corals with dinoflagellate endosymbionts of the genus Symbiodinium represents both the success and the Achilles’ heel of these ecosystems. Fueled by their endosymbiont’s provision of energy in the form of photosynthates, corals provide the structural foundation for the ecologically and economically important reef ecosystems ( Roberts et al. , 2002 ). Although reef ecosystems cover only a small percentage of the world’s oceans, almost a third of global marine biodiversity is dependent on their functional integrity ( Reaka-Kudla, 1997 ). But ocean warming and other impacts lead to the breakdown of the coral– Symbiodinium symbioses (coral bleaching) and the ongoing decline of coral reef habitats on a global scale ( Hoegh-Guldberg et al. , 2014 ; Hughes et al. , 2017 ). Based on this premise, fundamental questions to help forecast the future of coral reefs emerge. For instance, how are Symbiodinium communities structured in their respective coral hosts and across reef ecosystems, how will this change under future ocean conditions, and does this affect coral host and ecosystem resilience? Research on the physiology of Symbiodinium is typically focused on the most abundant symbiont(s) within the host ( Cooper et al. , 2011 ; Ziegler et al. , 2015 ) or on cultured clonal isolates ( Schoenberg and Trench, 1980 ; Suggett et al. , 2015 ; Parkinson et al. , 2016 ). For instance, the bleaching response and stress tolerance of the coral host are affected by the associated Symbiodinium type ( Berkelmans and van Oppen, 2006 ; Abrego et al. , 2008 ), where a type is defined based on ribosomal ITS2 DNA sequences, which, in combination with other genetic markers, is used to distinguish Symbiodinium species ( LaJeunesse et al. , 2012 , 2014 ). However, these host–symbiont associations are not static and may change as a mechanism to increase the size of the ecological niche ( Ziegler et al. , 2015 ) or in response to environmental perturbation ( Silverstein et al. , 2015 ; Boulotte et al. , 2016 ). The replacement of stress-susceptible with stress-resistant Symbiodinium according to the adaptive bleaching hypothesis ( Buddemeier and Fautin, 1993 ) is accomplished by the substitution of dominant symbionts with rare types already present within hosts and within the broader environment ( Berkelmans and van Oppen, 2006 ; Boulotte et al. , 2016 ). Although most coral colonies are primarily associated with one Symbiodinium type at a given time, a large cryptic diversity of low abundant Symbiodinium is becoming increasingly apparent through the combined use of next-generation sequencing (NGS) and large-scale sampling of coral–symbiont assemblages ( Quigley et al. , 2014 ; Thomas et al. , 2014 ; Hume et al. , 2016 ; Ziegler et al. , 2017 ). These NGS studies have in common that they describe the existence of a rare and diverse Symbiodinium community, in addition to an abundant community composed of one or a few symbiont taxa. This observation of few abundant and many rare algal symbionts associated with a coral host parallels the structure of bacterial communities where the so-called rare bacterial biosphere ( sensu \n Sogin et al. , 2006 ; Pedrós-Alió, 2012 ) supposedly represents the predominant sector of bacterial diversity ( Skopina et al. , 2016 ). The rare bacterial biosphere has been shown to fulfill essential functions associated with nutrient cycling, the degradation of pollutants, host health, and rare microbes may further also enhance functionality of abundant microbes (reviewed in Jousset et al. , 2017 ). In ocean and freshwater environments, rare members of bacterial and protist communities tend to be more active (as measured by RNA to DNA ratios) and thus contribute overproportionally to ecosystem function ( Campbell et al. , 2011 ; Debroas et al. , 2015 ). In the case of corals and their algal endosymbionts, the putative significance of a rare background Symbiodinium community remains largely unresolved ( Boulotte et al. , 2016 ; Lee et al. , 2016 ). A key question is whether the rare biosphere of microbial symbionts in general, and that of algal endosymbionts in particular, has a functional role in the coral holobiont. Building onto approaches and insights from other study systems, we further investigated this research question. For instance, the diversity of species and the complexity of their interactions are linked to the stability of ecosystems in a dynamic relationship with environmental drivers ( Kondoh, 2003 ; Ives and Carpenter, 2007 ). This has been particularly well-studied for plant communities in which phylogenetic and functional species diversity promote ecosystem stability (for example, Cadotte et al. , 2012 ). Mathematical approaches to describe and understand relationships introduce the concepts of (1) ‘networked buffering‘, which relies on community components to perform multiple functional roles and a partial overlap between their functional capabilities ( Whitacre and Bender, 2010 ) and (2) ‘buffered qualitative stability', which predicts robustness of a network (of genes or other biological entities) by its responses to arbitrary (that is, modeled) perturbations ( Albergante et al. , 2014 ). Following this, in this study we used network modeling to understand the composition and structure of coral host-associated algal endosymbiont communities. Our aim was to explore the potential role of rare Symbiodinium community members in regard to Symbiodinium community composition as a whole and in regard to the robustness (that is, stress tolerance) of the coral holobiont community composition under disturbance.",
"discussion": "Discussion The finding of a rare and diverse Symbiodinium community within coral hosts parallels the existence of a rare bacterial biosphere ( sensu \n Sogin et al. , 2006 ). This cryptic and diverse community structure of Symbiodinium within corals was only uncovered recently through the application of NGS approaches. Further, we found this community to be characterized by a rank-abundance pattern following a power-law relationship that matches those reported for planktonic bacterial assemblages (cf. Figure 2 in Pedrós-Alió, 2012 ). A key question is whether the rare biosphere in general, and that of coral–algal symbionts in particular, have a functional role in the ecosystem ( Lee et al. , 2016 ). Several lines of evidence suggest that the uneven Symbiodinium community structure, composed of few abundant and many rare members, may potentially present a so-far overlooked aspect of coral holobiont functioning. First, the prevalent and dominant components of the symbiont community cannot explain the breadth of responses of corals to stress or their resilience to locally restricted environmental perturbations, precisely because the dominant components are shared across wide environmental gradients and across reefs with contrasting status. This further points to the differentiating elements of the assemblage, composed of the rare symbionts, as the components that may underpin such diversity in response. Interactions between members of the bacterial symbiont community can support holobiont functions and traits that cannot be supplied by single taxa in isolation, for example, synergistic effects between pairs of bacterial taxa increase fungal resistance in the freshwater polyp Hydra compared to polyps colonized by single bacterial taxa ( Fraune et al. , 2015 ). Similarly, the functional roles of Symbiodinium assemblages are presumably not carried out by individual OTUs but by assemblages of OTUs, as suggested by the existence of distinct co-occurrence networks involving rare components of the symbiont pool revealed by our analysis. This is in accordance with patterns of co-occurrence of diverse micro- and macroorganisms, which suggests that non-random community assembly supporting ecosystem function and stability may be a general feature of all domains of life ( McCann, 2000 ; Horner-Devine et al. , 2007 ). Second, rare bacterial community members are shown to fulfill essential functions while at low abundance. They are often more active than abundant members as revealed by rRNA:rDNA ratios and respiration rates ( Dimitriu et al. , 2010 ; Campbell et al. , 2011 ). Rare microbes sustain a vast functional gene pool and can indirectly enhance functionality of abundant microbes (reviewed in Jousset et al. , 2017 ). For example, for coral-associated bacterial communities it was suggested that two widely distributed, but low abundant bacterial strains may contribute large proportions of defined metabolic processes within the network of coral host–bacterial functions ( Ainsworth et al. , 2015 ). Furthermore, functional diversity of rare species is well documented in assemblages of macroorganisms such as rainforest trees, reef and stream fishes, or birds ( Mouillot et al. , 2013 ; Leitão et al. , 2016 ). Symbiodinium also represent a highly diverse group that was described on the level of differences between species and between strains within a species ( Suggett et al. , 2015 ; Parkinson et al. , 2016 ). Given the large (functional) diversity of Symbiodinium and the similarity of Symbiodinium community structure to that of bacterial communities ( Pedrós-Alió, 2006 , 2012 ), we suggest that similar assumptions can be made regarding the repertoire and contribution of the rare Symbiodinium biosphere to community function, and further research should probe the functions provided by the rare Symbiodinium biosphere to the coral holobiont while at low abundance. Further, the network co-occurrence analysis suggests the potential for the rare Symbiodinium community to play a role in conferring resistance to the coral host by replacing dominant Symbiodinium types lost under environmental stress. The emergence of a regionally cryptic Symbiodinium species as prevalent in the heat-selected corals of the Persian/Arabian Gulf ( Hume et al. , 2016 ) supports the notion that a phylogenetically diverse and rare algal biosphere may serve as a source of genomic innovation and standing genetic diversity providing the raw material for adaptation to changing environmental conditions ( Sogin et al. , 2006 ). As such, an otherwise rare component can become the dominant member and represents a source of adaptation to the coral host–symbiont system ( Hume et al. , 2016 ); albeit this reorganization may sometimes come with a trade-off in general ecological performance, such as decreased coral growth rates ( Pettay et al. , 2015 ). The extensive data set used in our study entails samples with mixed communities of Symbiodinium that are comprised of combinations of various clade C symbionts, combinations of clade C and clade D symbionts, or clade A with clade C or clade D symbionts in varying abundances. The patterns in these data largely correspond to previously reported switching or shuffling scenarios for clade C to clade D dominance ( Jones et al. , 2008 ; Silverstein et al. , 2015 ; Boulotte et al. , 2016 ) or vice versa ( Jones et al. , 2008 ), between clade C types ( Boulotte et al. , 2016 ), and between clade A and C ( Hume et al. , 2015 ). In addition, some coral species, such as Pocillopora verrucosa and Porites lutea (both abundant in this data set) associate with a range of different symbionts throughout reef locations in the Red Sea ( Sawall et al. , 2015 ; Ziegler et al. , 2015 ). Thus, they may at least potentially be able to undergo switching or shuffling events under stress scenarios. Despite the large gap of knowledge under which environmental conditions shuffling and switching may occur, which coral species can change their symbionts, or whether this is a selective mechanism, our modeling approach helps exploring the role of rare Symbiodinium OTUs in host– Symbiodinium community dynamics, thereby providing directions for future research. The role that Symbiodinium background diversity plays in the robustness of the host–symbiont community assembly is a fundamental, but yet unresolved question. Previous models suggested high robustness of highly uneven host–symbiont communities ( Fabina et al. , 2013 ), but these conclusions were drawn without the exclusion of intragenomic ITS2 sequence variants that are prone to confound diversity estimates. In our study, we aimed to remove redundancy from intragenomic diversity by applying an OTU-based clustering approach. Consequently, our results suggest a rather low robustness of a static system when dominant Symbiodinium components are lost, while rare members remain unchanged and cannot become dominant. We integrated recent studies that suggest that rare Symbiodinium community members can become dominant over short time scales following stress events ( Boulotte et al. , 2016 ) and during long-term environmental adaptation ( Hume et al. , 2016 ) by conducting an adaptive robustness analysis (that is, allowing co-occurring endosymbionts to replace removed members). Comparing the results from both modeling approaches revealed a significant increase in the stability of the host–symbiont community when rare members can become dominant and replace the lost components. Thus, our model highlights a potential role for rare symbionts upon environmental change. It should be noted, however, that although our findings support a potential role of Symbiodinium community composition for holobiont functioning at large and for background symbionts as mediating components that buffer the community against stressors in particular, detailed and rigorous experiments should be conducted to confirm this notion."
} | 4,289 |
37672573 | PMC10482344 | pmc | 29 | {
"abstract": "Bioconversion of a heterogeneous mixture of lignin-related aromatic compounds (LRCs) to a single product via microbial biocatalysts is a promising approach to valorize lignin. Here, Pseudomonas putida KT2440 was engineered to convert mixed p-coumaroyl– and coniferyl-type LRCs to β-ketoadipic acid, a precursor for performance-advantaged polymers. Expression of enzymes mediating aromatic O -demethylation, hydroxylation, and ring-opening steps was tuned, and a global regulator was deleted. β-ketoadipate titers of 44.5 and 25 grams per liter and productivities of 1.15 and 0.66 grams per liter per hour were achieved from model LRCs and corn stover-derived LRCs, respectively, the latter representing an overall yield of 0.10 grams per gram corn stover-derived lignin. Technoeconomic analysis of the bioprocess and downstream processing predicted a β-ketoadipate minimum selling price of $2.01 per kilogram, which is cost competitive with fossil carbon-derived adipic acid ($1.10 to 1.80 per kilogram). Overall, this work achieved bioproduction metrics with economic relevance for conversion of lignin-derived streams into a performance-advantaged bioproduct.",
"introduction": "INTRODUCTION Lignin, an aromatic polymer that comprises 15 to 45% of terrestrial plant biomass ( 1 ), is an underutilized by-product in lignocellulosic biorefineries. Conversion of lignin to valuable chemicals is therefore of interest to maximize the carbon efficiency, economic viability, and sustainability of lignocellulosic biorefining ( 2 , 3 ). A central challenge in lignin valorization is the intrinsic chemical heterogeneity of lignin and its depolymerization, regardless of the approach, results in a mixture of chemicals ( 4 – 7 ). Biological funneling, the use of a microbial biocatalyst to convert a chemical mixture to a convergent bioproduct, has emerged as a promising strategy toward overcoming the challenge of heterogeneity ( 8 – 12 ). In addition, using this approach, the bioproduct is tunable by engineering the microbial host ( 9 – 12 ). Biological funneling for the conversion of lignin-related aromatic compound (LRC) mixtures to performance-advantaged bioproducts ( 13 , 14 ) has been demonstrated, and some of the associated bioproducts include 2-pyrone-4,6-dicarboxylic acid ( 15 – 17 ), itaconic acid ( 18 ), polyhydroxyalkanoates ( 8 , 19 , 20 ), cis , cis -muconic acid ( 21 – 24 ), vanillin ( 25 , 26 ), substituted styrene molecules ( 27 ), pyridine-2,4-dicarboxylic acids ( 28 , 29 ), and β-ketoadipic acid ( 30 ), among others, reviewed recently by Weiland et al. ( 12 ). β-Ketoadipic acid is a six-carbon dicarboxylic acid with a β-ketone ( Fig. 1 ) that can be polymerized with hexamethylenediamine into a nylon-6,6 analog with reduced water permeability relative to nylon made using petroleum-derived adipic acid ( 31 , 32 ), making it a promising chemical precursor for a performance-advantaged bioproduct. Fig. 1. Metabolic pathway for the biological conversion of p -coumarate and ferulate to β-ketoadipate. The genetic modifications applied to Pseudomonas putida are depicted with an “X” or a filled circle with a “+” for gene deletion or gene overexpression, respectively. N/E, nonenzymatic. The β-ketoadipate pathway is found in diverse soil microbes ( 33 ) and enables convergent and atom-efficient (1 mol/mol) conversion of aromatic LRCs, such as the hydroxycinnamic acids p -coumarate and ferulate ( 34 ), to β-ketoadipate, which is further catabolized to enter the tricarboxylic acid cycle ( Fig. 1 ). Pseudomonas putida KT2440 (hereafter P. putida ) is a Gram-negative soil bacterium with fast growth, high tolerance to cytotoxic compounds ( 35 ), extensive genome engineering tools ( 36 , 37 ), and a native β-ketoadipate pathway. For these reasons, P. putida is widely regarded as a promising chassis for industrial bioconversion processes ( 38 , 39 ), including for complex substrates such as those derived from lignin ( 12 ) and plastics ( 40 , 41 ). P. putida was previously engineered for the conversion of protocatechuate ( 42 ), 4-hydroxybenzoate ( 31 ), or p -coumarate ( 43 ) to β-ketoadipate by the deletion or repression of pcaIJ, encoding the 3-oxoadipate coenzyme (CoA) transferase PcaIJ. However, metabolic bottlenecks now limit β-ketoadipate productivity, and therefore, further strain and bioprocess development are necessary to achieve economically viable LRC bioconversion to β-ketoadipate. In this work, we engineered P. putida and developed bioprocesses for the conversion of mixed LRCs to β-ketoadipate to improve cell performance and overall productivities, as this has been identified as a primary cost driver ( 6 ). Specifically, we report that overexpressing genes that encode enzymes mediating 4-hydroxybenzoate hydroxylation and vanillate O-demethylation and deletion of the gene encoding the Crc global regulator reduces intermediate accumulation. Bioprocess development with strains that harbor these changes resulted in conversion of p -coumarate and ferulate to β-ketoadipate at productivities exceeding 1 g/liter per hour and yields of 1.0 mol/mol (100% of theoretical). Last, LRCs extracted from lignin-rich liquor derived from the alkaline pretreatment of corn stover [alkaline pretreated liquor (APL)] were fed as solids to bioreactor cultivations, resulting in β-ketoadipate production of 25 g/liter at 0.66 g/liter per hour, with an overall yield of 0.10 g/g from corn stover–derived lignin. Given these process parameters, technoeconomic analysis (TEA) predicted a β-ketoadipate minimum selling price (MSP) of $2.01/kg and life cycle assessment (LCA) predicted a greenhouse gas (GHG) emission of 1.99 kg carbon dioxide equivalent (CO 2e )/kg, a cumulative fossil energy consumption of 24.8 MJ/kg, and a water usage of 32.4 liter/kg with the process.",
"discussion": "DISCUSSION This work demonstrates a bioconversion process from mixed model aromatic compounds and lignin-derived substrates to β-ketoadipate at high titers, rates, and yields. The high productivity coupled with theoretical yields from LRCs to β-ketoadipate led to an estimated MSP of $2.01/kg in a facility with an ouptut of 100,000 metric tons (MT)/year and operating with the best experimentally demonstrated metrics reported in this study. This price should be competitive with most diacid precursors used in the manufacturing of nylon-6,6, either fossil- or bio-based, such as adipic acid [global market of >3 million MT/year ( 32 )]. Cost competitiveness was also observed for biological conversion of glucose to β-ketoadipate ( 32 ). Moving forward, reduced glucose demand, by-product recovery, and increased yields of bioavailable monomers from lignin are among the top process considerations for improvement. In addition, previous studies in our team demonstrated that β-ketoadipate can be isolated and purified from bioreactor broths for downstream polymerization efforts ( 31 , 32 ). In this work, we achieved β-ketoadipate productivities higher than 1 g/liter per hour from p- coumarate and ferulate ( Fig. 7 ). At the highest feeding rate tested, these LRCs accumulated immediately and, to a greater extent, than quantified intermediates, which suggests that to exceed a productivity of ~1 g/liter per hour, the potential bottlenecks derived from Fcs- or Vdh-mediated reactions and/or import of aromatic compounds to the cell need to be relieved. To evaluate the former bottleneck, free CoA and adenosine 5′-triphosphate should be quantified, along with the CoA-esters, as these may be limiting feruloyl-CoA or p- coumaryl–CoA biosynthesis. To evaluate substrate import-related bottlenecks, transporter engineering to modulate the expression of the MFS sympoter hcnK ( 60 ) and/or the putative porin PP_3350 ( 61 ) could be considered. Previously, the deletion of PP_3350 was found to be beneficial at high (20 g/liter) concentrations of p -coumarate ( 61 ), although ideal expression levels may be dependent on the bioprocess strategy (e.g., batch versus fed-batch). As expected, many of the genetic interventions previously shown to improve muconate production were also beneficial for β-ketoadipate production. An exception was deletion of gacA/S, which has been shown to be beneficial for production of muconate and indigoidine from p -coumarate or glucose ( 53 , 54 ), but was unsuccessful here likely due to the mixing of substrates [e.g., fitness defects are observed in p -coumarate and glucose ( 62 )]. While titers are similar for β-ketoadipate from p -coumarate and ferulate, β-ketoadipate from 4-hydroxybenzoate ( 31 ), and muconate from p -coumarate ( 46 ), all at ~40 g/liter, β-ketoadipate productivity was >2× higher in this study compared to muconate productivity [0.5 g/liter per hour; ( 46 )]. The underlying reason for this difference in productivity is unknown; postulations include that cofactors uniquely required for muconate production [e.g., prenylated FMN for the protocatechuate decarboxylase, AroY; ( 63 , 64 )] hinder cellular productivity. Further process development efforts could follow multiple tracks, many of which focus on reduction of consumable (e.g., LRC, glucose, and base) costs. To reduce glucose costs, using alternative carbon sources already present in the feed, such as acetate or glycerol in APL, for cell biomass generation should be explored. To reduce LRC costs, which can vary greatly by biomass source, biorefining configuration, depolymerization strategy ( 1 , 7 , 65 ), and the use of lower-purity lignin sources should be considered ( 66 ). Aqueous LRC feeds require NaOH, or another base, to solubilize the substrates, which is both an added cost and potential source of cytotoxicity. Strain engineering for sodium tolerance and the use of powder (solid) feeders may help reduce the amount of salt required, thereby reducing cost and mitigating sodium stress. In situ product removal is another approach to increase production by mitigating product toxicity ( 67 ). TEA and LCA results predict that ammonium sulfate recovery could reduce the MSP and improve the bioprocess environmental sustainability by >15%. Recycling part of the ammonium sulfate back to bioreactors is a potential strategy to offset fresh nitrogen demand. However, to validate the GHG and FEC results for ammonium sulfate as a coproduct, a consequentional LCA that takes into account the market saturation and the rebound effect would need to be performed. Alternatively, the current baseline market value–based system-level allocation method for this multifuntional system can better capture each respective process train’s technical attributes in quantifying each individual product’s GHG emissions and other impacts on its own accord ( 68 ). The development and integration of scalable upstream processes that provide high yields of bioavailable monomers will enable increased overall lignin conversion, a critical factor in the process economics ( 6 ). While this study displays some of the highest product titers and productivities from a lignin stream, the LRC extraction/separation approach used here was not optimized. The emergence of industrially relevant lignin monomer-oligomer separations methods based on membrane technologies could further advance the industrial viability of this approach ( 69 ). In addition, dimer cleavage has been reported in soil bacteria and engineered into P. putida for increased LRC conversion to product as a biological strategy to further use oligomeric lignin fractions and enhance conversion yields ( 11 , 70 ). Thermal and catalytic methods have been shown to enable rapid, extensive, and scaleable lignin deconstruction ( 6 ). Chemocatalytic oxidative depolymerization of lignin provides bioavailable monomers at higher yields than alkaline pretreatment methods ( 71 ). Notably, recent advances in oxidative chemistry have enabled the cleavage of lignin C─C bonds, not accessible by traditional catalytic lignin depolymerization approaches ( 72 ). Complete lignin conversion to bioavailable monomers would be a transformational advance."
} | 3,020 |
36882224 | PMC10045912 | pmc | 30 | {
"abstract": "Abstract Corals live in a complex, multipartite symbiosis with diverse microbes across kingdoms, some of which are implicated in vital functions, such as those related to resilience against climate change. However, knowledge gaps and technical challenges limit our understanding of the nature and functional significance of complex symbiotic relationships within corals. Here, we provide an overview of the complexity of the coral microbiome focusing on taxonomic diversity and functions of well-studied and cryptic microbes. Mining the coral literature indicate that while corals collectively harbour a third of all marine bacterial phyla, known bacterial symbionts and antagonists of corals represent a minute fraction of this diversity and that these taxa cluster into select genera, suggesting selective evolutionary mechanisms enabled these bacteria to gain a niche within the holobiont. Recent advances in coral microbiome research aimed at leveraging microbiome manipulation to increase coral’s fitness to help mitigate heat stress-related mortality are discussed. Then, insights into the potential mechanisms through which microbiota can communicate with and modify host responses are examined by describing known recognition patterns, potential microbially derived coral epigenome effector proteins and coral gene regulation. Finally, the power of omics tools used to study corals are highlighted with emphasis on an integrated host–microbiota multiomics framework to understand the underlying mechanisms during symbiosis and climate change-driven dysbiosis.",
"conclusion": "Conclusion Despite the recent development of molecular tools for understanding the diversity and function of coral holobionts, a mechanistic knowledge of the coral microbiome and its role in coral evolution and adaptation is still missing. Here, we charted an overview of the taxonomic diversity and function of microbiota associated with corals, the latest updates on coral microbiome research, and insights into the possible mechanisms through which coral–microbiota interactions could occur. We highlight that while most ‘omics techniques developed for corals have been powerful, integrated ‘coral–microbiota’ multiomics data are needed for holistic and systems-level understanding of the mechanisms underpinning the symbiotic and dysbiotic interactions within the coral holobiont. We also point the importance of the coral regulatory circuits and elements in responding to the microbiome and environmental change and how this knowledge can be used in an integrated multiomics framework. These advances, especially when combined with specific manipulative experiments and/or field samples and correlated with physiological status, promise to push the boundaries of knowledge of coral–microbiome research and may help global efforts to preserve corals in the future.",
"introduction": "Introduction Corals are metaorganisms that depend on dynamic multipartite symbioses with diverse microbes. These interkingdom interactions between the multicellular eukaryotic coral host and its associated microbiota maintain homeostasis within this complex system and has underpinned its resilience for >500 million years (Jaspers et al. 2019 , Robbins et al. 2019 , Peixoto et al. 2021 ). Associations within the metaorganism comprise a large diversity of viruses, prokaryotes, and microeukaryotes that collectively are termed the coral holobiont (Rohwer et al. 2002 , Rosenberg et al. 2007 , Rosenberg and Zilber-Rosenberg 2018 , Zilber-Rosenberg and Rosenberg 2021 ). Chief among the holobiont microbes, the primary endosymbiotic dinoflagellate of the family Symbiodiniaceae provides the bulk of the required nutritional needs to their coral hosts (Muscatine 1990 , Morris et al. 2019 ). In addition, an increasing body of evidence is unraveling the key roles particular bacterial species in specific and general prokaryotic communities play in maintaining holobiont fitness, potentially via exchanging essential metabolites, recycling nutrient, and providing protection against pathogenic microbes (Bourne et al. 2016 ). In the Anthropocene era, climate change disrupts these symbiotic relationships, leading to dysbiosis that is characterized by the overgrowth of opportunistic and putatively pathogenic microbes and results in a compromised coral immune system, inevitably causing the onset of coral bleaching and/or disease (van Oppen and Blackall 2019 ). Most coral microbiome work has been exclusively focused on either endosymbiotic algae or bacteria, while ignoring the other, largely underexplored members of the coral microbiomes due to difficulties associated with studying their role in the holobiont. This imbalance hinders our detailed understanding of the coral holobiont system. In this review, we provide the latest information on the taxonomic and functional diversity of members of the coral microbiome, focusing on (a) specific microbes that engage in beneficial or harmful interactions with their host, (b) the role these microbes presumably play in coral health or disease, (c) the potential mechanisms of coral–microbiome crosstalk and communication, and (d) new techniques and approaches to further our understanding of the coral holobiont. Our aim is to provide insights into the potential mechanisms through which coral–microbiome interactions occur, and how these mechanisms can be studied to unravel the governing principles of the coral holobiont ecology in a warming ocean. The coral holobiont Corals are a reservoir for microbes that includes diverse species of bacteria, archaea, viruses, and microeukaryotes (Bourne et al. 2016 ), some of which are well-characterized while others are cryptic (Fig. 1 ). The best-known coral symbionts are select bacteria and members of the dinoflagellate family Symbiodiniaceae (collectively known previously as the genus Symbiodinium ). Bacteria colonize all coral microhabitats including the surface mucous layer (SML), different tissue layers, and skeleton (Sweet et al. 2010 , Pollock et al. 2018 , van Oppen and Blackall 2019 ), while Symbiodiniaceae inhabit specific host-derived membrane structures called symbiosomes (Davy et al. 2012 , Mohamed et al. 2016 , Rosset et al. 2021 ) within the gastrodermis layer (Fig. 1 ). In addition, a plethora of under-explored microbes are also associated with corals, including newly discovered apicomplexan-related chromerids and corallicolids (Ainsworth et al. 2017 , Clerissi et al. 2018 , Kwong et al. 2019 ), endolithic algae, viruses, archaea, and fungi all with mostly unknown function (Fig. 1 ). Figure 1. The diverse microbiome of corals. The coral-associated microbiome is distributed across specific locations in a coral colony and is composed of diverse microbes spanning the three domains of life. Symbiodiniaceae and bacteria are among the most-studied coral symbionts (left). Symbiodiniaceae are localized within specialized coral structures called symbiosomes within the gastrodermis layer and are by far the best-studied symbiont of corals. Resident bacteria are found in most coral microhabitats, including the SML, coral tissue, and skeleton. Recent work on resident bacteria focuses on the importance of Endozoicomonas spp. as putative obligate coral symbionts. Most of the coral microbiome is considered cryptic, with mostly unknown roles in holobiont homeostasis (right). Some of these members can be endosymbiotic (e.g. corallicolids), while others appear to be epibionts (e.g. Chromerids). Evidence suggests some archaea may be involved in nitrogen cycling, viruses may be important in maintaining microbiome homeostasis, while endolithic communities ( Ostreobium and fungi) are implicated in primary production. Complex associations with the Symbiodiniaceae family Symbiodiniaceae were the first and most important symbionts of corals to be recognized (Muscatine and Porter 1977 , Bourne et al. 2016 , LaJeunesse et al. 2018 ). They live exclusively in a host-derived compartment known as the ‘symbiosome’ that originates from the early endosome of the coral host following phagocytosis of these algal symbionts (Fig. 1 ) (Davy et al. 2012 ). The symbiosome membrane protects Symbiodiniaceae cells from lysosome degradation by the host (Mohamed et al. 2016 ) and mediates the mutual transport of nutrients between both taxa (Davy et al. 2012 ). Considered initially as a single species, Symbiodinium microadriaticum (Freudenthal 1962 ), this group of symbionts now comprises the recently established dinoflagellate family Symbiodiniaceae, which currently includes seven distinct genera ( Symbiodinium —formerly known as Clade A, Breviolum —clade B, Cladocopium —clade C, Durusdinium —clade D, Effrenium —clade E, Fugacium —clade F, and Gerakladium —clade G) (LaJeunesse et al. 2018 ). Multiple reference genomes for Symbiodiniaceae are available (Shoguchi et al. 2013 , Lin et al. 2015 , Aranda et al. 2016 , Gonzalez-Pech et al. 2021 ), including a chromosome-scale genome for S. microadriaticum (Nand et al. 2021 ). These genomes reveal that this family is taxonomically and functionally divergent, a fact reflected in their diverse functional repertoire (González-Pech et al. 2021 ). The coral–Symbiodiniaceae symbiosis is likely highly complex as other endosymbiotic associations. Research indicates that metabolite exchange between Symbiodiniaceae and corals involve sugars, lipids, and nitrogen compounds (reviewed in Davy et al. 2012 ). However, these metabolites may vary in identity and importance among the various associations given the high divergence reported in the Symbiodiniaceae; despite this knowledge, the molecular mechanisms (pathways or molecules) that establish and maintain this interaction is unknown. Few insights into functions and pathways that could enable this symbiotic relationship were established using comparative genomics and transcriptomics (Aranda et al. 2016 , Liu et al. 2018 , Mohamed et al. 2020a ). For example, Aranda et al. ( 2016 ) showed that Symbiodiniaceae genomes possess an extensive repertoire of carbon and nitrogen transporters that likely underpin their symbiotic lifestyle and ultimately influence their hosts’ physiology. Comparative analysis of four Symbiodiniaceae draft genomes against other dinoflagellate genomes revealed identification of gene families under positive selection that included genes involved in photosynthesis, transmembrane ion transport, amino acid synthesis and transport, and stress responses (Liu et al. 2018 ). These functions may enable Symbiodiniaceae to be ideal partners to corals. In addition, these processes were shown to be activated during early interactions with coral larvae. Metatranscriptomics revealed upregulation of specific algal genes involved in carbohydrate, lipid, and nitrogen metabolism, and transport of various metabolites (glycerol, glutamate, choline) during colonization of coral larvae (Mohamed et al. 2020b ). More recently, simultaneous transcriptome, metabolome, and proteome data for three ecologically important Symbiodiniaceae isolates have become available (Camp et al. 2022 ). The availability of such large-scale omics data will inevitably increase our understanding of the molecular characteristics that underpin Symbiodiniaceae responses during their lifestyle changes and environmental stress. Diverse bacterial symbionts associated with corals Corals harbour a diverse bacterial microbiome (Blackall et al. 2015 ), spanning 39 phyla (Huggett and Apprill, 2019 ), more than one-third of the bacterial phyla found in seawater (Chen et al. 2021 ). A proportion of these coral-associated bacterial assemblages are thought to support the health and resilience of corals (Bourne et al. 2016 , Ziegler et al. 2019 , Voolstra and Ziegler 2020 , Meunier et al. 2021 ). Among the numerous bacterial phyla associated with corals, Proteobacteria, Bacteroidetes, Cyanobacteria, and Firmicutes are among the most abundant based on 16S rRNA gene phylogeny of 21 100 sequences derived from a public database (Huggett and Apprill, 2019 ). Moreover, a recent meta-analysis of 3055 bacterial isolates from 52 coral studies identified that most cultivable bacteria belonged to the Proteobacteria, Firmicutes, Bacteroidetes, and Actinobacteria phyla (Sweet et al. 2021 ). The collective reef microbiome may rapidly respond to environmental stressors, such as ocean warming, eventually leading to reef microbialization. Reef microbialization is characterized not only by a shift in abundance and biomass towards microbes, but more of a shift towards a pathogenic assemblage that can trigger major declines (Haas et al. 2016 ). Coral-associated bacteria inhabit several compartments within the coral, such as the SML, tissues, gastric cavity, and skeleton (Fig. 1 ) (Pollock et al. 2018 , van Oppen and Blackall 2019 , Vanwonterghem and Webster 2020 ). Distinct physiochemical properties and environmental gradients, including pollution (Wangpraseurt et al. 2016 , Pernice et al. 2020 ) play an important role in shaping the microbial composition within these compartments (Sweet et al. 2010 , Leite et al. 2018 , Pollock et al. 2018 ). Bacterial composition varies across these different niche compartments with some bacteria preferentially colonizing specific compartments. For example, bacteria belonging to the genera Chloroflexi, Sphingobacterium, Roseobacter , and Pseudoalteromonas were found exclusively in the SML (Sweet et al. 2010 ), while Endozoicomonas were found within aggregates inside coral tissues (Neave et al. 2017 ). This niche specificity suggests certain bacteria are adapted to the local microenvironment within the coral colony (Ritchie and Smith 2004 ), which ultimately leads to particular interactions with the host within each microenvironment. More diverse bacterial communities have been reported in the coral skeleton compared to those in the coral tissue or the SML (Pollock et al. 2018 ). Nitrogen cycling is common within corals [reviewed in Rädecker et al. ( 2015 )]. Diazotrophs are consistently associated with coral tissues (Rohwer et al. 2002 , Lema et al. 2012 , Olson and Lesser 2013 ), particularly in early life stages (larvae and juveniles) (Lema et al. 2014 ), indicating the potential importance of nitrogen fixation in the coral holobiont. Ammonium generated from nitrogen fixation may be partially oxidized by communities of ammonia oxidizing bacteria and archaea (Beman et al. 2008 , Siboni et al. 2008 , Yang et al. 2013 ). Likewise, denitrifying bacteria have also been reported in corals (Kimes et al. 2010 , Yang et al. 2013 ). More recently, Rädecker et al. ( 2022 ) have reported the tight relationship between disturbance in nitrogen cycling and coral bleaching. However, the molecular mechanisms by which the nitrogen-related activities of these microbial communities are coupled are largely unknown. Like all phytoplankton, Symbiodiniaceae associate with bacteria that play a role in their physiology and influence nutrient availability (Seymour et al. 2017 ). Members of the family Rhodobacteraceae are universally found among many phytoplankton lineages, including numerous Symbiodiniaceae cultures, and have been shown to play major roles in providing essential nutrients, hormones, and cofactors to phytoplankton (Cirri and Pohnert 2019 ). For example, Mameliella alba was shown to enhance the growth of Symbiodiniaceae in co-culture (Varasteh et al. 2020 ), similar to how other Rhodobacteraceae bacteria enhance the growth of diatoms (Amin et al. 2015 ) and coccolithophores (Segev et al. 2016 ). While Symbiodiniaceae–bacteria co-culture experiments are reductive, interactions between both taxa are hypothesized to occur within coral symbiosomes (Garrido et al. 2021 ). Indeed, nitrogen transfer between bacterial isolates labelled with 15 N and Symbiodiniaceae cells has been observed at the single-cell level in the coral Pocillopora damicornis (Ceh et al. 2013 ). Despite these findings, information on specific coral symbionts, opportunists, parasitic, and commensal bacteria and their importance to the coral holobiont is scarce. Cryptic diversity of the coral microbiome In addition to Symbiodiniaceae and bacteria, corals are home to a plethora of other microorganisms, including viruses, archaea, fungi (Bourne et al. 2016 , Ainsworth et al. 2017 , Clerissi et al. 2018 ), and microeukaryotes, including the apicomplexan-like Chromerids (Moore et al. 2008 , Janouškovec et al. 2012 ) and the recently discovered apicomplexans Corallicolids (Kwong et al. 2019 , Keeling et al. 2021 ). Below is a brief discussion of these mostly cryptic organisms. Apicomplexans and related organisms Early work has isolated and identified several lineages of apicomplexans associated with corals. Gemmocystis cylindrus was isolated from the gastrodermal cells of multiple corals (Upton and Peters 1986 ). Further molecular evidence of the existence of related apicomplexans included detection of DNA fragments in Caribbean corals (Toller et al. 2002 , Kirk et al. 2013 ). Analysis of plastid rRNA sequences derived from coral reef environments revealed eight distinct, novel apicomplexan-related lineages associated with corals (Janouškovec et al. 2012 ). Two of these lineages, Chromera velia (Moore et al. 2008 ) and Vitrella brassicaformis (Oborník et al. 2012 ), comprise photosynthetic alveolates of the phylum Chromerida that are commonly associated with corals worldwide, and are considered the closest known photosynthetic relatives of Apicomplexan parasites (Moore et al. 2008 , Janouškovec et al. 2013 ). More recently, a third apicomplexan taxon has been found ubiquitously associated with corals, potentially being the second most abundant microeukaryotic group in coral tissues after Symbiodiniaceae (Kwong et al. 2019 ). This taxon belongs to corallicolids (Kwong et al. 2019 ), a lineage that may be ubiquitous in the oceans including in metagenomes of deep-sea corals (Vohsen et al. 2020 ). Little is known about the biology of corallicolids or their influence on coral health/fitness, but it is unlikely that corallicolids have a mutualistic relationship with corals (Keeling et al. 2021 ). A recent transcriptomic study revealed that the coral host response to C. velia inoculation was similar to that of parasite or pathogen infection in vertebrates, suggesting that their relationship with corals is not beneficial (Mohamed et al. 2018 ). Further work is needed to employ inoculation experiments and subsequent time-series multiomics analyses to elucidate the nature of the coral–corallicolids association. Endolithic algae and fungi Endolithic algae form dense bands visible to the naked eye in the skeleton of many coral species and are often dominated by the filamentous green alga Ostreobium spp. (Chlorophyta) (Fig. 1 ) (Kornmann and Sahling 1980 ). Molecular studies revealed highly diverse communities within this group of green algae (Marcelino and Verbruggen 2016 , Del Campo et al. 2017 , Verbruggen et al. 2017 , Marcelino et al. 2018 ). More than 120 operational taxonomic units at the near-species level have been reported from 132 coral skeleton samples collected from multiple coral species (Marcelino and Verbruggen 2016 ). These endolithic communities were shown to substantially vary in identity among coral species. Marcelino et al. ( 2018 ) reported a more diverse endolithic community in the massive coral Porites spp. compared to the branching species Seriatopora hystrix and Pocillopora damicornis , suggesting that endolithic algae contribute to the resilience of the former to environmental stress. Ostreobium colonizes the skeleton of coral juveniles during their development (Masse et al. 2018 ) and can interact with the coral tissue through transfer of photosynthates (Schlichter et al. 1995 , Fine and Loya 2002 , Pernice et al. 2020 ), particularly after bleaching (Iha et al. 2021 ), and by enhancing coral recovery post-bleaching via reducing skeletal light reflectance (Galindo-Martinez et al. 2022 ). Fungi are known to be associated with many sessile marine invertebrates including corals and sponges (Yarden 2014 ). In corals, fungi are found in newly deposited coral skeleton along with Ostreobium (Le Campion-Alsumard et al. 1995 , Golubic et al. 2005 ), exhibiting rapid growth to match skeletal accretion (Le Campion-Alsumard et al. 1995 ). Fungi were identified as the most abundant microbes in the metagenome of Porites astreoides , contributing more than a third of the total microbial sequences (Wegley et al. 2007 ). Despite their ubiquitous associations with corals (Yarden 2014 ), including deep-sea corals (Marchese et al. 2021 ), their functions remain largely underexplored (Roik et al. 2022 ). Histological studies show widespread fungal invasion in corals infected with the coral disease white syndrome (Work and Aeby 2006 , Howells et al. 2020 ) [reviewed in Sexton and Howlett ( 2006 )]. In our recent work, reads belonging to the phylum Ascomycota in metagenomes of Acropora spp. under heat stress were observed, with higher relative abundance in colonies infected with white syndrome, potentially implicating them in disease manifestation (Amin et al., personal communication). Similarly, more diverse fungal communities were found in Acropora hyacinthus colonies living in warm pools compared to colder pools; these communities were also more transcriptionally active in warmer conditions (Amend et al. 2012 ), implicating them in responses to heat stress. Archaea Corals are associated with diverse archaeal species, mainly representatives from the phyla Thermoproteota (also known as Crenarchaeota) and Euryarchaeota. Members of the Thermoproteota are the most commonly reported followed by Marine Group II and Thermoplasma of the Euryarchaeota (Kellogg 2004 , Siboni et al. 2008 ). Archaea can comprise up to half of the prokaryotic community on the SML of some corals (Wegley et al. 2004 ). Despite their abundance, the functional roles of archaea within the coral holobiont have not been experimentally validated. However, they are often implicated in nitrogen recycling and ammonia oxidation within the SML (Siboni et al. 2008 , 2012 ). Two metagenomically assembled genomes (MAGs) that belong to the Nitrososphaerota (syn. Thaumarchaeota) phylum were assembled from metagenomic reads of the coral Porites lutea (Robbins et al. 2019 ) and revealed the presence of symbiosis-related metabolic pathways, including a reductive tricarboxylic acid cycle and cobalamin biosynthesis, suggesting these archaeal genomes might contribute essential vitamins or dissolved carbon to the host (Robbins et al. 2019 ). Viruses A wide range of coral species and associated microbes are reported to harbour virus-like particles (Wilson et al. 2004 , Marhaver et al. 2008 , Brüwer et al. 2017 ). Metagenomic sequencing show a high diversity of coral-associated DNA and RNA viruses (Weynberg et al. 2014 ). Large metagenomic and metatranscriptomic sequencing efforts towards establishing a ‘coral virome’ conducted across 101 cnidarian samples from the Red Sea documented DNA and RNA viral assemblages associated with corals (Cardenas et al. 2020 ) [for a recent review on the roles of viruses in corals see Ambalavanan et al. ( 2021 )]. While the functional roles of coral-associated viruses are still unclear, they likely play important roles in the coral holobiont. The presence of some bacteriophages in the coral SML may regulate the abundance of specific bacteria via targeted infection/lysis (Barr et al. 2013 ). Viral genes can encode for complementary functions that may be beneficial to the holobiont (Thurber et al. 2017 ). For example, some coral-associated viruses have genes related to photosynthesis that may alleviate and/or delay damage to Symbiodiniaceae photosystems at higher temperatures (Weynberg et al. 2017 ). In addition, Knowels et al. ( 2016 ) unexpectedly reported a decrease in viral abundance in reefs with high microbial abundance and suggested a lytic-to-lysogenic shift with increased microbial densities. This novel host-viral dynamic has been proposed as a mechanism of reef microbialization (Haas et al. 2016 ). However, viruses can also be detrimental to corals. Under various stress conditions, the coral-associated viral consortium exhibits an increase in herpes-like viruses, similar to other cnidarians (Thurber et al. 2008 ). Temperature-induced latent infection is also suggested to confer virulence to specific coral pathogens that could lead to the onset of coral disease (Weynberg et al. 2014 , Work et al. 2021 ). Functions of coral-associated microbes Coral–Symbiodiniaceae symbiosis, the engine of the holobiont The symbiotic relationship between corals and Symbiodiniaceae enabled the construction of the reef (calcium carbonate skeleton) via bidirectional nutrient exchange (Pogoreutz et al. 2020 ). The symbiosis relies on reciprocal metabolite exchanges, where Symbiodinaceae share excess photosynthetically derived dissolved organic matter with the coral host in exchange for access to inorganic nutrients and CO 2 generated from respiration (Muscatine 1990 , Falkowski et al. 1993 , Cunning et al. 2017 ). Indeed, the sharing of organic photosynthates by Symbiodiniaceae (e.g. glucose) is energetically sufficient for the host to meet 100% of its respiratory requirements (Muscatine and Porter 1977 , Bourne et al. 2016 ). Although corals are capable of assimilating ammonium to acquire nitrogen, Symbiodiniaceae are responsible for most inorganic nitrogen uptake in the forms of nitrate and ammonium (Pernice et al. 2012 ). A proportion of this nitrogen is shared with the coral host in the form of dissolved organic nitrogen (e.g. amino acids) (Wang and Douglas 1999 , Yellowlees et al. 2008 , Reynaud et al. 2009 ). However, high concentrations of inorganic nitrogen have been shown to destabilize the symbiosis. Increasing nitrogen fixation leads to an increase in nitrogen availability that subsequently increases cell division rates of the symbiont; this increase alters the N:P ratio within corals and causes phosphate limitation (Wiedenmann et al. 2013 ). Thus, corals control the growth of their symbionts by regulating access to inorganic nitrogen (Wooldridge 2013 ). Indeed, this nitrogen-budget balance is critical for the maintenance of the symbiotic relationship and further ‘fine-tuning’ of its outcome is evident from prokaryotic members of the coral microbiome (see below) (Cui et al. 2019 ). More recently, Rädecker et al. ( 2021 ) showed that coral bleaching can be correlated with disrupted nutrient cycling during heat stress, where the increased energetic demand of the host during heat stress leads to increased catabolism of amino acids, a more rapid release of ammonium concomitant with promotion of the growth of algal symbionts and retention of photosynthates. In addition to central metabolites, such as sugars and amino acids, Symbiodiniaceae produce mycosporine-like amino acids, pigments (e.g. fucoxanthin) and carotenoids, which collectively protect against UV radiation and reactive oxygen species (ROS) (Rosic and Dove 2011 , Rosic 2019 , Roach et al. 2021 ). Symbiodiniaceae-derived glucosides can serve as energy storage molecules, osmolytes, and antioxidants (Ochsenkühn et al. 2017 , Gegner et al. 2019 ), which may protect photosystem II from free radicals. Other metabolites such as glycerolipids, betaine lipids, and tocopherols that are produced by both host and symbiont are hypothesized to stabilize cellular membranes, assist protein renaturation, and act as antioxidants during heat stress (Hillyer et al. 2017b , Rosset et al. 2017 , Roach et al. 2021 ) and disease (Deutsch et al. 2021 ). Symbiodiniaceae possess the necessary genes to produce essential steroid precursors like squalenes and lanosterols that corals and other cnidarians either acquire through heterotrophic feeding or through their symbionts (Baumgarten et al. 2015 ). In addition to these exchanges, it is likely that the coral–Symbiodiniaceae relationship involves dozens to hundreds of metabolites that regulate their complex symbiosis, akin to most well-studied symbiotic systems, for which our knowledge is lacking. Further work is needed to shed light on these chemicals and their role in nurturing healthy corals. Beneficial coral-associated bacteria can boost coral health and resilience While the putative symbioses between the coral host and specific bacterial species is obscure when compared to Symbiodiniaceae, recent evidence suggests there are specific bacterial symbionts that benefit their coral host. Mining the literature, we assembled a list of bacteria that are hypothesized to be beneficial to corals based on experiments either in situ or in the laboratory ( Supplementary Table S1 ). Interestingly, thus far evidence shows that most bacteria that confer benefit to the holobiont belong mostly to the α - and γ -proteobacteria and to a lesser extent the Actinobacteria, Actinomycetia, Cytophagia, Flavobacteriia, Bacilli, and Oligoflexia classes ( Fig. 2 and Supplementary Table S1 ). The limited number of classes that have been found to be beneficial to corals relative to bacterial orders in seawater suggests there are selection mechanisms that enable corals to form beneficial interactions with such bacteria. Below is a discussion of some of these bacteria and the potential roles they play in benefitting the holobiont. Tissue-localized members of the coral microbiome, such as bacteria of the genus Endozoicomonas (Fig. 2 ), are hypothesized to be a core symbiont of corals as they are ubiquitously found across a wide range of coral species from diverse geographic locations (Neave et al. 2017 , Ziegler et al. 2017 , Pogoreutz et al. 2018 ), including deep-sea corals (Kellogg and Pratte 2021 ) and form highly stable association with such corals even during bleaching [reviewed in Hernandez-Agreda et al. ( 2016 , 2018 , 2019 )]. In addition, the relative abundance of Endozoicomonas is often strongly correlated to coral health (Bayer et al. 2013 , Roder et al. 2015 , Neave et al. 2016 ) with abundance generally high in healthy corals and lower in stressed, bleached, and diseased corals (Bourne et al. 2008 , Meyer et al. 2014 , Morrow et al. 2015 ). Based on these observations, it has been suggested that Endozoicomonas may be important for coral holobiont health, but its symbiotic exchanges with the holobiont have yet to be identified. Endozoicomonas harbour large numbers of genes involved in amino acid synthesis and carbohydrate cycling, prompting suggestions that it is involved in holobiont nutrient cycling (Neave et al. 2017 ). A recently published Endozoicomonas MAG from a Porites deep shotgun metagenome study contained genes essential for the biosynthesis of cobalamin, which is a vitamin required for methionine synthesis by both corals and Symbiodiniaceae (Robbins et al. 2019 ). This MAG also encoded the enzyme DMSO reductase that converts DMSO to DMS, providing a means to recycle dissolved organic sulphur (Robbins et al. 2019 ). More recently, the genome of Endozoicomonas acroporae was shown to encode a DMSP CoA-transferase/lyase gene (dddD), capable of metabolizing DMSP into DMS (Tandon et al. 2020 ). DMSP metabolism may play a role in structuring the holobiont microbial community (Raina et al. 2013 ). Finally in response to coral tissue extract additions, E. marisrubri was shown to differentially express genes putatively involved in symbiosis establishment, e.g. flagellar assembly, ankyrins, ephrins, and serpins. Proteins involved in vitamin B 1 and B 6 biosynthesis were also upregulated (Pogoreutz et al. 2022 ). Figure 2. Relationship and diversity of bacterial species associated with corals based on previous research. Coral microbiome data were retrieved from SILVA (Quast et al. 2013 ) based on a literature search and mapped onto the prokaryotic tree of life (Hug et al. 2016 ). Additional sequences were merged into the original alignment by Hug et al. ( 2016 ) using MAFFT (Katoh et al. 2002 ). A maximum likelihood tree was constructed using FastTree (Price et al. 2010 ). In cases where 16S rRNA sequences were not available in the database, the branches were substituted with their closest available neighbour. Only taxa with some experimental evidence as to their relationship to corals were included. Beneficial, antagonistic (opportunists and pathogens), and Symbiodiniaceae-associated bacterial species (highlighted in the outer ring) are clustered in different parts of the tree. The tips of the branches are colour-coded according to the taxonomic classification. Generally, clusters of closely related bacterial species have similar relationships within coral microbiomes. For example, Symbiodiniaceae-associated bacteria mostly belong to the family Rhodobacteraceae, species belonging to the genus Vibrio are mostly antagonistic, while species belonging to the genera Alteromonas and Photobacterium are beneficial. Interestingly, members of the same genus do not always have the same relationship. For example, members of Pseudoalteromonas can be beneficial or antagonistic. A full list of bacteria depicted here and their references are provided in Supplementary Table S1 . Many studies have examined the ability of natural bacteria in coral microbiomes to inhibit or prey on coral pathogens. Several strains of Ruegeria spp. (Fig. 2 ) were found to inhibit the growth of the coral pathogen Vibrio coralliilyticus and other Vibrio spp. (Miura et al. 2019 ). Ruegeria spp. have also been implicated as indicator species in healthy coral microbiomes (Rosado et al. 2019 ). Reef-building corals were challenged with V. coralliilyticus in the presence or absence of the Vibrio predator Halobacteriovorax sp. PA1 (Fig. 2 ), which is commonly found at low abundance on coral surfaces (Welsh et al. 2016 , Zaneveld et al. 2016 ). Inoculation of corals with V. coralliilyticus induced major changes in the microbiome, especially a large increase in relative abundance of Vibrio spp., a reduced microbiome stability and proliferation of opportunists, such as Rhodobacterales and Cytophagales. In contrast, co-inoculation of the corals with both bacteria eliminated the increase in Vibrio spp. and prevented the proliferation of opportunists (Welsh et al. 2017 ). Pseudovibrio sp. P12 (Fig. 2 ) was shown to produce the antimicrobial metabolite tropodithietic acid, potentially through metabolizing coral DMSP, to inhibit the growth of V. coralliilyticus and V. owensii (Raina et al. 2016 ). More recently, the beneficial role of degrading excess DMSP during heat stress has been validated (Santoro et al. 2021 ). Enrichment of a DMSP degrading bacterium is associated with a significant increase of DMSP degradation and a concomitant coral holobiont physiological improvement, resulting in higher survival rates. Scavenging of ROS is a common mechanism by which bacteria can benefit a eukaryotic host, and has also been recently suggested as a beneficial mechanism for corals (Peixoto et al. 2017 ). In corals, several bacterial species were shown to detoxify radicals and ROS mainly by producing ROS-reactive pigments. These include Fabibacter pacificus, Paracoccus marcusii , and Pseudoalteromonas shioyasakiensis (Fig. 2 ) (Varasteth et al. 2021 ). Six strains of bacteria belonging to Alteromonas macelodii, A. oceani, Roseibium aggregata, Marinobacter salsuginis, Micrococcus luteus , and M. yunnanensis (Fig. 2 ) were shown to remove oxygen radicals from the coral model Exaiptasia diaphana (Dungan et al 2021 ). SML-assoicated bacteria (Fig. 1 ) can be an important source of antibiotics that fend off pathogen colonization (Ritchie 2006 , Engelen et al. 2018 ) and isolates of Pseudaltermonas strains from O. patagonica SML were active against the coral pathogens V. shiloi, V. coralliilyticus , and Thalassomonas loyana (Shnit-Orland et al. 2012 ). Symbiodiniaceae–bacterial interactions, a forgotten partnership in a complex symbiotic network Much of our understanding of the symbiotic relationships within the holobiont stems from host interaction with either Symbiodiniaceae or cultivated bacteria; in contrast, little is known about Symbiodiniaceae–bacteria interactions. Recent work has revealed complex metabolite exchanges between phytoplankton and their associated microbiome occurring within a microscale diffusive boundary layer surrounding phytoplankton cells, known as the phycosphere (Amin et al. 2012 , Seymour et al. 2017 ). Symbiodiniaceae strains in culture harbour different bacterial communities in the phycosphere with abundances exceeding those of the algal cells by almost two orders of magnitude (Ritchie 2012 , Lawson et al. 2018 ). Among these diverse communities, members of the genera Marinobacter, Roseibium (formerly Labrenzia ), Muricauda, Hyphomicrobium, Methylobacterium , and members of the families Rhodobacteraceae/Roseobacteraceae (e.g. Ruegeria, Mameliella ) have been consistently detected in Symbiodiniaceae cultures (Ritchie 2012 , Lawson et al. 2018 , Camp et al. 2020 , Varasteh et al. 2020 , Maire et al. 2021 ) [reviewed in Matthews et al. ( 2020 )]. Some of these taxa have been shown to be symbiotic with different phytoplankton lineages. For example, Marinobacter spp. provide a bioavailable source of iron to dinoflagellates and some diatom species in iron-limited environments (Amin et al. 2009 ). In Roseobacteraceae, Ruegeria pomeroyi has been shown to provide vitamins to diatoms in exchange for organic sulphur compounds (Durham et al. 2015 ). Sulfitobacter pseudonitzschiae and Phaeobacter inhibins convert the diatom- and coccolithophore-secreted amino acid tryptophan to the hormone indole acetate, respectively, which enhances the algal cell division rate (Amin et al. 2015 , Segev et al. 2016 ). Sulfitobacter pseudonitzschiae and Phaeobacter spp. have also been shown to successfully colonize the phycosphere of diatoms (Fei et al. 2020 ) by efficiently responding to host secondary metabolites (Shibl et al. 2020 ). Among the Roseobacteraceae, Mameliella alba has been consistently isolated from dinoflagellate cultures (Li et al. 2019 , Varasteh et al. 2020 , Lin et al. 2021 , Ren et al. 2022 ) and appears to enhance the growth rate of the dinoflagellates Symbiodinium sp., Alexandrium catanella , and Karenia brevis (Varasteh et al. 2020 , Lin et al. 2021 , Ren et al. 2022 , Amin et al., personal communication), suggesting they produce a growth-promoting hormone. Axenic Symbiodiniaceae cultures originating from the coral Galaxea fascicularis have been shown to exhibit a decrease in photosystem II maximum quantum yield and an increased production of ROS. A Muricauda sp. (Fig. 2 ) isolated from xenic Symbiodiniaceae was subsequently shown to protect Symbiodiniaceae Photosystem II from ROS via production of the antioxidant Zeaxanthin (Motone et al. 2020 ). Despite these advances, no information is currently available on microbial communities associated directly with the Symbiodiniaceae phycosphere within coral symbiosomes [for a recent review see Garrido et al. ( 2021 )]. Matthews et al. ( 2020 ) hypothesized that Symbiodiniaceae-associated bacterial consortia regulate Symbiodiniaceae productivity and thus the symbiotic interactions with corals. Future research should exploit recent advances in microfluidics, single-cell sequencing, and metabolomics to uncover the metabolic interaction between Symbiodiniaceae and bacteria within the coral host that are likely central to the holobiont fitness. Microbial dysbiosis is triggered by environmental stress and could drive the onset of coral disease During environmental stress, many putative opportunistic and pathogenic taxa increase in abundance, such as members of the Vibrionaceae, Roseobacteraceae, and Rhodobacteraceae, due to the immune-compromised state of the host (Cardenas et al. 2012 , Ziegler et al. 2016 , Pollock et al. 2017 , Certner and Vollmer 2018 ). It is noteworthy to point out that generalizations about genera or families being beneficial or harmful to corals should be avoided as interspecies interactions rely on a highly specific sets of genes that enable a bacterium to behave one way or another. As pointed out below, some families of bacteria contain species that are both beneficial and harmful to corals. Across coral species, during dysbiosis, specific bacteria have been shown to increase in abundance and activity; concomitantly, larger changes in the coral microbiome that involves one or more groups of bacteria were also shown to change in abundance and some of these have direct repercussions on coral physiology. For example, during coral exposure to elevated temperatures, diazotrophic bacteria were shown to increase in abundance (Santos et al. 2014 , Lesser et al. 2018 , Mohamed, personal communication) [for reviews see Radecker et al. ( 2015 ), Benavides et al. ( 2017 )], which has a direct effect on nitrogen availability. More recent data show that despite an increase in nitrogen fixation that is correlated with an increase in diazotrophs during heat stress, fixed nitrogen is not assimilated by either the coral tissue or the algal symbionts (Rädecker et al. 2022 ). Below examples of specific parasitic and opportunistic bacteria that have been reported are discussed. Forty coral diseases have thus far been described (Sweet et al. 2012 , Bruckner 2015 ); however, only few coral pathogens have been described (Pollock et al. 2011 ) [for a list of putative causative agents of coral disease, see Mohamed and Sweet ( 2019 )]. Among proposed coral pathogens, V. coralliilyticus (Fig. 2 ) is the most well-characterized with direct implication in the onset of both coral bleaching and the infectious disease white syndrome (Ben-Haim et al. 2003 , Pollock et al. 2011 , Ushijima et al. 2014 ). Several other Vibrio species, such as V. harveyi, V. owensii , and V. alginolyticus (Fig. 2 ), have also been implicated in white syndrome (Luna et al. 2010 , Ushijima et al. 2012 , Zhenyu et al. 2013 ) and V. tubiashii (Fig. 2 ) in white patch syndrome (Sere et al. 2015 ). Pseudoalteromonas piratica (Fig. 2 ) has also been implicated in white syndrome (Beurmann et al. 2017 ). Other bacteria have been proposed as causative agents of coral diseases. For example, white pox in A. palmata has been proposed to be caused by the enteric bacterium Serratia marcescens (Fig. 2 ) (Patterson et al. 2002 ). White plague type II in scleractinian corals has been proposed to be caused by Aurantimonas coralicida (Fig. 2 ), a relative of Rhizobiales (Denner et al. 2003 ), while T. loyana (Fig. 2 ) was proposed to cause a white plague-like disease (Thomposon et al. 2006 ). Black band disease was originally thought to be caused by the cyanobacteria Pseudoscillatoria coralii (Rasoulouniriana et al. 2009 ) and Roseofilum reptotaenium (Fig. 2 ) (Casamattaet et al. 2012 ); however, further research described this disease as a lesion of a complex microbial consortium composed of cyanobacteria and other microbes, including the sulphate-reducing bacterium Desulfovibrio sp. (Fig. 2 ), and a diverse array of heterotrophic bacteria, archaea, fungi, and other microeukaryotes (Sato et al. 2016 ). Finally, Candidatus Aquarickettsia rohweri (Fig. 2 ) is suspected of being implicated in white syndrome type I. This putative parasitic bacterium possesses several tools to benefit from the coral host, including an antiporter to exchange host ATP for ADP, a type IV secretion system, and appears to be using host nutrients, particularly nitrogen (Klinges et al. 2020 ). Other causative agents and almost all molecular factors of disease remain obscure. More recently, during the onset of grey-patch disease, a ‘microbiome-to-pathobiome’ shift occurs that favours multiple specific pathogens that may be involved in degrading coral tissues (Sweet et al. 2019 ). This shift is hypothesized to be caused by bacterial quorum sensing molecules, such as homoserine lactones (Certner and Vollmer 2015 ). Homoserine lactones are small molecules produced by many bacteria to regulate their gene expression based on population density. Genes related to pathogenesis in bacteria, e.g. biofilm formation, siderophore production, toxin secretion, are typically regulated by quorum sensing (De Kievit and Iglewski 2000 , Winzer and Williams 2001 , Visca et al. 2007 ). Disease symptoms were induced in healthy A. cervicornis colonies exposed to bacteria supplemented with exogenous homoserine lactones, which correlated with a ‘healthy’ to ‘disease-causing’ microbiome switch and leading to white band disease-like symptoms. Indeed, microbial consortia isolated from white band disease-infected colonies and treated with homoserine lactone inhibitors lost their ability to develop the disease (Crenter and Vollmer 2018 ). These observations suggest that quorum sensing can modulate bacterial regulatory networks that then reshape the microbial community during disease onset, though the mechanism of how this occurs is still unclear. Leveraging the coral microbiome to boost resilience of the holobiont Inoculation of corals with probiotic microbes has been proposed to protect corals from the harmful impact of oil spills. This bioremediation approach was successful in mitigating the impacts of pollution and improved the health of affected corals (Fragoso Ados Santos et al. 2015 ). Several approaches have been proposed to aid corals in increasing their fitness, such as experimental evolution in coral photosymbionts (van Oppen et al. 2015 , 2017 ) and bacterial probiotics application (Peixoto et al. 2017 ). Introducing heat-tolerant Symbiodiniaceae into corals It is widely accepted that coral thermal tolerance is largely dependent on the physiology of their associated Symbiodiniaceae partners (Berkelmans and van Oppen 2006 ). In-vitro exposure of Symbiodiniaceae cultures to elevated temperatures increases their thermal tolerance after ∼40 generations (Chakravarti et al. 2017 , Chakravarti and van Oppen 2018 ). Despite this acclimation, reintroducing heat-tolerant strains into corals yielded no significant benefit for the holobiont (Chakravarti et al. 2017 ). In contrast, a small minority of heat-tolerant Symbiodiniaceae strains derived from the same wild-type clone increased the thermal tolerance of coral larvae (Buerger et al. 2020 ). A mechanistic understanding of how heat tolerance in Symbiodiniaceae occurs and how in turn it influences the coral holobiont is needed to improve the efficacy of this approach. Using probiotics to help increase corals’ resilience to climate change Bacterial symbionts of corals represent an opportunity to increase the resilience of the coral in response to ocean warming (Zielger et al. 2019 , Voolstra and Zielger 2020 , Voolstra et al. 2021a ). Coral microbiomes, especially those inhabiting the coral SML, are thought to rapidly respond to the surrounding environment and may contribute to the resilience and health of the holobiont (Bang et al. 2018 , Ziegler et al. 2017 , 2019 ). Recent efforts have been focusing on coral ‘probiotics’ applications to boost adaptation to climate change (Peixoto et al. 2017 ). This approach involves isolation and screening of native bacterial associates of corals for functions beneficial to coral health, and subsequently carry out physiological assays to determine holobiont performance after inoculation with these putatively beneficial microorganisms for corals (BMCs) (Rosado et al. 2019 ). Experimental manipulation using mixed consortia of native coral bacterial isolates harbouring beneficial genes such as nitrogen fixation ( nifH ) and DMSP-degradation ( dmdA ) genes (Fig. 3 ) resulted in partial mitigation of coral bleaching compared to controls or corals challenged with the pathogen V. coralliilyticus (Rosado et al. 2019 ). Mesocosm experiments coupled with multiomics revealed an increase in coral resilience following probiotic application and was followed by a reprogramming of coral transcriptional machinery to activate the immune system and stress pathways during the recovery period (Santoro et al. 2021 ). Despite the successful applications described so far, most marine bacteria remain uncultivable (Lok 2015 , Hofer 2018 , Jiao et al. 2021 ) and even with cultured ones, our knowledge of their benefit to the coral is limited. As our understanding of the role bacterial symbionts play expands, e.g. through the use of culturomics (Schultz et al. 2022 ), more targeted engineering of beneficial microbial communities may be valuable in supporting recovery of coral reefs. Figure 3. Depiction of the influence of BMCs to reduce mortality after heat stress (left) and pathogenic bacteria of the genus Vibrio on the coral host (right). Coral cells are depicted as the eukaryotic cell at the bottom of each side of the figure, while the microbiome or Vibrio are depicted by bacterial cells on the top. Function names in red font indicate an increase, while those in blue font indicate a decrease. After being exposed to heat stress and inoculated with BMCs, the coral host increases sterol biosynthesis while decreasing apoptosis and inflammation. BMCs support the coral host via N 2 fixation, DMSP–DMSO degradation, ROS scavenging, proteins related to B-complex vitamins, nitric oxide detoxification, and production of antibiotics. Vibrio competes with members of the coral microbiome via production of type VI secretion system (T6SS) proteins, toxins, and siderophores. These interspecies competition mechanisms increase virulence factors that may confer an advantage to these pathogens over resident taxa and induce changes in the coral holobiont. Corals exposed to vibrios show high levels of platelet-activating factors (such as Lyso-PAFs) as a defence mechanism as they have antimicrobial properties. DMSP; dimethylsulfoniopropionate, DMSO; dimethyl sulfoxide, Lyso-PAFs; platelet-activating factors, and AMPs; antimicrobial peptides. Another approach inspired by fecal microbiota transplantation in humans, called field-based coral microbiome transplantation (CMT) (Doering et al. 2021 ), has successfully shown the feasibility of microbiome transplantation of homogenized coral tissues from healthy colonies to bleached colonies to increase coral heat tolerance. This approach has many advantages as it circumvents ethical issues associated with introducing new bacterial isolates into the environment, avoids the daunting task of screening bacterial function in the laboratory, and enables the transmission of the large uncultivable fraction of the microbiome. In both probiotic and CMT approaches, the mechanisms underlying the microbiome–host interaction and stress tolerance are yet to be established. Understanding host–pathogen interactions in the coral- Vibrio system It is now widely accepted that many coral diseases are caused by a diverse polymicrobial consortium [Roder et al. 2014 , Sweet et al. 2019 , for a review see (Mohamed and Sweet 2019 )] though the mechanisms underlying these infections are largely unknown. During responses to environmental stressors, the microbiome of the reef undergoes shifts towards an increasing microbial diversity (Haas et al. 2016 ). During this reef microbialization, the coral decline is attributed to the preferential increased abundance of pathogens and their virulence factors. Although most infection experiments examine coral pathogens separately, simultaneous inoculation of V. coralliilyticus and V. mediterranei in the coral Oculina patagonica leads to increased virulence and higher coral tissue damage, suggesting the cumulative effect of both bacteria accelerate pathogenicity (Rubio-Portillo et al. 2014 ). Recently, Rubio-Portillo et al. ( 2020 ) attempted to understand the mechanisms underpinning the interaction between these pathogens and their interaction with corals during infection. When co-cultured together, these bacteria overexpress genes related to virulence factors, such as siderophores, type VI secretion system, and toxins (Fig. 3 ). These transcriptional responses towards a related competing species suggest these pathogens may favour the colonization of the host when they are present in a mixed population. Moreover, during coral exposure to a coculture of V. coralliilyticus and V. mediterranei , virulence factors (product of interspecies competition between the two coral pathogens) led to shifts in the coral microbiome favouring specific opportunistic groups. These in turn caused increased production of Lyso-PAFs (Fig. 3 ) by the coral to fight the pathogens back, which led to increased production of ROS and tissue necrosis (Rubio-Portillo et al. 2020 ). Coral–microbiome crosstalk from recognition to gene regulation Little is known about the mechanisms of coral–microbiome interactions mainly because of the lack of a genetically tractable coral model system that can be manipulated in the laboratory, the lack of cultivable strains of certain coral microbial symbionts (e.g. some bacteria, certain Symbiodiniaceae strains, and cryptic species) and their genomic resources. All these aspects hinder our understanding of the gene regulatory circuits within members of the coral holobiont. Other symbiotic systems, such as human–microbiome interactions, provide an opportunity to learn more about coral holobiont interactions. In this section, identified recognition mechanisms that corals use to interact with its microbiome, the potential for the coral microbiome to produce epigenome-effector proteins, delivery of such microbially derived signals to the coral host, and mechanisms that enable coral–microbiome interactions with a focus on noncoding RNAs (ncRNAs), akin to human–microbiome interactions, are discussed. Recognition mechanisms in coral–microbiome interactions Strong evidence supports the key roles of the host innate immune system in all aspects of the symbiotic association, from recognition, maintenance, and collapse (dysbiosis) (Weis 2008 ). Recognition is the first step that enables corals to determine whether a microbe is beneficial or not. Microbial cell membranes (cell walls in the case of Symbiodiniaceae) are decorated with a variety of microbial-associated molecular patterns (MAMPs), including glycans, that are recognized by the coral host via the pattern recognition receptors (PRRs) on phagocyte cell surfaces (Weis 2008 ). A wide variety of PRRs has been recognized in cnidarians, including toll-like receptors (TLRs), the intracellular pattern recognition receptor nucleotide-binding oligomerization domain 2 (NOD2), complement and its receptor (CRs), scavenger receptors (SRs), and lectins (Fig. 4 ) (Weis 2008 , Davy et al. 2012 , Weis 2019 ). These PRRs can be activated to detect beneficial microbes, e.g. the detection of Symbiodiniaceae by lectins (Mohamed et al. 2016 ), while being suppressed during encounters with parasites presumably as a host evasion mechanism (Mohamed et al. 2018 ). These MAMP–PRR interactions in corals include: lectin–glycan interactions (Wood-Charlson et al. 2006 , Bay et al. 2011 , Parkinson et al. 2018 ), scavenger receptors (Neubauer et al. 2016 ), thrombospondin type 1 repeat proteins (Wolfowicz et al. 2016 , Neubauer et al. 2017 , Mohamed et al. 2020b ), glycoprotein2 (GP2) (Mohamed et al. 2016 , 2020b ), toll-like/nucleotide oligomerization domain (NOD)-like receptors (TLRs/NLRs) (Hamada et al. 2013 , Weiss et al. 2013 , Mohamed et al. 2020b ), and complement systems (Poole et al. 2016 ) (Fig. 4 ). Figure 4. Potential mechanisms underpinning host–microbiota crosstalk in corals. Coral-associated microbes can alter the host gene expression by modifying the host epigenome. The coral host recognizes microbes (lectin–glycan interaction) and their products through interaction with extracellular receptors or through exosome-based delivery of microbiome-derived molecules. (A) These microbial-derived signals may cause specific changes in the host nucleus directly via epigenome-modifying proteins or indirectly through NOD2 signalling, following binding of various pattern recognition receptors to microbial signals (e.g. metabolites). (B) This leads to changes in host gene expression through epigenetic mechanisms, such as changing chromatin accessibility and DNA methylation that will lead to differential transcription factors binding, altered expression of certain genes, transcription factors, and ncRNAs such as long noncoding RNAs (lncRNAs). LncRNAs can further regulate gene activity at the epigenetic, transcriptional, or post-transcriptional levels. (C). The interaction is bidirectional as host-derived signals (proteins, ncRNAs, or metabolites) may be delivered to the coral microbiome (D) leading to differential microbial growth. NLR: NOD-like receptor, lncRNP: lncRNA-protein complex, and miRNA: microRNA. Once PRRs are activated in the case of Symbiodinaceae, the host facilitates the persistence and proliferation of symbionts inside symbiosomes via suppression of its immune response and arrest of phagosomal maturation (Davy et al. 2012 , Mohamed et al. 2016 , 2020b ). In contrast, immune response and phagosomal maturation are activated to reject and clear the symbionts out during dysbiosis and bleaching (Downs et al. 2009 ) and during encounters with parasites (Mohamed et al. 2018 ). Activation of MAMP–PRR interactions leads to downstream innate immune signalling cascades and production of effector proteins (Fig. 4 ) such as tolerogenic TGFβ pathway (Detournay et al. 2012 , Berthelier et al. 2017 ), sphingolipid signalling (Kitchen and Weis 2017 , Kitchen et al. 2017 ), and the master immunity regulator NFκB (Mansfield et al. 2017 , Jacobovitz et al. 2021 ). These interactions are well supported in corals and sea anemone by many ‘omics studies that implicate innate immune genes (Shinzato et al. 2011 , Mohamed et al. 2016 , 2020b , Cunning et al. 2017 ,, Jacobovitz et al. 2021 ). Microbially derived host epigenome-effector proteins Many bacteria, including the pathogens Helicobacter pylori and V. cholerae deliver effector proteins into a wide range of host cells, including humans, plants, and invertebrates using type IV or VI secretion systems to interfere with host signalling pathways (Green and Mecsas 2016 ). Not surprisingly, secretion systems have been highlighted as putative mediators of symbiotic associations (Coombes 2009 ). Interestingly, several secretion systems, including type IV, have been identified in the microbiomes of several coral species (Weber et al. 2019 ). These secretion systems can be used to deliver effector proteins from the associated microbiomes to the host or other members of the holobiont. Moreover, living cells can send and receive packages of information that are enclosed by cell membranes in the form of extracellular vehicles (EVs) such as exosomes (Fig. 4 ). EVs are lipid bilayer nanoparticles that act as key messengers in cell-to-cell communication and can be produced by unicellular microbes and multicellular metazoans alike. EVs contain diverse molecules, including effector proteins, such as microbes-derived epigenetic-modifying proteins (Yang et al. 2022 ). Once delivered to the host, these epigenetic-modifying proteins target the host cell nucleus to affect host responses through epigenetic mechanisms. These epigenetic modulations are hypothesized to directly or indirectly influence phenotypic responses in the host (Barno et al. 2021 , Morovic and Budinoff 2021 ). Barno et al. searched for putative homologs of known epigenome-modifying proteins from other host model systems in 18 bacterial genomes and 52 prokaryotic MAGs associated with two coral species. They identified homologs of the histone modification proteins ankyrin-repeat protein A and internalin B, a histone methyltransferase, and several DNA methyltransferases, suggesting that the coral microbiome has the machinery to modify the host epigenome (Barno et al. 2021 ). Molecular mechanisms of host–microbiota crosstalk in corals The mechanisms enabling the microbiome to influence genetic and physiological responses of the coral host are lacking. Many studies in model organisms, including humans, show an association between the microbiome and host gene expression. However, it is unclear what the direction of causality is with these associations. Disentangling this relationship is crucial for understanding homeostasis of normal symbiosis and dysbiosis, leading to disease etiology (Nichols and Davenport 2021 ). Upon delivery of microbiome-derived signals, many signalling cascades are likely activated to influence the host epigenome, which ultimately reprograms the host transcriptional machinery towards specific host phenotypes (Fig. 4 ). Transcription factor-mediated gene regulation Gene regulation is usually mediated through transcription factors (TFs) that can link host gene expression and its microbiome (Fig. 4 ). Host TFs bind to specific DNA motifs (regulatory elements such as promoters and enhancers) to control the transcription of certain genes. Previous research on zebrafish demonstrated interactions between the TF HNF4A and the microbiome promote gene expression patterns associated with inflammatory diseases (Davison et al. 2017 ). In mice, microbiota colonization of intestinal epithelial cells leads to drastic genome-wide reduction of the HNF4A occupancy, a measure of TF binding to its DNA motif (Davison et al. 2017 ), suggesting that the microbiota negatively regulate HNF4A. This indicates a conserved role for HNF4A in maintaining homeostasis of the intestine in response to the microbiome. Similarly, in Metazoa, the master regulator of innate immunity, NF-κB (Gilmore and Wolenski 2012 ) has been implicated during both the onset and breakdown of the coral–Symbiodiniaceae symbiosis. NF-κB activation leads to the upregulation of various effector pathways that drive an innate immune response. Numerous MAMP–PRR interactions that trigger the activation of NF-κB have been a recent focus in cnidarian genomic studies (Poole and Weis 2014 , Baumgarten et al. 2015 , Williams et al. 2018 ). The presence of Symbiodiniaceae in Aiptasia triggers a strong suppression of the host immune response (Perez and Weis 2006 , Detournay and Weis 2011 ), but the exact mechanism is still unclear. Inoculation experiments in Aiptasia suggest that NF-κB is playing a role in this immune suppression, as the addition of symbionts leads to decreases in NF-κB expression in aposymbiotic larvae inoculated with mutualistic Symbiodiniaceae (Wolfowicz et al. 2016 , Mansfield et al. 2017 ). Inversely, NF-κB expression increases during bleaching in adults (Mansfield et al. 2017 ). These results suggest that during symbiosis establishment, the algal symbionts modulate the host immune response by repressing the expression of NF-κB to enable colonization of the host. However, the link between NF-κB and coral-associated bacteria has not been established yet. Epigenetic modifications Other mechanisms of gene regulation include epigenetic modifications that can influence gene activity, including DNA methylation (addition of methyl groups at specific genomic CpG loci) and histone acetylation (addition of acetyl groups at specific histone sites). Epigenetics is a rapidly growing field and of great interest in the context of ‘environmental memory’ (Eirin-Lopez and Putnam 2019 ) that may explain phenotypic plasticity and acclimatization (Torda et al. 2017 , Liew et al. 2018 ). The host epigenetic profiles are thought to be influenced by its microbiome (Yu et al. 2015 , Krautkramer et al. 2017 , Miro-Blanch and Yanes 2019 ). In mice, germ-free animals have lower DNA methylation levels across the genome in the colon cells compared to animals with microbiomes (Yu et al. 2015 ). Fecal transplants also increase global DNA methylation in germ-free mice (Krautkramer et al. 2017 ). The microbiome can additionally remodel host responses at the chromatin level (Fig. 4 ) in intestinal epithelial cells. Profiling chromatin states via ATAC-seq in human colonic epithelial cell culture demonstrated that specific microbes regulate genome-wide accessibility of chromatin and TF binding in the host tissues (Richards et al. 2019 ). In mice, the presence of the microbiome results in increased levels of chromatin accessibility in intestinal epithelial cells compared to a gnotobiotic mouse (Semenkovich et al. 2016 ). These microbiome-induced changes to host epigenomes play a major role in various aspects of health and metabolic disease, including responding to diet, inflammation, obesity, and diabetes [for a review see Sharma et al. ( 2020 )]. However, this microbiome–host epigenome relationship has not been examined in corals yet. Research should be focusing on establishing the link between the coral-associated microbes and the host epigenetic profiles. While changes in the coral epigenome or microbiome is mostly correlated to environmental stressors (Putnam and Gates 2015 , Roder et al. 2015 , Dixon et al. 2018 , Liew et al. 2018 , Ziegler et al. 2019 , Tong et al. 2020 ), similar interaction patterns between the microbiome and the coral epigenome are expected, despite the current lack of such information. Interestingly, a recent shotgun metagenomic study has hypothesized a possible chromatin interaction between corals and their associated microbiomes. MAGs derived from healthy colonies of the coral P. lutea were shown to be enriched in genes encoding ankyrin repeat proteins (Robbins et al. 2019 ). One such protein, AnkA, presumably disrupts host antimicrobial responses against its producer, the intracellular pathogenic bacterium Anaplasma phagocytophilum (Garcia-Garcia et al. 2009 ). AnkA is hypothesized to translocate into the host cell nucleus and bind to regulatory regions of the host chromatin, silencing key host defence genes involved in ROS production. These patterns suggest intracellular pathogens may directly regulate host gene expression by changing host chromatin structure. Similarly, pathogens like V. coralliilyticus may be able to modify host responses using similar mechanisms. Intracellular mutualistic bacteria may also use these mechanisms to suppress the host immune response to establish a symbiotic relationship. ncRNAs as mediators of immune priming and potential communication signals RNA molecules constitute a common ancient language encoded in all living organisms across all domains of life. Genome-wide transcription in all living cells is responsible for producing two main classes of RNA molecules: coding RNAs (<2%) and ncRNAs comprised of RNA molecules with reduced coding potential that have instead regulatory functions (Hangauer et al. 2013 ). Among the different classes of ncRNAs, lncRNAs share structural similarities with mRNAs such as polyadenylation at the 3′ ends and the cap structure at the 5′ ends [for a recent review see Statello et al. ( 2021 )]. Mechanisms of lncRNA-mediated gene regulation can happen at the epigenetic (by recruiting or sequestering epigenetic modifiers), transcriptional (forming triple helical-structures with DNA via Hoogsteen base paring), and post-transcriptional (forming duplexes such as lncRNA–mRNA or lncRNA–miRNA duplexes) levels. (Fig. 4 ). At the molecular level, lncRNAs can sequester or recruit epigenetic modifying proteins, such as DNA methyltransferase and chromatin-remodelling complexes (Yu et al. 2017 , Canzio et al. 2019 , Liu et al. 2019 ). They can also stabilize their target transcripts (Geisler and Coller 2013 ) to increase mRNA expression (Ebert and Sharp 2010 ) and sequester proteins by forming lncRNA-protein complexes that alter mRNA splicing (Peng et al. 2020 ) (Fig. 4 ). LncRNAs have been proposed as putative regulators of diverse biological processes, including immune responses during host–pathogen interactions in mammals [for a review see Agliano et al. ( 2019 )]. They are also implicated in developing innate immune memory ‘priming’ (Zhang and Cao 2019 ) in invertebrates (Gourbal et al. 2018 ). In sea anemone, pre-exposure to a sublethal pathogen dose enhances short-term survival upon subsequent lethal exposures (Brown and Rodriguez-Lanetty 2015 ). Despite the availability of several reference genomes for corals, the mechanistic and genomic basis for immune priming in corals is currently completely unknown. In the context of the coral holobiont, more work is urgently needed in order to understand the functions of lncRNA regulatory networks. The role of ncRNAs in coral holobiont homeostasis is readily teased out from available transcriptomic data, demonstrating the fluctuations in the holobiont transcriptional output during disease progression, bleaching (Libro et al. 2013 , Daniels et al. 2015 , Pinzon et al. 2015 ), or host–pathogen/parasite interactions (Burge et al. 2013 , Mohamed et al. 2018 , van de Water et al. 2018 ). Preliminary data support the idea that ncRNAs act as mediators of both biotic or abiotic stress responses and symbiosis establishment. Early evidence of the presence of functional ncRNAs in corals was obtained from small RNA-seq experiments performed in Stylophora pistillata (Liew et al. 2014 ), where five microRNAs (miRNAs) among 50 were conserved in other metazoans. Other families of ncRNAs, such as lncRNAs, have also been recently described as putative players in coral responses to microbiome imbalance. In Palythoa caribaeorum , transcriptome analysis detected >10 000 expressed lncRNAs, some of which were conserved in higher eukaryotes (Huang et al. 2017 ). Investigation of differentially expressed lncRNAs in healthy colonies compared to bleached colonies indicated that upregulated lncRNAs in P. caribaeorum could act as post-transcriptional modulators of the Ras-mediated signal transduction pathway and components of the innate immune system, as part of the molecular response of corals to bleaching (Huang et al. 2017 ). ncRNAs have also been implicated in the establishment of endosymbiosis in cnidarians, e.g. miRNAs in the sea anemone Aiptasia (Baumgarten et al. 2018 ) and lncRNAs in the coral A. digitifera (Mohamed et al. 2016 , Huang et al. 2019 ). Using high-throughput sequencing cross-linking immunoprecipitation (HITS-CLIP), miRNAs were found to be differentially expressed and subsequently targeted genes implicated in Symbiodiniaceae colonization (e.g. FGFR, TGFβR, and components of the TNFR/TRAF pathways, arrest of phagosomal maturation, and sterol/peptide transporters) in response to endosymbiont infection (Baumgarten et al. 2018 ). Twenty-one out of 815 lncRNAs were differentially expressed 4 h post-colonization of algal symbionts (Huang et al. 2017 ) for which gene co-expression networks identified 6395 coral transcripts potentially regulated. Many of these transcripts were involved in the early stages of coral–algal interactions. Cumulatively, these studies provide preliminary evidence that ncRNAs modulate specific pathways related to symbiosis onset and breakdown of the cnidarian–Symbiodiniaceae relationship. The transfer of communication signals (ncRNA, proteins, metabolites) between cells via exosomes (Fig. 4 ) was first observed in mouse and human cells (Valadi et al. 2007 ). Functional transfer of ncRNAs among tissues of the same organism was demonstrated in animal models, where exosome-mediated transfer of miRNA from adipose tissue were released and transported through circulating fluids to their final target distant tissues (Thomou et al. 2017 ). Exosomes may be microbiota-derived to deliver microbial signals that can control diverse pathways in the host (Ahmadi Badi et al. 2017 ). Our current understanding of ncRNAs sorting and processing via exosomes stems from miRNAs in human and mammalian systems (Kogure et al. 2019 , Lee et al. 2019 ). However, the mechanisms of ncRNAs processing are not well understood in nonmodel species such as cnidarians and bacteria. ncRNAs are promising communication signals within the holobiont due to their common mechanisms of action among different compartments of the holobiont. The regulatory roles of molecules like miRNAs in eukaryotes and sRNAs in prokaryotes have been previously established (Blenkiron et al. 2016 , Viennois et al. 2019 ). On the other hand, the regulatory roles of other ncRNAs such as lncRNAs within holobiont context are still unknown. Integrated host–microbiota multiomics to understand holobiont biology Over the past decades, technical and computational advances have allowed the collective use of ‘omics tools to better understand different aspects of host–microbiome systems. In corals, however, most of these ‘omics tools have been used separately to understand a particular molecular level at a time. In this section, we describe the recent omics tools used in the field that shape our understanding of the coral microbiome and discuss the need for integrated coral–microbiota multiomics data to unravel the different layers of the coral microbiota interplay. Characterizing the structure and functional potential of the coral microbiome Recent advances in coral holobiont research have been possible due to the rise of ‘omics techniques, particularly genomics (Cooke et al. 2019 , Jasper et al. 2019 , Engelberts et al. 2021 ). Among the most widely used methods to characterize the diversity of coral microbiomes and understand their role in light of climate change is amplicon sequencing studies (Fig. 5 ). Most of these efforts focus on examining the diversity of coral-associated microbes using sequencing amplified variable regions of marker genes such as the 16S rRNA gene for prokaryotes or internal transcribed spacer 2 (ITS2) and the 18S rRNA gene for eukaryotes like Symbiodiniaceae and fungi. Amplicon sequencing is affordable, enabling the incorporation of large sample numbers; however, classification of prokaryotes using 16S rRNA amplicons is limited mostly to the family or genus level while species-/strain-level classification is mostly not attainable (Johnson et al. 2019 ). Shallow whole-metagenome shotgun sequencing has been proposed as a cost-effective tool to study microbial diversity at higher accuracy compared to amplicon sequencing (Xu et al. 2021a ). Even when species-level classification is possible using the full-length 16S rRNA gene (Matsuo et al. 2021 ), connecting such species to functional potential is difficult (Jasper et al. 2019 ). Inherent biases in PCR that are used in amplicon sequencing is another limitation (Aird et al. 2011 ). A workaround uses phylogenetic investigation of communities by reconstruction of unobserved states (PICRUSt, PICRUSt2) (Langille et al. 2013 , Douglas et al. 2020 ) and Tax4fun (Aßhauer et al. 2015 , Wemheuer et al. 2020 ), which enable prediction of gene function from 16S rRNA gene information based on publicly available reference genomes (Ainsworth et al. 2015 , Röthig et al. 2016 , Ziegler et al. 2017 , Hernandez-Agreda et al. 2018 ). While this is a useful approach for well-studied and sequenced microbes, prediction of function for microbes with unresolved phylogeny, cryptic microbes, and microeukaryotes is largely limited (Sun et al. 2020 ). In addition, genomic islands and plasmids in bacteria that are often horizontally transferred among bacteria and typically contain antibiotic resistance genes and other genes for rapid responses to environmental change are not resolved using PICRUSt (Sun et al. 2020 ). Figure 5. The different omics approaches deployed for studying the coral holobiont. An integrated coral–microbiota multiomics approach is needed to fully understand the biology of the coral holobiont. New approaches yet to be explored are highlighted in red. A handful of studies examined the coral microbiome using metagenomics and the subsequent acquisition of MAGs to study the functional potential of coral-associated microbes (Cai et al. 2017 , Meyer et al. 2017 , Robbins et al. 2019 , Keller-Costa et al. 2021 , Wada et al. 2022 ). Shotgun metagenomics enables the characterization of the diversity and functional potential of microbial communities, including the coral host with less bias than PCR-based amplicon sequencing. The assembly of MAGs enables direct examination of the role specific microbes may play in the holobiont. Metagenomics data can also integrate with other ‘omics techniques (discussed below) to provide a more holistic understanding of the holobiont. Despite these advantages, shotgun metagenomic data still suffers from some limitations. Shotgun metagenomics is significantly more expensive, though the cost is constantly decreasing, and computationally more demanding than amplicon sequencing. Due to the larger genomes and biomass of the coral host and Symbiodiniaceae cells compared to other microbial cells, shotgun metagenomics produces mostly coral and symbiont reads, with microbial reads becoming a small fraction of the total output. While in-silico separation of these microbial reads from other reads is possible, it is technically challenging and requires sequencing to high depths, especially when lacking reference genomes for the coral and Symbiodiniaceae. Physical size fractionation to enrich prokaryotic cells and DNA has been successfully applied in selected reef species, albeit with limitations (Littman et al. 2011 , Robbins et al. 2019 ). Nonetheless, more metagenomic sequencing is urgently needed to advance our understanding of what coral symbionts contribute to the host and the role of cryptic microbes in this relationship. Metatranscriptome profiling to study holobiont transcriptional responses The availability of draft genomes for several coral species paved the way to a wave of coral transcriptome-wide sequencing studies (Fig. 5 ). RNA-seq is the most widely used among the ‘omics methods to understand coral responses during exposure to disease (Libro et al. 2013 , Wright et al. 2015 , Anderson et al. 2016 , Frazier et al. 2017 ), establishment of coral–algal symbiosis (Mohamed et al. 2016 , Yuyama et al. 2018 , Mohamed et al. 2020b , Yoshioka et al. 2021 ), adaptation to the deep sea environment (Yum et al. 2017 ), responses to natural bleaching (Pinzon et al. 2015 , Rose et al. 2015 , Seneca and Palumbi 2015 ), and heat stress (Savary et al. 2021 , Voolstra et al. 2021b ). In addition to characterizing the transcriptional responses of coral hosts via mRNA differential analysis, metatranscriptomics has been scarcely utilized in corals to study the transcriptional responses of the microbiome primarily because of the need to overcome the relatively high abundance of host RNA relative to the microbiome. Studies that utilized metatranscriptomics have discovered new patterns in the holobiont. For example, Daniels et al. ( 2015 ) identified shared and distinct transcriptional responses to disease among different holobiont compartments, where innate immunity, cytoskeletal integrity, cell adhesion, and oxidative stress characterized the coral response, heat shock proteins, genes related to oxidative stress, and DNA repair characterized the bacterial response, and photosynthesis, and metal transport characterized the algal symbiont’s response. These results highlight a functional integration across the holobiont in response to disease. Metatranscriptome data from three different coral species identified host and algal symbiont genes that exhibited different changes in gene expression in a lineage-specific way (among the Robust and Complex coral clades) and showed higher bacterial diversity, bacterial metabolic capabilities, and transcriptional activity in the thermo-tolerant to -susceptible species suggesting potential roles for the bacterial microbiome in supplementing the metabolic needs of the holobiont during heat stress (Avila-Magaña et al. 2021 ). Metatranscriptomic data have also been generated from in-hospite Symbiodiniaceae in both adult and larval stages (Bellantuono et al. 2019 , Maor-Landaw et al. 2020 , Mohamed et al. 2020b ) highlighted a genealized transcriptome-wide suppression that includes photosynthesis and protein synthesis during symbiosis onset. Coupling gene co-expression networks with identifying key regulators in metatranscriptomics Gene networks have been recently utilized to identify gene expression patterns in corals using weighted gene coexpression network analysis (WGCNA), that quantifies the co-expression patterns among DE genes, to identify clusters of highly correlated genes (gene modules) (Langfelder and Horvath 2008 ). These efforts have mainly focused on coral gene expression data to identify expression modules or ‘clusters of co-expressed genes’ (Rose et al. 2015 ). Indeed, WGCNA-inferred gene networks revealed a potential adaptive mechanism named ‘transcriptional frontloading’, which means the constitutive higher baseline of expression levels of stress response genes in well-adapted ‘stress resilient’ corals compared to less-adapted ‘stress susceptible’ counterparts (Brashis et al. 2013 ). WGCNA analyses were implemented to study the response to experimental heat stress of A. hyacinthus colonies that had been transplanted between two differing reef environments (Rose et al. 2015 , Bay and Palumbi 2017 ). These experiments identified modules of coexpressed genes where some of which correlated strongly with the bleaching responses of individual colonies, hence called ‘bleaching modules’. These genes were proposed as potential biomarkers for predicting coral survival under environmental stress. However, the applied WGCNA approach is best for identifying expression modules but has not been used to pinpoint to master regulators that could control the transcriptome remodelling due to a given perturbation. Other information can be readily extracted from coral transcriptomes beyond differentially expressed genes, such as the regulatory potential of TFs and ncRNAs including lncRNAs. Despite the availability of the needed coral transcriptomic data, these analyses are not widely performed in the coral field. Amongst other promising network approaches, the Partial Correlation and Information Theory (PCIT) algorithm (Reverter and Chan 2008 ) aims at identifying key regulatory factors within the gene network by applying. PCIT combines partial correlation coefficients with information theory to explore all the correlations between possible triplets of genes within the dataset prior to the identification of significant correlations. This approach has been coupled with the concept of differential networks (the difference in connections per node from one network to another) (Ideker and Krogan 2012 , Hu et al. 2016 ) to understand various traits in other species and identify master regulators (Cánovas et al. 2014 , Wouters et al. 2020 , Botwright et al. 2021 , Mohamed et al. 2022 ). Master regulators undergo substantial changes in connectivity to genes during the transition between physiological states so that differential connectivity may identify highly differentially connected genes between the networks (Hudson et al. 2012 ). These master regulatory genes within a network may act as key regulatory components of transcriptional networks that could be used further in functional assays (e.g. gene editing) to understand their functions beyond correlations. A shift from differential expression to differential networks in coral molecular studies would allow an understanding of the gene regulatory circuits underlying various traits in corals. Proteomics and metabolomics potential to uncover the metabolic activity of the holobiont While metatranscriptomics alone has proven useful in the few examples it has been applied, integrating with other meta-’omics techniques can further confirm gene expression patterns, quantify their impact and shed light on new metabolic and symbiotic patterns. Particularly, mass-spectrometry based techniques, like proteomics and metabolomics (Fig. 5 ), are elucidating new discoveries in holobiont research. Proteomics was recently successfully used to distinguish between host and symbiont-related responses due to heat stress, showing downregulation of symbiosis signals in the host, and photosynthesis breakdown in the symbiont (Mayfield et al. 2021 , McRae et al. 2021 , Petrou et al. 2021 ). Proteomics also shed light on the potential role of bacterial symbionts, such as vitamin supply by Endozoicomonas to their host (Pogoreutz et al. 2022 ). New molecules such as the dipeptides lysine–glutamine and arginine–glutamine have been identified as molecular biomarkers for coral thermal stress (Williams et al. 2021a ). Although proteomic data correlates well with physiological data, there is a temporal delay in transcript-to-protein responses in most organisms that often renders proteomic data in apparent disagreement with transcriptomic data if this delay is not taken into consideration (Mayfield et al. 2018 ). However, proteomic data are considered more representative of observed phenotypes, unlike transcriptomics where epigenomic regulation and post-translational modifications may skew transcriptome–phenotype comparisons (Manzoni et al. 2018 ). Therefore, proteomic studies need to become a larger component of coral studies. Compared to other ‘omics techniques, metabolomics (Fig. 5 ) has largely lagged due to major limitations in detection of low-abundance metabolites and the large number of cryptic metabolites that are often observed in biological systems (Vohsen et al. 2019 ). Therefore, most studies resolve to compare overall metabolic profiles at various conditions, while highlighting only putative identifications (Sogin et al. 2017 , Hillyer et al. 2017a , Lohr et al. 2019 ). Several metabolomics studies have shed light on new mechanisms of interactions within the coral holobiont, such as increases in platelet activation factors at coral–algal interfaces (Quinn et al. 2016 ). The correlations of various lipid classes like betaine–lipids and diacyl–glycerides were able to indicate previous bleaching events of coral colonies (Roach et al. 2020 ), can be markers of disesases (Deutsch et al. 2021 ). Likewise, dipeptides were reported as indicators of heat stress (Williams et al. 2021a ). By sampling coral colonies at different distances from the coral animal surface, dozens of metabolites and chemical features were found to form a gradient around coral colonies; these metabolites were composed of diverse chemical classes that may be important in structuring the SML microbiome (Ochsenkühn et al. 2018 ), such as the hormone estrogen (Vilela et al. 2021 ), which can restructure the microbiome in addition to controlling stress responses (Stien et al. 2020 ). In-silico analysis of mass shifts in holobiont mass spectrometry datasets, like the addition/substraction of e.g. hydroxy (−OH) or methyl groups (−CH3), can be used to reveal genetic differences and even can be correlated to transcriptomic variances of closely related organisms (Hartmann et al. 2017 ). Beyond predicting metabolites, examination of elemental exchanges within the holobiont using stable-isotope labelling (e.g. 13 C, 15 N) of metabolites or nutrients has been useful in determining the role of diazotrophs in the holobiont as a source of nitrogen in a stable symbiotic state but disregards an oversupply of microbial derived N during heatstress (Rädecker et al. 2021 ). Isotope labelling of coral fragments in 13 C-bicarbonate enriched media of e.g. amino acids, fatty acids or lipogenesis intermediates confirmed under severe heat stress conditions decreases of de novo biosynthesis of fatty acids in the symbiont leads to a consequent decrease in fatty acids in the host (Hillyer et al. 2017b ). Despite its promises and relatively low cost compared to sequencing, metabolomics suffers from several limitations. The immense diversity of chemical formulae and resulting structures poses the biggest challenge in mass spectrometry-based metabolomics, where annotations can only be confirmed reliably with comparisons to known standards. Further, extraction protocols, e.g. differences in solvent mixtures, disruption techniques, or even cooling, have an influence on detected molecules as some metabolites are prone to degradation (Lu et al. 2020 ), while others are not efficiently extracted (Andersson et al. 2019 ), which calls for standardization. In the context of the holobiont, a major challenge is linking metabolites with their producing organism. Statistical correlation of metabolite and amplicon sequencing data enables linking of an organism with its metabolites (Jorissen et al. 2021 ); however, most central metabolites are shared across different organisms within the holobiont and so this approach only applies to unique secondary metabolites. Using metagenomic, metatranscriptomic, or proteomic data can enhance metabolomics significantly since enriched organisms/genes/proteins from a specific organism is often indicative of a corresponding increase in metabolite abundance. For example, steroids depend on the presence of algal symbionts, which cnidarians are unable to produce. Therefore, the expulsion of symbionts during bleaching, consequently leads to a decrease of steroids in the host (Jiang et al. 2021 ). Another approach relies on isolating and culturing microbes from the host, characterizing metabolites from these microbes, and combining this information with data acquired from coral samples. This approach identified antioxidants and osmolytes, like betaine–lipids or glucosides from the coral microbiome, which are hypothesized to increase holobiont stress resistance (Gegner et al. 2017 , Ochsenkühn et al. 2017 , Roach et al. 2021 ). Despite its limitations, metabolomics is steadily gaining traction with improvements in protocols, instrumentation and analysis and is becoming a complementary tool with high predictive power (Lu et al. 2017 , Greene et al. 2021 , Wegley Kelly et al. 2021 ). Multiomics data are urgently needed for a holistic understanding of the holobiont A single layer of omics is not usually adequate to understand a complex system such as the coral holobiont. An integrated host–microbiota multiomics framework has been developed and proposed to understand other holobiont systems (Nyholm et al. 2020 ). Nyholm et al. recently coined the term ‘holo-omics’ to describe experiments that aim to obtain multiple omics data from both host and microbiota domains. This holistic approach has been recently applied to study the plant microbiome (Zolti et al. 2020 ; for a review, see Xu et al. 2021b ), the human microbiome (Heintz-Buschart et al. 2016 , Lloyd-Price et al. 2019 , Park et al. 2022 ) [for a review, see Zhang et al. ( 2019 )]. For example, Lloyd-Price et al. ( 2019 ) provided a comprehensive multiomics data during the functional dysbiosis in the human gut microbiome upon the progression of the inflammatory bowel disease. In the Lloyd-Price paper, metagenomic, metatranscriptomic, and stool metabolomic profiles were combined to show a unique microbiome restructure characterized by an increase in facultative anaerobes at the expense of obligate anaerobes and identify biochemical and host factors central to this dysregulation. In corals, most of the omics data generated were obtained solo; there have been attempts to attain multiomics data (Cziesielski et al. 2018 , Maruyama et al. 2021 , Santoro et al. 2021 , Voolstra et al. 2021b , Williams et al. 2021b , Pogoreutz et al. 2022 ). However, most of these studies have exclusively relied on descriptive microbiome tools such as amplicon sequencing (Table 1 ), that have been significant in shaping our understanding of the coral microbiome composition; but limited in providing mechanistic insights into mechanisms of coral–microbiome interactions. However, a few examples showed the power of integrating metatranscriptomes and metagenomes to understand microbial processes during onset and progression of the black band disease (Arotsker et al. 2016 , Sato et al. 2017 ). Williams et al. ( 2021 ) combined polar metabolomics with host transcriptomics to investigate gene–metabolite interactions in the coral Montipora capitata exposed to a 5-week period of thermal stress. The gene–metabolite integrated analysis revealed thermal stress affects reproductive activity evidenced by the downregulation of CYP-like genes and the irregular production of sex hormones. Despite these data being focused on the coral host, they provided a set of genes and metabolites that can be used as markers of coral thermal stress. Indeed, leveraging the integrated metatranscriptomic and metagenomic data would greatly enhance our understanding of onset of coral disease that will help in coral disease prognostics and coral bleaching management. Santoro et al. ( 2021 ) was a clear example that made use of manipulative experiments, physiology to define the phenotype before collecting multiomics data, 16S rRNA amplicon sequencing, host gene expression, and metabolomes that were well correlated with physiological responses associated with health status. However, we currently lack comprehensive functional insights into coral–microbiome interactions, despite the recent advances in coral microbiome research. To reach a more comprehensive, systems-level view of coral–microbiome interactions, experiments should focus on pairing the host-centred omics data such as host transcriptome, epigenome, and metabolome/proteome with microbially centred data such as shotgun metagenomes, metatranscriptomes, and meta-metabolome/meta-proteome (Fig. 5 ). Table 1. Recent coral studies adopting multiomics to investigate the molecular basis of disease onset, thermal tolerance, and symbiosis. The coral species, the ‘omics approaches utilized, and the scope of the study are shown. Study Coral species ‘Omics approach Scope Daniels et al. 2015 \n Orbicella faveolata \n Metatranscriptomics White plague disease Host, algal/bacterial symbionts Sato et al. 2017 \n Montipora hispida \n Metatranscriptomics Black band disease Shotgun metagenomics Meyer et al. 2017 \n Montastraea cavernosa \n 16S rRNA amplicon sequencing Black band disease Shotgun metagenomics Cziesielski et al. 2018 \n Aiptasia pallida , a coral model Host transcriptomics Thermal stress Proteomics Cleves et al. 2020 Review article Multiomics and reverse genetics Thermal stress Roach et al. 2020 \n Diploria strigosa and O. faveolata Shotgun metagenomics Coral–turf algal interactions Metabolomics Williams et al. 2021b \n M. capitata \n Host transcriptomics, metabolomics Thermal stress Santoro et al. 2021 \n M. hispida \n 16S rRNA amplicon sequencing Microbiome-enabled thermal tolerance Host transcriptomics Metabolomics Voolstra et al. 2021b \n Stylophora pistillata \n Host + algal transcriptomics and ITS2/16S rRNA amplicon sequencing Thermal tolerance Savary et al. 2021 \n S. pistillata \n Host + algal transcriptomics and ITS2/16S rRNA amplicon sequencing Thermal tolerance Maruyama et al. 2021 2021 \n Acropora tenuis \n Host + algal transcriptomics, 16S rRNA amplicon sequencing, and bacterial genome sequencing Coral–microbiota interactions Pogoreutz et al. 2022 \n A. humilis \n 16S rRNA amplicon sequencing, host transcriptomic, proteomics, and bacterial genome sequencing Coral– Endozoicomonas symbiosis Current limitations and new directions Despite years of significant work on cnidarian symbiosis [for reviews see Weis ( 2019 ), Rossett et al. ( 2021 )], insights into the onset of coral–algal symbiosis (Davy et al. 2012 , Mohamed et al. 2016 , 2020b , Yoshioka et al. 2021 ), and more recently the role of nutrient cycling in the breakdown of coral–algal symbiosis during heat stress (Rädecker et al. 2022 ), a comprehensive mechanistic understanding of the synthesis, homeostasis, and demise of symbiosis due to coral bleaching is still lacking. Many of the current insights stem from studies conducted on other cnidarian species such as Aiptasia [see e.g. Celeves et al. ( 2020 )] as a coral model, due to difficulties in experimentally manipulating corals in the lab. Only recently, the tropical stony coral species Galaxea fascicularis was proposed as a candidate coral model system (Puntin et al. 2022 ). This will indeed help us understand the molecular underpinnings of coral symbiosis in the near future. Despite the high potential of multiomics data, the presence of cryptic biological information (‘biological dark matter’, e.g. the noncoding part of the genome (Eisenstein 2021 ) and hypothetical proteins with unknown function (Stephens et al. 2018 ) and cryptic or undescribed microbes (Lok 2015 )) is a major limiting factor for understanding the coral holobiont. On the other hand, traditional molecular biology and biochemistry techniques to characterize gene/protein function are time consuming and are low throughput. Gene networks have been proven to detect hypothetical proteins with presumable importance (Cleves et al. 2020 ). By applying the ‘guilt by association’ principle, these genes/proteins or unclassified microbes connected with other genes/microbes of known function (Fig. 5 ) may shed light on their importance. A few such genes were co-expressed along other genes in the same module that was upregulated shortly after thermal stress (Cleves et al. 2020 ). These hypothetical proteins of interests can be then folded using Alpha-fold to gain more insights into their functionality (Ma et al. 2022 ). More methods such as RNA interference and CRISPR/CAS9-mediated genome editing that enable identification and characterization of microbial and gene functions are needed to overcome this major hurdle in biology. Recent advances in genomics have allowed the construction of detailed cell type atlases for a soft coral ( Xenia sp.) and a stony coral ( Stylophora pistillata ) using single-cell RNA sequencing (scRNA-seq) (Hu et al. 2020 , Levy et al. 2021 ), chromosome-level genome assemblies (Fig. 5 ) for the soft coral Xenia sp. and the hard corals A. millepora and M. capitata (Fuller et al. 2020 , Hu et al. 2020 , Stephens et al. 2021 ), and the coral endosymbiont S. microadriaticum (Nand et al. 2021 ). Recent initiatives such as the Aquatic Symbiosis Genomics (ASG) project (McKenna et al. 2021 ) will provide chromosome-level genomes for ∼500 symbiotic systems including corals and their microbiomes. Within the ASG project, a combination of long-read and long-range genomic data will be generated using the Pacific Biosciences sequencing platform to generate high fidelity reads in the 15–20 kb range, along with Oxford Nanopore long-read technologies. Transcriptome data will be generated to help annotate those genomes. Chromatin conformation capture sequencing, known as 3C or Hi−C, will be used to link sequences within chromosomes and organelles (Belton et al. 2012 ). The same proximity ligation strategy has been developed recently to study the microbiome (meta3C). The ProxiMeta platform is designed to deconvolute chromosomes and plasmids in a mixed microbial sample into complete genomes (Stadler et al. 2019 ). More recently, spatial transcriptomics (based on 10× Genomics data) has been developed to couple gene expression data with spatial information (spatially resolved transcriptomes) that enable measuring all the gene activity in a given tissue sample and mapping where the activity is occurring (Larsson et al. 2021 , Rao et al. 2021 ). These new approaches will enable disentangling functions of different microbiome members within the holobiont. These data will provide baseline genomic resources that will revolutionize the way researchers look at and design symbiosis experiments to understand the molecular mechanisms underlying homeostasis of the coral holobionts and their responses to climate change. The Assay of Transposase Accessible Chromatin sequencing (ATAC-seq) has become increasingly popular for detecting chromatin accessibility (Yan et al. 2020 ). The quest of identifying regulatory elements across different cell types and developmental stages has led to large international efforts mostly focusing on model organisms, such as the human Encyclopedia of DNA Elements (ENCODE; Moore et al. 2020 ), and the Functional Annotation of Animal Genomes (FAANG; Giuffra et al. 2019 ), to unravel the regulatory elements in model/nonmodel organisms. Spatiotemporal changes in the epigenome at the chromatin level are crucial to development, cellular differentiation, health, and disease (Gorkin et al. 2020 , Fang et al. 2021 ). In corals, epigenetic mechanisms have been exclusively studied through DNA methylation. The study by Liew et al. ( 2018 ) reported that DNA methylation levels in corals are highly sparse, where only 9% of genome-wide CpG loci were methylated, most of which are co-located within gene bodies, in contrast to higher-level genome-wide methylation levels in promoters and enhancers in vertebrates. ATAC-seq data, paired with RNA-seq, was utilized for the first time to study the cnidarian–dinoflagellate model Exaptasia pallida to reveal a role of chromatin dynamics in response to thermal stress (Weizman and Levy 2019 ). Compared to DNA methylation, chromatin accessibility data (Fig. 5 ) hold a great promise towards understanding epigenetic mechanisms and their role in regulating gene expression in corals as shown for other nonmodel species (Alexandre et al. 2021 , Mohamed et al. 2022 ). Indeed, defining the coral regulatory vocabulary would allow understanding many aspects of coral biology including responses to climate change at an unprecedented level. The use of the forementioned omics technologies and their integration will be only valuable when combined with specific manipulative experiments with accurate physiological and/or environmental monitoring to be correlated with the molecular and/or microbial data. The correlation of the multiomics data with the holobiont’s physiological status (phenotype) would provide valid hypothesis to be developed to understand coral thermal thresholds, their response to stress, susceptibility to pathogens, and resilience. Finally, these hypotheses have to be validated with further genetic or microbial manipulative experiments to confirm these sequencing-based findings (Fig. 6 ). Figure 6. Strategies for conducting experiments aimed at understanding coral holobiont functions. The collection of multiomics data representing both the coral (host omics) and its microbiome (metaomics). The integrated multiomics analyses need to be paired with accurate metadata measurements for correlations. Hypotheses drawn from the integrated multiomics data should be tested by further validation experiments.\n\nIntroducing heat-tolerant Symbiodiniaceae into corals It is widely accepted that coral thermal tolerance is largely dependent on the physiology of their associated Symbiodiniaceae partners (Berkelmans and van Oppen 2006 ). In-vitro exposure of Symbiodiniaceae cultures to elevated temperatures increases their thermal tolerance after ∼40 generations (Chakravarti et al. 2017 , Chakravarti and van Oppen 2018 ). Despite this acclimation, reintroducing heat-tolerant strains into corals yielded no significant benefit for the holobiont (Chakravarti et al. 2017 ). In contrast, a small minority of heat-tolerant Symbiodiniaceae strains derived from the same wild-type clone increased the thermal tolerance of coral larvae (Buerger et al. 2020 ). A mechanistic understanding of how heat tolerance in Symbiodiniaceae occurs and how in turn it influences the coral holobiont is needed to improve the efficacy of this approach."
} | 26,093 |
35694519 | PMC9178609 | pmc | 31 | {
"abstract": "Self-healing materials\nplay an essential role in the field of organic\nelectronics with numerous stunning applications such as novel integrated\nand wearable devices. With the development of stretchable, printable,\nand implantable electronics, organic field-effect transistors (OFETs)\nwith a self-healable capability are becoming increasingly important\nboth academically and industrially. However, the related research\nwork is still in the initial stage due to the challenges in developing\nrobust self-healing electronic materials with both electronic and\nmechanical properties. In this mini-review, we have summarized the\nrecent research progress in self-healing materials used in OFETs from\nconductor, semiconductor, and insulator materials. Moreover, the relationship\nbetween the material design and device performance for self-healing\nproperties is also further discussed. Finally, the primary challenges\nand outlook in this field are introduced. We believe that the review\nwill shed light on the development of self-healing electronic materials\nfor application in OFETs.",
"conclusion": "4 Conclusions and Outlook With the rise of\nnovel portable and wearable electronic technology,\nmaterials with preferable self-healing ability have broad prospects\nin OFETs. In this mini-review, we have highlighted the recent advances\nof SHMs applied in OFETs. In general, the popular patterns are divided\ninto extrinsic and intrinsic self-healing modes for the SHMs. The\nintrinsic SHMs have become a research hot spot due to their stable\nand reliable self-healing ability, which could also accomplish multiple\nrecoveries via reversible dynamic covalent or noncovalent bonds. We\nintroduced the recent progress in intrinsic SHMs applied in OFETs\nfrom three aspects, including conductor, semiconductor, and insulator\nmaterials. Moreover, the existing challenges and perspectives of SHMs\nin the OFETs have also been discussed. The SHMs could significantly\nextend the service lifetime, improve the reliability, availability\nand affordability, and decrease the replacement costs of OFETs. However,\nthe performance of reported self-healable OFETs is not yet comparable\nto that of their nonhealable counterparts. There are still several\nkey challenges toward the creation of efficient SHMs with satisfying\nelectronic properties. First, multiscale theoretical simulation\nand transient analysis\ntechniques are required to elucidate the healing mechanism at the\nmolecular level, which may provide new guidelines to synthesize novel\nmaterials, especially tolerable conductors and semiconductors with\nlimited availability. Second, the self-healing process is relatively\nslow and nonautonomous. Thus, it is highly important to develop an\nautonomous self-healing system avoiding the use of heat, light exposure,\nor solvents. Third, the selective triggering of the self-healing process\nin designated locations is being pursued in integrated electronic\nsystems. In addition to that, the self-healing behavior is usually\nlimited to small-area damages. It is urgent to explore SHMs with the\ncapability of large-area recovery. Fourth, the mechanical properties\nare mostly sensitive to temperature due to the dynamic noncovalent\ncross-linked bonds in intrinsic SHMs. Therefore, the thermal stability\nneeds to be improved by constructing thermodynamically stable self-healing\nsystems. We anticipate that further advances in electronic SHMs\nwill be\nmade to promote the application of self-healing OFETs in wearable\nelectronics and bioelectronics by the deep collaboration of scientists\nin the fields of chemistry, material science, electronics, theoretical\ncomputation, and engineering.",
"introduction": "1 Introduction With the growing demands\nof renewable energy and the fast development\nof organic electronics, materials with self-healing ability have attracted\nincreasing interest due to their vast prospects in the fields of electronic\nskins (E-skins), sensors, supercapacitors, OFETs, solar cells, and\nso on. 1 − 4 However, it has been a persistent problem that the accumulation\nof uncontrolled damages by abrasion, breakage, aging, degradation,\nmechanical damage, or operational fatigue in the process of actual\nuse would lead to an attenuation of the device performance and an\nincreasing shortening of the service life. 5 In terms of this issue, with the remarkable ability of human skin\nto restore itself from wounds as inspiration, novel intelligent self-healing\nmaterials (SHMs) that possess the ability to autonomically repair\ndamages inflicted during the operation procedure are becoming increasingly\nimportant. 6 , 7 These functional materials with self-healable\nability could dramatically increase the durability and prolong the\nlifetime of the devices. Although very promising progress has been\nmade in soft self-healing fields such as wearable sensors, E-skins,\nfabrics, and so on, there is still a long way to go for self-healing\nmaterials to be used in practical applications or in commercial demands.\nSelf-healing materials should be optimized and improved in three aspects,\nincluding fast healing efficiency, biocompatibility, and low cost,\nwhich could enhance the performance and lifespan of electronic devices. 7 In general, the current SHMs can be broadly\ndivided into two categories\nbased on the trigger requirements and the essential attributes of\nthe self-healing process: nonautonomous healing systems and autonomous\nhealing systems. 5 The nonautonomous SHM\nsystems require external stimulation such as temperature, light, heat,\npH and so on, while the autonomous healing systems promptly initiate\nthe self-healing behavior when they suffer from damages. 8 In addition to the above classifications, the\nSHMs can also be divided into extrinsic self-healing and intrinsic\nself-healing depending on whether healing agents are added or not\n( Figure 1 ). The extrinsic\nself-healing strategy is implemented by releasing the healable agents\nthat are encapsulated in the carriers (such as microcapsule and microvascular\ncarriers) to restore the original function in damage locations. In\ncomparison, the intrinsic self-healing procedure depends on the reconstruction\nof noncovalent supramolecular interactions or dynamic covalent bond\nnetworks in SHMs. In other word, the external self-healing ability\nrelies on polymerization or chemical reactions, while the intrinsic\nself-healing functions usually depend on the chemical cross-linking\nformed by dynamic covalent bonds or physical cross-linking generated\nby supramolecular interactions, which could benefit the fulfillment\nof multiple healing processes at the same location. 9 , 10 However,\nfor comparison, in general device systems, there are neither chemical\ncross-linking effects nor external polymerization or chemical reactions.\nThe reversible mutual effect cannot occur when a device suffers damage,\nand thus it cannot exhibit self-healing behavior. Figure 1 Schematic illustration\nof self-healing polymer systems through\n(a) an extrinsic self-healing mode and (b) an intrinsic self-healing\nmode. Reproduced with permission from ref ( 10 ). Copyright 2017 Wiley-VCH. Organic field-effect transistors (OFETs) are three-terminal switching\nelectronic devices controlled by the gate voltage, providing adjustable\noutput current within a certain range. 11 An OFET is the basic building component in electronic circuits,\nwith the multiple advantages including low cost, readout integration,\nlarge-area coverage, and power efficiency, and it possesses wide applications\nin sensor arrays, active matrix displays, logic circuits, and radio-frequency\nidentification tags. 9 Actually, the performance\nof OFETs has been significantly improved in the last decades, but\nthere are still some difficulties in the areas of long-term stability\nand large-area uniformity. The key reasons are the interface compatibility\nproblems between organic semiconductors and amorphous polymer insulators\n(API). In addition, a functional API may expand the application fields\nfor OFETs, such as wearable electronics, high-density memory, and\nself-healing devices, which makes the investigation of APIs to become\nmore and more important. 11 Nevertheless,\nwith the development of flexible, wearable, and self-healing electronics,\nOFETs face more challenges due to mechanical damages such as cracks\nand scratches in a long-term use process, which inevitably lead to\na lower durability, shorter service life, and worse performance. Consequently,\nOFETs with self-healing ability have attracted increasing attention,\nespecially in an intrinsic self-healing procedure that can repair\nthe damages without external intervention (with reversible interactions\nof covalent or noncovalent bonds and/or the molecular movement and\nrearrangement within polymer networks). 8 To date, some important reviews have discussed the self-healing\nresearch development of SHMs. 3 For example,\nBao et al. summarized the developments of functional devices and integrated\nsystems based on self-healing electronic materials from the perspective\nof soft electronics. 4 The Haick group discussed\ncontemporary studies of self-healing soft sensors on material design,\ndevice structure, and fabrication methods. 12 Latif et al. discussed the potential advantages and challenges of\nself-healing materials. 13 Nevertheless,\nas far as we know, reviews on the systematic introduction of self-healing\nOFETs are rare, although it is very important for the development\nof future flexible organic electronic circuits. Therefore, it is necessary\nand urgent to summarize the self-healing OFETs. For this purpose,\nwe have summarized the common self-healing strategies and mechanisms\nfor SHMs and described the key breakthroughs and recent progress in\nSHM OFETs. 2 , 9 Finally, we address several points for further\nexploration in this flourishing field. We believe that this mini-review\ncould give some guidance and motivate the development of SHM OFETs.\n\n2 Brief Introduction of Intrinsic SHMs There is a long\nhistory of SHMs imitating living organisms to confront\ndeleterious damage and recover the original functionality. Numerous\nattempts have been made to exploit materials with a self-healing ability.\nTo assess the self-healing ability of a material, three important\nfactors have been proposed: localization (position of the damage),\ntemporality (recovery time), and mobility (dynamic interactions). 14 The recovery process can be achieved via the\ninteractions of dynamic bonds or the mutual effects of entanglement\nand diffusion of polymer chains. 15 White\net al. further defined healing efficiency (η) as a ratio of\nchanges in material properties as where f is the property of\ninterest. All of these strategies bring positive help and effective\ncomparison to allow us to evaluate the self-healing ability of materials\nand/or devices. Despite significant progress in the development of\nthe self-healing field, there are great differences in the methods\nand characterizations to measure self-healing, which require further\nresearch and discussion. 16 A specific scheme\ncan be designed and optimized according to the damage of the material. 8 It is noteworthy that the development of\nreliable and durable intrinsic\nSHMs cannot occur without a reversible transformation of dynamic bonds,\nwhich can avoid complex problems of integration and compatibility\nof an extrinsic self-healing behavior. As researchers continue to\nexplore, a variety of novel self-healing strategies have been designed:\nfor example, the dynamic covalent bonding interactions ( Figure 2 ) with a self-healing ability,\nincluding Diels–Alder reactions, imine bonds, disulfide exchanges,\nsilyl ether linkages, acylhydrazone bonds, diarybibenzofuranone bonds,\nalkoxyamine bonds, borate ester bonds, diselenide bonds, and hindered\nurea bonds. These covalent bonds usually have a strong bond energy,\nso that the material can exhibit a satisfying mechanical property\nand healing ability. 2 , 8 , 17 However,\nmost of the SHMs with dynamic covalent bonds need external stimuli\n(such as light, heat, and pH change) to accelerate the repair rate. 18 In comparison, noncovalent bonding interactions\nwith self-healing ability involving host–guest interactions,\nmetal–ligand coordination, hydrogen bonds, π–π\nstacking interactions, electrostatic interactions, dipole–dipole\ninteractions, and van der Waals forces ( Figure 3 ) 14 usually have\na lower kinetic stability and a reversible process of dissociation\nand generation without huge energy consumption. By modification of\nthe types of reversible dynamic bonds and the mobility of chains,\nthe self-healing conditions, efficiency, and the modulus and liquidity\nof the related polymer can be regulated and optimized. 1 As a matter of fact, dynamic behaviors of SHMs on multiple\nlength scales are important for implementing a spontaneous intrinsic\nhealing process. At the macroscopic level, the interfaces of damaged\nposition must be adequately close to each other, thus promoting the\ndynamic reorganization process. At the molecular level, the obtained\npolymers must provided plentiful dynamic interactions to exhibit a\nsufficient dynamics of the polymer chains. 15 Figure 2 Chemical\nstructures of various dynamic covalent bonds used in the\nself-healing process. Reproduced with permission from ref ( 15 ). Copyright 2020 American\nInstitute of Physics. Figure 3 Chemical structures of\nvarious dynamic noncovalent bonds used in\nself-healing processes. Reproduced with permission from ref ( 15 ). Copyright 2020 American\nInstitute of Physics. In comparison to extrinsic\nSHMs, intrinsic SHMs can achieve multiple\nrecoveries, avoiding encapsulation and integration of the healing\nagent in the matrix. Moreover, the intrinsic SHMs are more reliable\nand durable due to the recombination and reconstruction of intrinsic\nreversible dynamic bonds. 18 Moving forward,\nvarious ideas based on intrinsic self-healing behavior have appeared\nin numerous fields, and we mainly focus on the advances in intrinsic\nSHMs. In 2008, Leibler et al. first employed the intrinsic self-healing\nmethodology into a supramolecular material, which could be associated\ntogether to generate chains and cross-linked networks via reversible\ndynamic hydrogen bonds. 19 When it suffers\nfrom mechanical damage, the supramolecular material can be repaired\nby bridging the fractured surfaces at room temperature, and the recovery\nprocess can be accomplished many times. Since then, the intrinsic\nSHMs have gained extensive attention and gradually became the focus\nof research. In 2012, Bao et al. demonstrated the first electronic\ncomposite material with an ambient, repeatable self-healing property\nby embedding nanotextured nickel (Ni) microparticles into a supramolecular\norganic polymer. 20 The healing efficiency\nof the composite material decreased with the surface exposure time\ndue to the hydrogen-bonding reassociation between the cut surfaces.\nThen, Bao et al. introduced a novel chemical moiety to enhance the\ndynamic noncovalent cross-linking of the polymer, which could restore\nits initial morphology and charge mobility upon thermal and solvent\nannealing treatments when it was damaged by mechanical strain. 21 The prepared fully stretchable transistor exhibited\nsatisfying stretchability and healing performance, paving the way\nfor the development of self-healing OFETs. In recent years,\nmaterials with intrinsic self-healing ability\nare becoming a research hot spot due to their significant role and\nbroad demand in future artificial applications with long durability.\nThere are still some challenges to be solved. For example, SHMs with\nbetter self-healing efficiency, rapid healing speed, and stable conductivity\nare pressingly urgent for the fabrication of devices with a long device\nlife. In addition, the synergistic effect of good conductivity and\nmechanical performance as well as a simple fabrication process and\nmild experimental conditions are required to be considered for the\ndesign of SHMs. Therefore, we believe that the progress in SHMs will\ngreatly improve the development of flexible electronics."
} | 4,015 |
30533318 | PMC6282938 | pmc | 32 | {
"abstract": "The coral symbiosis is the linchpin of the reef ecosystem, yet the mechanisms that promote and maintain cooperation between hosts and symbionts have not been fully resolved. We used a phylogenetically controlled design to investigate the role of vertical symbiont transmission, an evolutionary mechanism in which symbionts are inherited directly from parents, predicted to enhance cooperation and holobiont fitness. Six species of coral, three vertical transmitters and their closest horizontally transmitting relatives, which exhibit environmental acquisition of symbionts, were fragmented and subjected to a 2-week thermal stress experiment. Symbiont cell density, photosynthetic function and translocation of photosynthetically fixed carbon between symbionts and hosts were quantified to assess changes in physiological performance and cooperation. All species exhibited similar decreases in symbiont cell density and net photosynthesis in response to elevated temperature, consistent with the onset of bleaching. Yet baseline cooperation, or translocation of photosynthate, in ambient conditions and the reduction in cooperation in response to elevated temperature differed among species. Although Porites lobata and Galaxea acrhelia did exhibit the highest levels of baseline cooperation, we did not observe universally higher levels of cooperation in vertically transmitting species. Post hoc sequencing of the Symbiodinium ITS-2 locus was used to investigate the potential role of differences in symbiont community composition. Interestingly, reductions in cooperation at the onset of bleaching tended to be associated with increased symbiont community diversity among coral species. The theoretical benefits of evolving vertical transmission are based on the underlying assumption that the host-symbiont relationship becomes genetically uniform, thereby reducing competition among symbionts. Taken together, our results suggest that it may not be vertical transmission per se that influences host-symbiont cooperation, but genetic uniformity of the symbiont community, although additional work is needed to test this hypothesis.",
"conclusion": "Conclusions This study investigated whether corals employing vertical symbiont transmission also exhibit enhanced cooperation and holobiont fitness. Contrary to theoretical predictions, we did not find significant support for the role of vertical transmission in spite of significant differences in cooperation among our six focal species. In a post hoc analysis of other drivers, we found that a greater initial symbiont cell density was associated with a greater bleaching intensity, but this association did not appear to result from an alteration of host-symbiont cooperation. Rather, the reduction in cooperation across species at the onset of bleaching was marginally associated with symbiont community diversity. The theoretical benefits of evolving vertical transmission are based on the underlying assumption that the host-symbiont relationship becomes genetically uniform, thereby reducing competition among symbionts. Taken together, our results suggest that it may not be vertical transmission per se that influences host-symbiont cooperation, but genetic uniformity of the symbiont community, though future work is needed to directly test this hypothesis.",
"introduction": "Introduction Cooperation between species has played a fundamental role in the evolution and diversification of life ( Friesen & Jones, 2013 ; Kiers & West, 2015 ; Maynard Smith & Szathmary, 1995 ; Moran, 2006 ). In the case of reef-building corals, the intracellular symbiosis between dinoflagellates in the genus Symbiodinium and a calcifying Cnidarian host forms the basis of one of the most biodiverse and productive ecosystems on the planet ( Hatcher, 1988 ; Knowlton et al., 2010 ). The process of host calcification, which builds the three dimensional structure of the reef, is largely powered by symbiont primary productivity ( Roth, 2014 ). However, climate change and other anthropogenic processes threaten reefs because of the sensitivity of the coral-dinoflagellate symbiosis to environmental stress ( Hoegh-Guldberg et al., 2007 ; Hughes et al., 2003 ), indicating that host-symbiont cooperation is not stable over ecological timescales. Recent work has suggested that a transition to parasitism may precipitate the breakdown of the host-symbiont relationship ( Baker et al., 2018 ), but this is not a unique feature of the coral symbiosis. Across taxa, mutualisms are better defined as a spectrum that ranges from negative parasitic interactions to mutually beneficial symbioses, both within the context of a focal inter-species interaction and when comparing relationships across taxa ( Doebeli & Knowlton, 1998 ; Lesser, Stat & Gates, 2013 ; Nowak, Bonhoeffer & May, 1994 ; Sachs, Essenberg & Turcotte, 2011 ). Less well-resolved, particularly for the coral- Symbiodinium symbiosis, are the mechanisms that promote and maintain positive interactions between partners ( Lesser, Stat & Gates, 2013 ; Sachs & Wilcox, 2006 ). One major factor predicted to influence levels of cooperation is the mode of symbiont transmission ( Anderson & May, 1982 ; Ebert & Bull, 2003 ). In corals as in other symbioses, two transmission modes predominate: symbionts can be acquired horizontally from the local environment, usually during a defined larval stage, or vertically from parents, typically through the maternal germ line (reviewed in Bright & Bulgheresi (2010) ). Virulence theory predicts that horizontal transmission allows symbionts to adopt selfish strategies, potentially harmful to the host ( Bull, 1994 ). A transition from horizontal to vertical transmission is predicted to align the reproductive interests of partners (via partner-fidelity feedback sensu \n Sachs et al., 2004 ) and optimize resource sharing to maximize holobiont (the combination of host and symbiont) fitness ( Ebert, 2013 ; Frank, 1994 ; Herre et al., 1999 ). Experimental manipulations of transmission mode in other systems have provided empirical support for a reduction in pathogen virulence under enforced vertical transmission scenarios ( Bull, Molineux & Rice, 1991 ; Dusi et al., 2015 ; Sachs & Wilcox, 2006 ; Stewart, Logsdon & Kelley, 2005 ). Bacteriophages forced into vertical transmission evolved lower virulence and lost the capacity to transmit horizontally ( Bull, Molineux & Rice, 1991 ). Similarly, Symbiodinium microadriaticum under an experimentally enforced horizontal transmission regime proliferated faster within their Cassiopea jellyfish hosts while reducing host reproduction and growth ( Sachs & Wilcox, 2006 ). However, studies attempting to quantify the evolutionary consequences of natural shifts in transmission mode remain rare, with Herre’s demonstration of a negative relationship between vertical transmission and virulence in nematodes that parasitize fig wasps a notable exception ( Herre, 1993 ). Reef-building corals are a potential system in which to study naturally occurring transitions in transmission mode in a mutualistic symbiosis. Corals (Cnidaria: Anthozoa: Scleractinia) are colonial animals that harbour intracellular populations of dinoflagellate algae in the genus Symbiodinium . This symbiosis is considered obligate as the breakdown of the relationship between host animals and their intracellular Symbiodinium algae, commonly known as coral bleaching, has major fitness consequences for both partners, and can be lethal (reviewed in Brown (1997) ). This inter-species partnership is ancient (evolved ∼250 Mya ( Stanley & Swart, 1995 ), prolific (600+ coral species worldwide ( Daly et al., 2007 ), and constitutes the foundation of one of the most bio-diverse ecosystems on the planet. The majority of coral species (∼85%) acquire their Symbiodinium horizontally from the local environment in each generation ( Harrison & Wallace, 1990 ). However, vertical transmission has independently evolved at least four times, such that both transmission strategies can be exhibited by different coral species within the same genus ( Baird, Guest & Willis, 2009b ; Hartmann et al., 2017 ). We compared physiological components of cooperation and fitness proxies between horizontal and vertical transmitters (VTs) in a phylogenetically controlled design using three pairs of related coral species exhibiting different strategies: (1) Galaxea acrhelia (VT) and Galaxea astreata (horizontal transmitter, HT); (2) Porites lobata (VT) and Goniopora columna (HT); (3) Montipora aequituberculata (VT) and Acropora millepora (HT). Species comparisons were drawn from the same or sister genera and replicate comparisons from more distantly related clades ( Hartmann et al., 2017 ). Species also represented a diversity of reproductive modes (e.g. broadcast spawner, brooder), sexual systems (e.g. hermaphroditic, gonochoric), and morphologies (e.g. massive, branching), and host different subclades of Symbiodinium ( Tonk et al., 2013 ), such that mode of symbiont transmission was the only consistent difference between pairs ( Baird, Guest & Willis, 2009b ; Franklin et al., 2012 ; Kerr, Baird & Hughes, 2011 ). We quantified changes in symbiont cell density, photosynthetic function and translocation of photosynthetically fixed carbon between symbionts and hosts. We defined cooperation as the proportion of photosynthetically fixed carbon translocated to the host, while the degree of symbiont parasitism was calculated as the difference in the proportion of photosynthetically fixed carbon translocated to hosts between control and heat-treated samples at the end of the experiment, sensu \n Baker et al. (2018) . Bleaching, or the reduction in symbiont density in response to sustained thermal stress was used as a proxy for holobiont fitness. While we observed differences in host-symbiont cooperation, both at a baseline level and during the onset of bleaching, vertically transmitting species did not exhibit universally elevated levels of cooperation. Additional post hoc analysis of Symbiodinium ITS-2 diversity among coral species was therefore used to investigate whether symbiont community composition could better explain physiological trait patterns. Symbiont community composition did not explain a significant portion of the variation in physiological components of cooperation or fitness proxies; however, diversity tended to be associated with the degree of symbiont parasitism at the onset of bleaching, suggesting that reduced genetic diversity of symbionts, rather than vertical transmission per se , may influence host-symbiont cooperation.",
"discussion": "Discussion Understanding variation in the degree of cooperation between corals and their Symbiodinium will be critical for assessing survival potential among species and populations in the face of increasing environmental change ( Lesser, Stat & Gates, 2013 ). As in other mutualisms ( Ebert, 2013 ; Frank, 1994 ; Herre et al., 1999 ), vertical transmission has been proposed as an evolutionary mechanism for enhancing holobiont fitness in the Cnidarian–algal symbiosis ( Putnam et al., 2012 ). However, we did not observe universally consistent differences in cooperation between vertical and horizontally transmitting species. The vertically transmitting P. lobata and Galaxea acrhelia exhibited the highest levels of carbon translocation in ambient conditions, which we interpret as symbiont-host cooperation, significantly more so than their respective horizontally transmitting counterparts, Goniopora columna and Galaxea astreata ; but cooperation was not different between M. aequituberculata and A. millepora , and tended to be higher in the latter horizontally transmitting species. However, species-specific photosynthesis-irradiance curves were not measured and the potential for interactions between photophysiology and baseline rates of carbon translocation must also be explored. We also expected that the degree of breakdown in the host-symbiont relationship under heat stress, which we interpret as a transition towards parasitism sensu ( Baker et al., 2018 ), would be comparatively intensified in horizontally transmitting species, but again, this was not the case. P. lobata did exhibit the least change in carbon translocation in spite of showing the same signs of bleaching as other species, but there was no difference in comparison to Goniopora columna which also largely maintained its baseline translocation under elevated temperature. In addition, the greatest relative declines in the proportion of carbon translocated from symbionts to hosts actually occurred in the other two vertically transmitting species, Galaxea acrhelia and M. aequituberculata . Although we did not find significant support for the role of vertical transmission we still observed significant differences in cooperation among our six focal species, both baseline differences under ambient conditions and in the degree of transition towards parasitism under elevated temperature stress. We therefore conducted a series of post hoc analyses to explore other putative drivers of differential cooperation and thermal tolerance: differences in Symbiodinium cell density and/or in symbiont community composition. The density of symbiont cells was recently proposed as a major driver underpinning the degree of cooperation between coral hosts and symbionts and the functional response of the coral symbiosis to environmental stressors ( Cunning & Baker, 2014 ). Studies in other species have shown that corals with greater initial Symbiodinium cell densities, as quantified by the symbiont to host cell ratio, are subsequently associated with greater bleaching severity in response to elevated thermal stress ( Cunning & Baker, 2013 ; Silverstein, Cunning & Baker, 2014 ). This association has been hypothesized to result from the proportional increase in oxidative cellular stress: more symbionts yield more reactive oxygen species when the photosynthetic machinery is overloaded ( Cunning & Baker, 2014 ), though the recent work of Baker et al. (2018) , adds another potential explanation. They showed that under non-limiting nutrient conditions, Symbiodinium cell division in Orbicella faveolata actually increased in response to sub-bleaching temperature exposure, but that the metabolic costs were borne by the coral hosts ( Baker et al., 2018 ), supporting the prediction that a transition to parasitism precedes unsustainable proliferation of the symbiont community which ultimately results in bleaching ( Wooldridge, 2009 ). In examining relationships between symbiont cell densities, the intensity of the bleaching response at the end of our 17-day temperature exposure and carbon translocation rates, we did observe a negative relationship between initial symbiont cell density on days 2 and 4 and subsequent bleaching response on day 17, which did not significantly differ among our six focal species. However, initial symbiont cell densities did not predict the degree of the subsequent transition to parasitism. In fact, the majority of species showed a trend in which greater initial cell densities were associated with a greater maintenance of cooperation under bleaching stress. We also found no relationship between the degree of bleaching and the degree of parasitism on day 17, nor did symbiont cell density explain variation in cooperation among species in ambient conditions. For some species, cooperation tended to decrease with increasing density of symbionts, whereas in others it increased. Taken together, these observations support the association between initial symbiont cell density and subsequent bleaching intensity, but disagree with the proposed role of an alteration of host-symbiont cooperation in mediating the bleaching response. The discrepancy in results among studies may be due to the importance of nutrient enrichment for observing a parasitic increase in symbiont communities, or in the difference in study duration and sampling design. Corals were exposed to a constant light environment, which likely did not provide a saturating irradiance across species. Consequently, alteration of carbon translocation as a function of species-specific photophysiology may explain baseline differences in ambient conditions, and future work should aim to test this hypothesis. In addition, while corals received supplemental feeding throughout the duration of this experiment, which likely introduced organic nitrogen, we did not explicitly manipulate inorganic nitrogen levels. It is also possible that variation among hosts in their ability to limit their symbionts’ nitrogen supply may have influenced the observed variation in the degree of parasitism ( Cunning et al., 2017 ; Wooldridge, 2009 ), although additional studies are needed to test this. In addition, Baker et al. (2018) exposed corals to a +5 °C temperature ramp over 8 days, sampling once 24 h after the final temperature was reached, whereas the present study increased temperatures by +4 °C over 5 days and maintained that difference for an additional 12 days, sampling at three time points, comparatively earlier and later. We did not observe an initial increase in symbiont cell density or decrease in carbon translocation on days 2 and 4 under elevated temperature, the experimental time-frame most analogous to that of Baker et al. (2018) . It is possible that these dynamics occurred during a time frame in which we did not sample; however, our final results argue against this explanation. On day 17, we observed symptoms of bleaching that did not differ across species: symbiont cell densities and rates of net photosynthesis were uniformly decreased. However, the transition to parasitism was not uniform, as some species exhibited significant differences in carbon translocation in response to heat stress whereas others did not. We therefore conclude that while symbiont density alone may be a reasonable predictor of the potential for observing a bleaching response under elevated temperature, it does not predict cooperative dynamics that likely also influence holobiont fitness. We also explored the role of symbiont community diversity on bleaching stress and cooperation. Predictions regarding the cooperative and fitness benefits of evolving vertical transmission are based on the assumption that the host-symbiont relationship becomes exclusive: symbiont population sizes are substantially reduced, resulting in genetic uniformity, more rapid co-evolution of partner traits and reduction in intra-symbiont community competition ( Herre et al., 1999 ; Maynard Smith & Szathmary, 1995 ). In this case, it may not be vertical transmission per se that influences host-symbiont cooperation, but the relative diversity of the symbiont community. A prior meta-analysis found that symbiont specificity was correlated with transmission mode, with horizontally transmitting species being more likely to interact with generalist symbionts ( Fabina et al., 2012 ). However, the relationship between transmission mode and overall community diversity was not explored. Other more recent studies have also shown the potential for ontogenetic shifts in Symbiodinium community composition of putative VTs, potentially indicating the capacity for mixed-mode or cryptic horizontal transmission ( Byler et al., 2013 ; Reich, Robertson & Goodbody-Gringley, 2017 ). Our results show that symbiont diversity does not partition by transmission mode. While communities in the vertically transmitting M. aequituberculata and P. lobata were largely uniform, consisting predominantly of C35 and C15-type sequence variants, respectively, Symbiodinium community diversity was more than twice as high in Galaxea acrhelia in comparison to Galaxea astreata (1/D = 3.6 vs. 1.7). In addition, community composition in the horizontally transmitting Goniopora columna was also highly uniform, second only to that of P. lobata (1/D = 1.08 and 1.00, respectively). In exploring the relationship between symbiont diversity at the ITS2 locus and metrics of host-symbiont cooperation and bleaching independent of transmission mode, we did not find any formally significant correlations, likely due to the fact that our symbiont community analysis was limited to the level of differences among the six species, greatly reducing our statistical power. However, there was a weak negative relationship between community diversity and the degree of parasitism under thermal stress. Species with the most genetically uniform symbiont communities, P. lobata and Goniopora columna , maintained the highest levels of cooperation in spite of showing signs of bleaching. Yet there was no relationship between community diversity and baseline differences in cooperation among species under ambient conditions, as high rates of translocation were observed in species with both the lowest ( P. lobata ) and highest community diversity ( A. millepora, Galaxea acrhelia ), though again, species-specific interactions between photophysiology and carbon translocation remain to be explored. The presence of particular symbiont types has also been shown to influence holobiont fitness and carbon translocation. For example, conspecific corals hosting clade D Symbiodinium exhibit greater thermal tolerance than those hosting C1 or C2-types ( Berkelmans & Van Oppen, 2006 ; Jones et al., 2008 ), however, they generally grow more slowly under non-stressful conditions ( Jones & Berkelmans, 2010 ; Little, Van Oppen & Willis, 2004 ) and receive less photosynthetically fixed carbon from their symbionts ( Cantin et al., 2009 ). We found no significant relationships between the proportional abundance of Clade D-type symbionts and metrics of host-symbiont cooperation or bleaching. Most species had no, or a low proportion of Clade D, but the species with the greatest proportion of Clade D symbionts ( Galaxea acrhelia ) exhibited both the highest carbon translocation under ambient conditions and the greatest transition to parasitism under elevated temperature stress. Similar to our observations regarding symbiont cell density, these results support prior observations that Symbiodinium community composition alone is not sufficient to explain variation in holobiont performance ( Abrego et al., 2008 ; Baird et al., 2009a ; Kenkel et al., 2013 ). However, we reiterate that our analyses are limited in comparing only averages among species. There was some variation observed in dominant symbiont types among individual coral colonies within species ( Fig. S6 ) and a priority for future study should be to investigate whether these conclusions hold when considering intraspecific variation in symbiont community composition in addition to these broader interspecific differences. Quantifying cooperation between symbiotic partners in terms of biologically realistic costs and benefits remains an outstanding question for many symbioses ( Herre et al., 1999 ). The transfer of photosynthetically fixed carbon has long been known as a major cooperative benefit to the coral host as up to 95% of a coral’s energy requirements can be met through this mechanism ( Muscatine, 1990 ); however, reciprocal products shared by hosts with their symbionts remain largely unknown ( Yellowlees, Rees & Leggat, 2008 ). Similarly, heterotrophic feeding can offset the need for symbiont-derived carbon in some species and in these cases other symbiont-derived metabolites may be more critical for host fitness ( Grottoli, Rodrigues & Palardy, 2006 ). Substantial variation in both intra- and inter-specific bleaching thresholds ( Marshall & Baird, 2000 ), suggests that levels of cooperation between host and symbiont may also vary. Significant work has gone into investigating coral bleaching over the past three decades, yet fundamental questions remain unresolved ( Edmunds & Gates, 2003 ). Ultimately, a greater understanding of both fine-scale interactions between coral hosts and symbionts and the evolutionary and ecological mechanisms that maintain and strengthen cooperation will be essential for managing these ecosystems ( Davy, Allemand & Weis, 2012 ; Lesser, Stat & Gates, 2013 )."
} | 6,063 |
25642161 | PMC4295533 | pmc | 33 | {
"abstract": "Resistive (or memristive) switching devices based on metal oxides find applications in memory, logic and neuromorphic computing systems. Their small area, low power operation, and high functionality meet the challenges of brain-inspired computing aiming at achieving a huge density of active connections (synapses) with low operation power. This work presents a new artificial synapse scheme, consisting of a memristive switch connected to 2 transistors responsible for gating the communication and learning operations. Spike timing dependent plasticity (STDP) is achieved through appropriate shaping of the pre-synaptic and the post synaptic spikes. Experiments with integrated artificial synapses demonstrate STDP with stochastic behavior due to (i) the natural variability of set/reset processes in the nanoscale switch, and (ii) the different response of the switch to a given stimulus depending on the initial state. Experimental results are confirmed by model-based simulations of the memristive switching. Finally, system-level simulations of a 2-layer neural network and a simplified STDP model show random learning and recognition of patterns.",
"introduction": "Introduction Brain-inspired computing is among the top challenges of the today's information and communication technology. The brain is capable of formidable tasks, such as learning, recognition of visual/auditory patterns, and adaptation in response to new information. To meet this grand challenge, a neuromorphic system should include a number of neurons and synapses similar to a biological human brain, featuring around 10 12 neurons and 10 15 synapses (Rajendran et al., 2013 ). Clearly, such a complex system can be realized only through advanced manufacturing techniques (e.g., 3D integration), and small circuit blocks for neurons and synapses. The latter, in particular, represents by far the largest area of the neuromorphic circuit due to the huge number of inter-neural connections, therefore scaling down the size and complexity of the artificial synapse is a key task in the design of a neuromorphic circuit. To this purpose, nanoscale resistive switches, or memristors, have been proposed as novel artificial synapses in neuromorphic systems (Likharev et al., 2003 ; Snider, 2008 ; Jo et al., 2010 ). Memristors have the capability of an inherent analog tuning, combined with a 2-terminal structure and a scalable device area and power, therefore they display strong advantage with respect to silicon-based synapses, such as floating gate memories (Diorio et al., 1996 ) and static RAM (Indiveri et al., 2006 ). Different switch technologies have been proposed for artificial synapses, including phase change memories (Wright et al., 2011 ; Bichler et al., 2012 ; Kuzum et al., 2012 ), organic-based switches (Bichler et al., 2010 ), chalcogenide-based switches (Ohno et al., 2011 ; Suri et al., 2013 ) and oxide-based resistive switching memories (Seo et al., 2011 ; Yu et al., 2011 , 2013 ; Park et al., 2012 ; Ambrogio et al., 2013 ). The latter approach provides analog switching, nonvolatile behavior, CMOS compatible materials, back-end process and scalable power consumption thanks to filamentary switching (Wong et al., 2012 ). A memristor naturally satisfies the requisites for electrically-tunable conductance, serving as a connection for communication between a pre-synaptic neuron (PRE) and a post-synaptic neuron (POST), and responsive to the individual spikes fed from both neurons. To achieve this multitask operation, a time-division multiplexing (TDM) approach was previously proposed, where neuron spikes obey a precise synchronous sequence for communication, long-term potentiation (LTP) and long-term depression (LTD) (Snider, 2008 ; Jo et al., 2010 ). The synchronous approach, however, may be too idealized with respect to the biological brain, where synapses are potentiated/depressed through asynchronous spike timing dependent plasticity (STDP) (Bi and Poo, 1998 ). Also, synchronous clocking may be practically difficult in the case of large neuromorphic systems (Zamarreño-Ramos et al., 2011 ). More recently, a fully asynchronous approach for communication/learning of neuromorphic synapses with leaky-integrate-and-fire (LIF) neurons was proposed (Zamarreño-Ramos et al., 2011 ; Serrano-Gotarredona et al., 2013 ). However, a conceptual demonstration of realistic memristor synapses for communication and learning has not been achieved so far. This work addresses the integration of memristors in neuromorphic systems by introducing a 2-transistor/1-resistor (2T1R) synapse for large scale neuromorphic systems. The transistors in the synapse block allow for (i) multiple-input control of the synapse, which must receive signals from both the PRE and the POST, and (ii) accurate control of the filament growth for analog switching and STDP behavior (Yu et al., 2011 ; Ambrogio et al., 2013 ; Subramaniam et al., 2013 ). STDP in the 2T1R synapse is experimentally demonstrated on bipolar resistive switching memories based on HfO 2 acting as memristive switches. We show that the memristive synapse is capable of communication of spiking signals between neurons and stochastic STDP due to both the natural switching variability in the switch, and to the variations of memristive response depending on the initial state. We finally show a conceptual demonstration of a simulated 2-layer neuromorphic network displaying stochastic pattern learning and recognition, thus further supporting memristive synapse as a scalable, high-functionality building blocks for large scale neuromorphic systems.",
"discussion": "Discussion The proposed synapse circuit allows for asynchronous transmission and plasticity controlled by the spiking delay between the pre- and post-synaptic neurons. The synapse circuit adheres to the conventional organization of the neural network, where integrate-and-fire neurons serve as both input and output of the communication and plasticity. In particular, the BE terminal, being connected to the virtual ground input of the neuron, serves as reference ground for the synapse circuit, while pulses of arbitrary voltage are applied to the other 3 terminals, namely TE, CG and FG. This is different from previous approaches, where the pre-synaptic pulse (spike) and the post-synaptic pulse (fire) where applied to the TE and BE, respectively, of the resistive synapse (Yu et al., 2011 ; Indiveri et al., 2013 ). It is also different from other approaches employing 1T1R structures, where STDP relied on a dynamic V T behavior of the transistor, achieved through nanoparticle-containing gate dielectric (Subramaniam et al., 2013 ). In fact, only standard transistor CMOS transistor are needed in the 2T1R synapse in this work. The transistors in the 2T1R structure are functional in achieving 2 necessary behaviors of the synapse array, namely STDP and communication. On the one hand, the FG transistor allows for a spike timing comparison between two pulses, namely the TE pulse from the pre-synaptic neuron and the FG pulse from the post-synaptic neuron (Ambrogio et al., 2013 ). Therefore, the FG transistor is functional to plasticity. On the other hand, the CG transistor allows for enabling communication from pre-synaptic neuron to post-synaptic neuron in the neural network. If there was no CG transistor, the TE pulse might affect the weight of the synapse even without any fire from the post-synaptic neuron. Note in fact that the CG voltage is high only during the initial part of the TE pulse, at relatively low voltage. Therefore, this transistor is functional to communication, while protecting the memristor from the rather large TE voltage used for plasticity. In addition, transistors allow to limit the current flowing in the memristive switch during the set transition, thus preventing uncontrolled switching and even irreversible breakdown of the device. These latter events may result in excessive power consumption due to low resistance value in the synapse, and/or in the impossibility to reset the memristor because of excessive growth of the conductive channel. Current limitation can be achieved by biasing the transistor in the saturated regime at relatively low gate voltage, which ensures that the maximum current after set transition is limited. Finally, the transistor serves as selector in the synapse array of Figure 9 , which otherwise would be plagued by significant sneak-path currents (Baek et al., 2005 ). Note that other types of selectors would allow better scalability of the array, e.g., p-n diodes (Baek et al., 2005 ), or threshold switch devices (Cha et al., 2013 ), thanks to the 2-terminal structure. However, 2 terminals would not be sufficient for the local comparison of spike timing which is needed for synapse plasticity control. It has been pointed out that the necessity to generate dedicated waveforms within the neuron circuit might lead to an excessive circuit overhead, thus conflicting with the need for very large scale arrays with high synaptic densities (Kornijcuk et al., 2014 ). Note however that the generator of the spike belongs to the neuron circuit, thus a complex waveform should not affect the density of synapses. Also, note that the waveforms in Figures 1 , 2 have been designed to achieve a bio-realistic STDP as shown in Figure 4 . Other waveforms and STDP characteristics can be used with no impact on the pattern recognition capability, while strongly alleviating the burden on the neuron circuit. This is demonstrated in Figure 12 , showing the square waveforms for V TE and V FG (a) and the corresponding statistical STDP characteristic (b) obtained from 7.5 × 10 4 random spikes. Note that the STDP characteristics reflects the simple shape of the spike and fire pulses, while we demonstrated that the pattern learning behavior is not affected. This further demonstrates the strength of the STDP process and the flexibility of our 2T1R circuit in realizing LTP and LTD with a variety of spike shapes. Note that pulse widths of the neuron spikes in the range of 100 ms, which are needed to achieve real-time bio-compatible neuromorphic behavior (Indiveri et al., 2011 ), do not necessarily require large capacitors. In fact, time responses in the 100 ms range are straightforwardly achieved in neuromorphic circuits through relatively small capacitances (e.g., 1 pF) charged/discharged by extremely low current in MOS transistors biased in the subthreshold regime (Mitra et al., 2009 ). Figure 12 Square-pulse STDP . The use of square pulses for V TE and V FG \n (A) allows to achieve STDP with square characteristics suitable to learning and recognition (B) , while requiring a simple circuit for pulse generation. Low-power operation is a fundamental property of neuromorphic circuits. The energy consumption of our 2T1R synapse for communication can be estimated to about 150 nJ from the voltage waveform in Figure 1B assuming I = 50 μA. Assuming an average spike frequency of 1 Hz, the power consumption for communication should be around 150 nW. This value can be reduced by decreasing the pulse-width of the V CG pulse and the current during communication. On the other hand, the energy consumption is slightly larger due to the larger voltage and current needed for resistive switching. For instance, the LTP energy is around 400 nJ for a current of 170 μA and a V TE of 2.4 V in correspondence of the positive peak. However, since the LTP frequency is expected to be smaller than the spiking frequency, the power consumption for LTP might be in the same range as the communication power. Similar to the communication case, LTP power can be reduced by properly decreasing the current (e.g., by up to a factor 10) and the pulse width (up to a factor 10 3 ). This allows for memristive-based synapses with relatively low power consumption. Other switching concepts might be used in alternative to oxide memristors, e.g., spin-transfer-torque (STT) elements (Locatelli et al., 2014 ) or phase change memory (PCM) elements (Kuzum et al., 2012 ; Eryilmaz et al., 2014 ). However, oxide memristors allows for a smaller power consumption since the switching channel area can be controlled through the transistor current during the set transition, whereas the switching current is controlled by the lithography-defined area of the device in both STT and PCM devices, which thus can hardly be reduced below 50 μA (Ielmini and Lacaita, 2011 ; Kim et al., 2011 ). The use of a HfO 2 memristor allows for CMOS compatible process in the back-end, however other metal oxides can be used in principle for the active switching layer, such as TaO x (Lee et al., 2011 ). A careful material engineering is needed to identify the best material properties for synaptic functionality, including, e.g., controllability of the synapse weight, stochastic switching and low power operation. Conflict of interest statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest."
} | 3,278 |
37687484 | PMC10488364 | pmc | 34 | {
"abstract": "Superhydrophobic coating ice suppression is an advanced and durable technology that shows great potential for application on pavements. Although many researchers have conducted experimental and theoretical validations to confirm the effectiveness of superhydrophobic surfaces in actively suppressing ice formation, there are still some who remain skeptical. They argue that the roughness of the surface may increase ice adhesion due to the mechanical interlocking effect of condensation droplets in low-temperature and high-humidity environments. In this study, we present a comprehensive investigation of a novel superhydrophobic coating specifically designed for pavement surfaces, aiming to address the question of its active anti-icing/ice-sparing capabilities in a condensing environment. The changes in contact angle before and after condensation for four material surfaces with varying wettability were investigated, as well as the morphology and ice adhesion of liquid water after it freezes on the material surface. The findings reveal that the proposed superhydrophobic coating for pavements effectively prevents condensate droplets from infiltrating the surface structure, resulting in delaying the surface icing time and reducing the attachment strength of the ice.",
"conclusion": "5. Conclusions Applications of superhydrophobic coatings as anti-icing materials for pavements can effectively delay the icing of static water before icing and reduce the adhesion of ice to the pavement surface after icing due to their unique micro- and nano-structures and excellent anti-condensation properties. Through both theory and experimentation, the study yielded important conclusions about the mechanisms behind these coatings: (1) Through theoretical discussions, we have uncovered the ice suppression capabilities of the superhydrophobic surface in relation to the condensate. It can be concluded that when the nano-gap of the superhydrophobic surface texture is lower than the critical nuclear radius of 145 nm under low humidity and high moisture environment, it prevents condensate droplets from entering the surface texture, allowing liquid water to remain in the Cassie state on the superhydrophobic surface. This inhibits the formation of mechanical interlocking effects, which can increase ice adhesion on the surface. (2) The surface of the superhydrophobic coating is composed of micro- and nano-papillae. Even in conditions of low temperatures and high humidity, condensation on the superhydrophobic surface does not affect its water contact angle, which is 151.9° with little change compared to the pre-condensation angle of 154.0°, ensuring continued strong superhydrophobicity. The large contact angle conceals the air in the gap of the coarse structure, which delays the icing time of the droplets and reduces the contact surface of the droplet with the superhydrophobic coating. (3) After investigating the adhesion of ice on a variety of wettability surfaces, our findings revealed that the adhesion force of hydrophilic surfaces with ice is the most robust, up to 947.75 N. In contrast, it can be observed that the presence of air within the microstructure of superhydrophobic surfaces significantly reduces the ice adhesion, with an adhesion of 214 N, without any observable mechanical interlocking effect.",
"introduction": "1. Introduction The accumulation of snow and ice on the pavement surface in winter leads to a significant reduction in the skid resistance of the pavement surface, seriously weakening road capacity and posing a grave threat to vehicle safety, potentially leading to severe traffic accidents [ 1 , 2 ]. This threat is particularly prominent in southern China, where freezing rain and black ice form in high-altitude, cold, and wet areas [ 3 , 4 ], further jeopardizing road safety. Therefore, addressing the issue of snow and ice on pavements during winter is crucial to ensure traffic safety. Currently, two main types of anti-icing technologies are employed for pavements, namely active de-icing technology and passive de-icing technology. Passive de-icing technology primarily relies on physical or chemical methods such as manual mechanical removal and the application of snow-melting agents. Although this approach is relatively efficient, it still has the potential to damage the pavement surface and bring about pollution to the surrounding ecological environment. What is more, the costs associated with passive de-icing methods are higher [ 5 , 6 ]. In contrast, active de-icing methods are more efficient and less damaging to pavement structures. Examples include self-heating pavements, pavements containing anti-freeze fillers, phase change energy storage pavements, carbon fiber conductive pavements, and emerging superhydrophobic materials [ 7 , 8 , 9 , 10 , 11 ]. However, the currently available active de-icing methods in the market are often not affordable. Notably, the utilization of superhydrophobic materials, serving as a potential active de-icing solution, has garnered increasing attention due to its simple nature of preparation, cost-effectiveness, and, more importantly, free from environmental contamination [ 12 , 13 ]. There are several studies [ 14 , 15 ] that have investigated ice adhesion mechanisms on superhydrophobic surfaces. In 2003, Pilotek [ 16 ] first proposed that superhydrophobic coatings might exhibit low ice adhesion. It promotes the application of superhydrophobic surfaces in the field of anti-icing coverage. Some researchers [ 17 , 18 ] have compared ice adhesion for different surface structures and concluded that nanostructured surfaces with relatively lower roughness exhibit the lowest ice adhesion. Subramanyam’s group at MIT, USA [ 19 ] fabricated four different surface structures, including smooth, micron, nano, and micro/nanocomposite surfaces, modified with surface energy substances. Their analysis revealed that the nanostructured and micro/nanocomposite surfaces exhibited excellent resistance to frost formation, with the nanostructured samples displaying the lowest ice adhesion. This finding has provided significant inspiration. Huang’s non-uniform nucleation ice crystal growth model experimentally verifies the ice inhibition mechanism of superhydrophobic materials. The results show that the prepared superhydrophobic coating material exhibited excellent properties, including a high contact angle (>150°), good anti-slip nature, and the ability to retard water droplet crystallization and maintain the droplet shape even after freezing [ 20 ]. These findings highlight that superhydrophobic surfaces achieve extremely low surface energy due to the presence of micro-nano rough structures. Also, the grooves on the substrate surface are filled with air, resulting in a Cassie state for droplets, where only about 10% of the total contact area is occupied by the droplet and substrate contact [ 21 ]. This, combined with the surface’s exceptionally low rolling angle, hinders the infiltration of droplets into pavement material upon impact, effectively reducing the icy surface area of the pavement. This, in turn, reduces the area of ice in contact with the pavement. The adhesion between the ice and the pavement material is reduced. The surface affects the wettability of superhydrophobic coatings in terms of ice adhesion remains controversial. Some researchers argue [ 22 , 23 , 24 , 25 ] that ice adhesion is linked to surface wettability and studied the relationship between surface wettability and ice adhesion force. Their findings indicate that ice adhesion force is proportional to smaller hysteresis angles, resulting in lower adhesion forces. While some researchers hold the opposite opinion [ 26 , 27 , 28 , 29 , 30 ], for instance, Chen [ 31 ] prepared 13 kinds of silicon wafers with superhydrophilic or superhydrophobic wettability to study surface morphology and adhesion. It was found that ice adhesion on superhydrophilic and superhydrophobic surfaces was similar. Chen’s research concluded that superhydrophobic coatings not only fail to reduce ice-cover bond strength but also exacerbate subsequent de-icing efforts due to their surface structure and mechanical interlocking effect between ice and the coating. Varanasi [ 32 ] investigated the application of superhydrophobic surfaces for de-icing and revealed that rough superhydrophobic surfaces in high humidity environments are susceptible to frost formation, therefore resulting in greater ice adhesion compared to smooth surfaces. These results raise questions about the effectiveness of superhydrophobic surfaces for de-icing applications. Actually, superhydrophobic surfaces are frequently exposed to low-temperature and high-humidity environments, where gas–liquid phase change condensation or fog coalescence, commonly observed in nature, inevitably occurs on solid surfaces [ 32 ]. As the supercooling increases, water vapor condenses and nucleates within the micro- and nano-structures of the superhydrophobic surface and grows gradually. Failure to expel the condensed droplets from the surface irreversibly transforms the surface into a partially or completely wetted Wenzel state, adhering to the surface [ 33 , 34 , 35 ]. At this stage, when external droplets come into contact with the condensed surface, they fuse with the droplets present on the surface, causing a reduction in the apparent contact angle and altering the surface wettability. Consequently, the superhydrophobic ice-suppressing performance also diminishes. Therefore, to effectively utilize superhydrophobic ice-suppressing coatings on pavements, theoretical discussions and feasibility studies regarding the anti-condensation performance of superhydrophobic materials are valuable, ensuring the desired active anti-icing/ice-sparing effects. This paper employs the “self-migration” movement of condensate droplets on superhydrophobic coatings to theoretically analyze factors influencing the anti-condensation performance of the coating. We characterize four material surfaces with varying wettability to investigate changes in the contact angle of different material surfaces before and after condensation and the morphology and ice adhesion of liquid water after freezing on the material surface to support the proposed theory. Condensed droplets do not affect the water contact angle of the superhydrophobic surface, and the superhydrophobic surface can delay the icing time of the droplets, as the droplets have a smaller contact area and adhesion force with the material surface after icing. The superhydrophobic coating has good anti-condensation properties and broad application prospects in active ice suppression on pavements.",
"discussion": "4. Results and Discussion 4.1. Microscopy of Superhydrophobic Materials The micromorphology of the superhydrophobic material was examined using FESEM, and the results are presented in Figure 5 . During the hydrothermal reaction, SiO 2 particles were observed to encapsulate the surface of TiO 2 particles. The nano-TiO 2 reacted with stearic acid, leading to a rough morphology of the nanostructures and the formation of a large number of papillae. The lotus leaf is a natural superhydrophobic surface due to its special micro-nanostructure and self-cleaning characteristics. It can be observed in the electron micrograph that the superhydrophobic surface has the same micronano papillary structure as the lotus leaf surface. Thus, the field emission scanning electron microscopy (FESEM, provided of Shanghai Sinu Optical Technology Co., Ltd., Shanghai, China) analysis confirmed the presence of superhydrophobic materials with low surface energy. 4.2. Effect of Condensation on the Water Contact Angle of Materials The contact angle test was performed to evaluate the wettability of water droplets on the surface of the superhydrophobic material. Multiple measurements were taken at different locations on the coated sample. A surface material can be classified as superhydrophobic when the contact angle of water droplets exceeds 150°. The contact angle results are summarized in Table 2 . The results show that the water drop contact angle on the prepared superhydrophobic coating 2 was 154.0° (>150°) at room temperature and 151.9° (>150°) after the condensation phenomenon. These findings demonstrate that the coating material still has a large water contact angle at low temperatures and high humidity, conforming to the characteristics of wettability as superhydrophobicity, thereby confirming its robustness in maintaining superhydrophobicity. 4.3. Water Droplet Icing on Different Wettability Surfaces Following the freezing test procedure described in Section 3.2.3 , the samples were placed in a temperature control box set to −4 °C. After 20 min, the droplets on the surface of the normal sample were observed to freeze, while those on the hydrophobic coating exhibited a mixed state of ice and water. Remarkably, the droplets on the superhydrophobic coating remained in a liquid state. As time progressed, after 35 min, the droplets on the superhydrophobic coating eventually froze. This indicates that at a temperature of −4 °C, the solidification time of droplets on superhydrophobic surfaces was significantly longer compared to the ordinary specimen surface, thus demonstrating the ability of the superhydrophobic coating to extend the solidification time of droplets. Next, the temperature of the high-temperature control box was changed to −8 °C, and another set of comparison specimens was placed inside. It was observed that the droplets on the surface of the normal specimen started to freeze after about 8 min, while the water droplets on the hydrophobic coating surface began freezing after about 11 min. In contrast, the water droplets on the superhydrophobic coating surface exhibited freezing after approximately 14 min. Finally, a final group of specimens was placed in the high and low-temperature control box set to −10 °C to observe the time required for water droplets to freeze. The results indicated that the droplets on the regular specimen surface started freezing after about 3 min, the droplets on the hydrophobic coating surface started freezing after approximately 6 min, and the droplets on the superhydrophobic coating surface started icing after approximately 8 min. Figure 6 shows the photos of water droplet states on the regular specimen surface and the superhydrophobic-coated specimen at the same moment. Among them, Figure 6 A–C represent the photos of droplets freezing on the surface of the regular specimen, Figure 6 D–F represent the photos of drops of water on the surface of the hydrophobic coating specimen, and Figure 6 G–I represent the photos of droplets on the surface of the superhydrophobic coating specimen. Analyzing the figures, under −4 °C conditions, Figure 6 A shows the state when the droplets on the regular specimen surface start to freeze, with small scattered ice crystals already present. Figure 6 D represents the hydrophobic coating specimen at the same moment, while Figure 6 G depicts the superhydrophobic coating specimen, where the surface water droplets remained in a liquid state. Under −8 °C and −10 °C conditions, Figure 6 B,E,H represent the state after freezing at the same moment, while Figure 6 C,F,I represent the state after icing at the same moment. It can be observed that the icing time of Figure 6 B,C is shorter, while the icing time of Figure 6 E,F,H,I is longer. Additionally, the ice droplet contact area with the wall of the dish after icing was larger in both regular specimens and hydrophobic coating specimens compared to the superhydrophobic coating. The water droplets in Figure 6 G–I gradually increased in size as the temperature decreased, suggesting nucleation of water vapor condensing on the large droplets in a low-temperature and high-humidity environment. 4.4. The Law of Ice Adhesion Cover on Different Surfaces To evaluate the ice cover adhesion of the new pavement superhydrophobic coating after icing, specimens of hydrophilic coating, hydrophobic coating, superhydrophobic coating modified with anatase titanium dioxide nanoparticles, and superhydrophobic coating modified with rutile titanium dioxide nanoparticles were prepared, corresponding to No. 1, No. 2, No. 3, and No. 4, as shown in Figure 7 a. These specimens were placed in molds, as depicted in Figure 7 b, inside high- and low-temperature control boxes set to −10 °C, allowing the water to condense into rectangular iced concrete specimens of dimensions 200 mm × 100 mm × 100 mm. The iced concrete specimens were positioned on a press, and two circular mat strips were inserted at both ends of the interface between the specimens and the ice. Pressure was applied until splitting occurred, and the experimental results are presented in Figure 8 , illustrating the ice adhesion to the four specimens. The flat hydrophilic surface exhibited the highest ice adhesion, measuring 947.75 N. This is due to the smoothness of the hydrophilic surface, which results in a larger actual contact area between the ice surface and the specimen. The ice adhesion of the hydrophobic surface and the superhydrophobic surface modified with rutile titanium dioxide was below that of the smooth hydrophilic surface, measuring 252.75 N and 237.75 N, respectively. Remarkably, the ice adhesion of superhydrophobic surfaces modified with rutile titanium dioxide was as low as 214 N. This can be attributed to the enhanced stability and superior weather resistance of rutile titanium dioxide compared to anatase titanium dioxide nanoparticles. Consequently, the impact of environmental changes on the superhydrophobic surface was minimized. For superhydrophobic surfaces, at room temperature, liquid water resided above the surface texture, known as the Cassie state, while the air trapped in the surface structure beneath the liquid water remained in thermodynamic equilibrium with it. As the temperature decreased, the surface liquid water gradually began to freeze, and the condensed droplets of air did not freeze within the superhydrophobic surface texture where the nano-gap fell below the critical nuclear radius. Instead, they condensed and nucleated on the surface of larger droplets. At this stage, the liquid water existed in a state of coexistence between Wenzel and Cassie states, as shown in Figure 9 . As a result, the actual contact area between the ice surface and the sample decreased, resulting in a reduction in the ice adhesion."
} | 4,626 |
37793161 | PMC10592315 | pmc | 35 | {
"abstract": "The superhydrophobic properties of material surfaces\nhave attracted\nsignificant research and practical development in a wide range of\napplications. In the present study, a superhydrophobic coating was\nfabricated using a vapor-phase sublimation and deposition process.\nThis process offers several advantages, including a controllable and\ntunable superhydrophobic property, a dry and solvent-free process\nthat uses well-defined water/ice templates during fabrication, and\na coating technology that is applicable to various substrates, regardless\nof their dimensions or complex geometric configurations. The fabrication\nprocess exploits time-dependent condensation to produce ice templates\nwith a controlled surface morphology and roughness. The templates\nare sacrificed via vapor sublimation, which results in mass transfer\nof water vapor out of the system. A second vapor source of a polymer\nprecursor is then introduced to the system, and deposition occurs\nupon polymerization on the iced templates, replicating the same topologies\nfrom the iced templates. The continuation of the co-current sublimation\nand deposition processes finally renders permanent hierarchical structures\nof the polymer coatings that combine the native hydrophobic property\nof the polymer and the structured property by the sacrificed ice templates,\nachieving a level of superhydrophobicity that is tunable from 90°\nto 164°. The experiments demonstrated the use of [2,2]paracyclophanes\nas the starting materials for forming the superhydrophobic coatings\nof poly( p -xylylenes) on substrate surfaces. In comparison\nto conventional vapor deposition of poly( p -xylylenes),\nwhich resulted in dense thin-film coatings with only a moderate water\ncontact angle of approximately 90°, the reported superhydrophobic\ncoatings and fabrication process can achieve a high water contact\nangle of 164°. Demonstrations furthermore revealed that the proposed\ncoatings are durable while maintaining superhydrophobicity on various\nsubstrates, including an intraocular lens and a cardiovascular stent,\neven against harsh treatment conditions and varied solution compositions\nused on the substrates.",
"conclusion": "Conclusion The introduced superhydrophobic coating\ntechnology is versatile\nand can be applied to a wide range of substrate conditions, from two-dimensional\n(2D) to 3D and curved surfaces. It is also applicable regardless of\nthe solution compositions. The technology is based on a vapor-phase\nsublimation and deposition process, which offers the advantages of\nbeing a dry and solvent-free process as well as providing conformal\nand precise coating fidelity on complex dimensions and geometries\nof substrates. The process can theoretically be applied to a wide\nrange of substrate materials. In this study, we used an ice template\nfor sublimation and poly( p -xylylene) (parylene, U.S.\nPharmacopeia Class VI as highly biocompatible) for deposition. These\nmaterials agreed well and are compatible with applications in sensitive\nand delicate substrate conditions. The fabricated coatings exhibited\na WCA of 164° compared to the conventional poly( p -xylylene) coating of approximately 90°, demonstrating superior\nhydrophobicity. The coatings also showed high durability, maintaining\nabove 150° after exposure to harsh conditions. The proposed fabrication\nprocess and coating technology have significant potential for unlimited\napplication developments in the fields of energy materials, aerospace\nand defense materials, and biomaterials that require stringent and\ntunable interface properties.",
"introduction": "Introduction Since the mechanism behind the non-wetting\nproperties of lotus\nleaves in the 1970s was discovered through the use of scanning electron\nmicroscopy, it began attracting more and more attention and interest\ntoward replication of such an appealing property for application in\nvarious fields. 1 To reach the non-wetting\nor superhydrophobic state, the water contact angle (WCA) must be greater\nthan 150° to enable the droplets to roll easily on surfaces.\nThe hierarchical micro- and nanoscale structures of the surfaces have\nbeen found to be the primary reason for superhydrophobicity, as seen\nin lotus leaves. 2 , 3 Over the last few decades, a variety\nof techniques have been studied and investigated to produce superhydrophobic\nsurfaces for different applications, including electrospinning, 4 phase separation, 5 emulsion, 6 plasma method, 7 template method, 8 solution immersion, 9 chemical vapor deposition (CVD), 10 sol–gel processing, 11 lithography, 12 spray coating, 13 three-dimensional (3D) printing, 14 etc. Among these methods, CVD possesses several\nsuperior features over other techniques, including an extremely uniform\nand conformal coating result, high purity, fine controllability, and\nsolvent-free processing, making it suitable for a wide range of material-processing\napplications, especially candidating a promising treatment for miniaturized\nand sensitive materials/devices. 15 Even\nwith those superior features of the CVD method, however, the previous\nstudies are focusing on applying the CVD method to a pre-made rough\ntemplate by means of decreasing its surface energy to create a superhydrophobic\nsurface, while the pre-made rough template that was not created by\nthe vapor-phase process made the advantages of the CVD process disappear\nor be limited. 16 Recently, some approaches\nhave been developed to create rough templates via the vapor deposition\nmethod; however, the requirement of specific materials (e.g., silane\nfor SiO 2 nanoparticles and C 2 H 2 for\ncarbon nanofiber or carbon nanotube) and harsh conditions (e.g., high\ntemperature and toxic chemicals) for creating roughness through the\nvapor-phase process limits the availability of various substrate materials. 17 Although some template-free approaches using\ninitiated chemical vapor deposition (iCVD) were demonstrated to overcome\nthese issues, 18 the requirement of using\nan initiator has a potential hazard to the environment. Herein,\nwe propose a simple and versatile fabrication process to\nproduce superhydrophobic coatings based on the vapor sublimation and\ndeposition polymerization process. Unlike the conventional ice templating\ntechnique that was mainly utilized to fabricate porous materials, 19 the process exploits self-sacrificing ice/water\ntemplates, which feature the control of the time-dependent ice condensation\nand crystallization structure on the substrate surface that provides\na native and tunable morphology from the surface and from bottom up.\nSubsequently, a transformation of the iced templates as well as the\nconstructed templated morphology occurs and resulted in a final and\npermanent polymeric morphology during the fabrication process, which\nincludes the following: (i) a sublimation of water/ice vapor escapes\nout of the coating system, at the same thermodynamic conditions; (ii)\na second vapor of monomers is then introduced to the system, in which\nan opposite physiochemical behavior of deposition upon polymerization\noccurs on the iced templates, and the resultant polymeric structures\nreplicate the same morphology of the evaporated ice templates; and\nfinally, (iii) the combination of the hydrophobic nature of the vapor-deposited\npolymer with the same time-constructed hierarchical structure and\nmorphology delivers the proposed superhydrophobic coating on the substrate\nsubface. In the demonstration, we used [2,2]paracyclophanes as the\nstarting materials to form the vapor monomers and produce superhydrophobic\ncoatings of poly( p -xylylenes) on the surfaces. 20 In contrast, in comparison to existing approaches\nof superhydrophobic treatments that required additional and complicated\nmethods to create surface roughness and are limited to flat substrates\nand simple geometry, our proposed fabrication process enables the\nproduction of controlled coating structures in hierarchy and from\nbottom up and yields a tunable surface superhydrophobicity on complex\nsubstrates with ease.",
"discussion": "Results and Discussion The proposed fabrication process\nutilized a home-built vapor deposition\nsystem, which allowed for a temperature-controlled sample stage in\nthe deposition chamber and to accommodate ice templates and the polymer\nprecursors for the described sublimation and deposition polymerization.\nTheoretically, commercially available equipment that is designed for\npolymer vapor deposition can also be used to realize the process.\nDuring the fabrication, 200 mTorr of pressure was ensured to maintain\nthroughout the process, and a key parameter of the temperature was\ncontrolled between −4 and 25 °C for the sublimation (ice\ntemplate, water molecule) and deposition (polymer precursor, quinodimethane\nmonomer) substances to behave. Under the thermodynamic conditions,\nthe ice template and water molecule tend to undergo an endothermic\nphase transition of a sublimation process and from a solid ice to\nvaporized water molecules. In contrast, the vaporized quinodimethane\nmonomers favor deposition and polymerization at the same thermodynamic\nconditions, thus forming solid-phase polymers of poly( p -xylylenes) on the substrates. 21 The two\nprocesses, namely, the sublimation of ice and the deposition of poly( p -xylylene), occur contentiously and simultaneously while\ncompeting with mass transfer flow and complimenting each other with\nthe same volume space. 22 , 23 With fine control of thermodynamic\nconditions, the deposited polymer on such an iced template would result\nin a porous and 3D polymer structure instead of a conventional dense\nthin film, as shown in previous work. 23 We therefore hypothesize that the controlling of morphology and\nroughness of the ice templates can result in the controls of the constructed\nstructures with thereafter deposited polymers based on the described\nvapor sublimation and deposition process and are applicable on a variety\nsubstrate materials with complex geometries. Also, to prevent the\nformation of porous structures that would affect the morphology of\nconstructed structures, we suppose that the manipulation of the ratio\nof the deposition rate to the sublimation rate during the fabrication\nprocess can result in the non-porous structure of deposited polymers.\nOn the basis of previous discoveries, 22 , 24 the architectural\nand structural fidelity was found precisely controllable; i.e., fine\nreplica structures were obtained for the deposited polymers compared\nto their ice templates. During the experiment, substrates were placed\nat a controlled humidity of approximately 75%, and lowering the substrate\ntemperature with a nitrogen bath can instantly obtain an iced template\nvia a condensation process. Interestingly, the structures of the formed\nice template by condensation appear to be fractal as a result of aggregation,\nand the degree of the aggregation was found to be a time-dependent\nprocess. 25 Illustration of the fabrication\nprocess is shown in Figure 1 a that revealed that a substrate surface is subjected to the\ncontrolling crystallization of an iced template, and the proposed\nsuperhydrophobic coating is constructed subsequently by the sublimation\nand deposition process. Theoretically, a wide range of substrates\nare applicable for the fabrication process, and for the demonstration\nin the study, substrates, including papers, ceramics, metals, and\npolymers, are used and tested. By examination of the surface morphology\nof the iced template (ice crystalline structure, before the process)\nand the superhydrophobic coating (replicated polymer structure, after\nthe process), the optical images in panels b and c of Figure 1 also showed high similarity\nfor two surfaces with a surface roughness ( R q ) of 2.89 and 2.52 μm, respectively, and a WCA of approximately\n164° was observed on the fabricated coating. The important vapor\ncompositions during the fabrication process were also examined in\ndetail, and as revealed in Figure 2 a, the analysis results of residual gas using a real-time\nmass spectrometric gas analyzer indicated that the characteristic\npeaks observed at 103 and 138 amu correspond to the presence of the\nmonomer resulting from the pyrolysis of the precursor. The peak observed\nat 18 amu is attributed to the water vapor escaping from the ice template\nunder reduced pressure, and the peak at 40 amu represents the carrier\ngas argon. To verify the chemical structures of the prepared coatings,\nFourier transform infrared spectroscopy (FTIR) analysis was also performed,\nand the results were compared to those of the coatings prepared using\nthe same vapor deposition procedure but without the ice template (see Figure S1 of the Supporting Information). The\nFTIR data for both coatings displayed similar characteristic bands\nof poly( p -xylylene), including an alkyl C–H\nstretch around 2900 cm –1 , an aromatic C–C\nstretch around 1500 cm –1 , and a C–Cl stretch\naround 800 cm –1 , indicating that the presence of\nthe ice template did not affect the chemical structures of the coatings.\nFurthermore, the absence of the O–H peaks around 3500 cm –1 confirms the consumption of the ice template by the\nsublimation process and that the overall FTIR data verified the formation\nof the same poly( p -xylylene) polymer during the proposed\nfabrication process. Verifications of the hypothesis that forming\ncondensed ice template fractal structures and the same deposited poly( p -xylylene) structures can render a tunable hydrophobic\nproperty on a substrate surface were further examined. The WCA of\nthe resulting coatings was found to increase when the condensation\ntime also increased ( Figure 2 b). Specifically, the contact angle increased from 135.1°\n± 4.7° for coatings with 1 min of condensation time to 164.4°\n± 1.9° for coatings with 4 min of condensation time. Notably,\nthe contact angle of coatings without ice templates was measured to\nbe 89.4° ± 0.7°, indicating a remarkable improvement\nin water repellency of conventional vapor deposition coatings with\nthe assistance of ice templates. Intriguingly, the analysis of contact\nangle hysteresis revealed that only coatings with a condensation time\nof 4 min exhibited a contact angle hysteresis of 8.3° ±\n1.7°, while coatings with a condensation time of less than 4\nmin displayed a “sticky” state, 26 in which water droplets adhered firmly to the surface without any\nslipping, irrespective of its tilt ( Figure S2 of the Supporting Information). These findings suggest that the\nwater repellency of the coatings is highly dependent upon the condensation\ntime and that the superhydrophobic properties of the coatings can\nonly be achieved under specific conditions. To further investigate\nthe wetting behavior of the coatings, the surface morphology of coatings\nwith varying condensation times was characterized by scanning electron\nmicroscopy (SEM) and 3D profile imaging. The results shown in Figure 3 b indicate that,\nwith longer condensation times, the profile features left by the ice\ntemplate became increasingly dense and obvious and the surface roughness\nof the coatings increased accordingly. Young’s model, the Wenzel\nmodel, and the Cassie–Baxter model are used to explain how\nthe variation in surface roughness lead to the various wetting behaviors 27 ( Figure 3 a). For smooth surfaces of the coatings without the ice template,\nthe wetting regime is physically and chemically homogeneous as a result\nof the low surface roughness. Therefore, the wettability is mainly\ndetermined by the surface chemistry, in accordance with Young’s\nmodel. As surface roughness increases as a result of the formation\nof ice templates, the real solid–liquid contact area gradually\nincreases, while the wetting regime remains homogeneous as a result\nof the relatively large spacing of profile features, as depicted in\nthe Wenzel model. This wetting behavior results in a higher contact\nangle and higher adhesive force between the solid surface and liquid\ndroplet, as observed in the “sticky” state of coatings\nwith the condensation time less than 4 min. As surface roughness continues\nto increase and the spacing of profile features keeps decreasing with\nrising condensation time, the droplets eventually cause the air to\nbe trapped in the spaces of profile features beneath it. This results\nin the wetting regime becoming heterogeneous with a decrease in the\nsolid–liquid contact area and an increase in the solid–gas\ncontact area. As a consequence, the coatings exhibit ultralow adhesive\nforce and ultrahigh contact angles, namely, the superhydrophobic state\nas predicted by the Cassie–Baxter model. More quantitatively,\nwe can apply the mathematic model regarding the hierarchical roughness\nproposed by Feng et al. 28 to simulate the\nsuperhydrophobicity. 1 The fractal dimension D in\n3D space is 2.2618. For the coating with 4 min of condensation time,\nthe upper and lower limit scales of the surface, L and l , were 10 and 1 μm, respectively. The\nsurface fraction under the water droplet occupied by solid material\nand air, f s and f v , was estimated to be 0.1 and 0.9, respectively. The value\nof θ was measured to be 89°. Therefore, the value of θ f can be calculated by eq 1 as 161.1°, which is close to the measured value (164.4°\n± 1.9°). The result also indicates that the surface roughness\ninduced by ice templates plays a key role in determining the wetting\nbehavior of the coatings. Figure 1 (a) Schematic illustration for the fabrication\nprocess of superhydrophobic\ncoatings via the ice template technique and vapor sublimation and\ndeposition process. Optical images of (b) an ice template fabricated\nby the condensation process and (c) a subsequent superhydrophobic\nsurface as well as water droplets with blue dye placed on the surface. Figure 2 (a) Analysis by real-time mass spectrometry during the\nfabrication\nprocess, which indicates the presence of water molecules from the\nsublimating ice template at 18 amu and p -quinodimethane\nmonomers from pyrolysis of the precursor at 103 and 138 amu. (b) Contact\nangle measurements of superhydrophobic coatings via the ice template\nunder different condensation durations. Figure 3 (a) Schematic illustration of different wetting models\nregarding\nsurface roughness and wetting behavior. (b) Topographical characterization\nof superhydrophobic coatings with different condensation times. The\n3D profile analysis shows the surface structure images of coatings\nwith different condensation times and the corresponding surface roughness,\nand the insets show the corresponding water contact angle images. One important aspect in achieving superhydrophobic\nstate is the\nmultiscale roughness of surfaces, because it helps increase the apparent\ncontact angles and reduce the contact angle hysteresis. 3 To better understand the mechanism underlying\nthe formation of hierarchical structures of ice templates, the growing\nprocess of ice crystals under different conditions was investigated.\nThe schematic illustration of two different patterns of the condensation\nprocess resulting from the different substrate temperatures is shown\nin Figure 4 a, with\nthe results presented in Figure 4 b. At a classic condition of the substrate temperature\naround −20 °C, ice crystals on the substrate were observed\nto grow slowly in terms of the crystal size over time. The results\nare in accordance with previous research that focused on the growth\npattern of ice crystals. Specifically, the growth of ice crystals\noccurs at the solid–gas interface when saturated vapors contact\nthe surface of ice crystals and subsequently condense to form a new\nlayer of ice crystal. 29 The size distribution\nanalysis of ice crystals using ImageJ also reveals that the dimensions\nof each ice crystal continued to develop over time, whereas the number\nof ice crystals in the same area remained approximately constant ( Figure S3 of the Supporting Information). However,\nwhen the temperature of substrates was reduced to below −100\n°C by immersion in a liquid nitrogen bath, ice crystals exhibit\na distinct growth pattern from that described above. Specifically,\nice crystals on the substrate begin to grow in terms of numbers rather\nthan size. The corresponding analysis of the size distribution also\nconfirmed the increasing number of ice crystals with an almost constant\ncrystal size ( Figure S3 of the Supporting\nInformation). The different growth patterns of ice crystals resulting\nfrom varying the substrate temperature can be explained by the different\nlocations where water vapor condenses to ice crystals. In the case\nof substrates at around −20 °C, the water vapor pressure\nwhile decreasing as it approaches the cooled surface remains sufficiently\nhigh (approximately 171 Pa) for condensation to occur on the surface\nof ice crystals. Conversely, the saturated vapor pressure in the case\nof substrates maintained below −100° is nearly 0 Pa (below\n0.02 Pa), implying that nearly all of the water vapor in proximity\nto the cooled substrate would instantaneously condense to ice crystals\nin the air and lead to the formation of a thin layer of ice fog near\nthe surface of the substrates. Subsequently, the newly formed ice\ncrystals will land and attach onto the cooled substrates as a result\nof gravitational force, gradually stacking to form larger structures.\nAs indicated in the SEM images in Figure 4 b, the structures left by the assembly of\nplate-like ice crystals resulted in disordered arrangement, confirming\nthat these small structures were constructed via random attachments\nof ice crystals floating in the air. Furthermore, the sizes of these\nsmall structures exhibit no significant variation between different\ncondensation times, directly verifying that, as the temperature of\nthe substrate remains below −100 °C, the position where\nwater vapor condenses to ice crystals will be located in the air instead\nof the solid–gas interface of ice crystals. Overall, by cooling\nsubstrates to below −100 °C, water vapor will condense\ninto small ice crystals in the air and stack to form larger structures\non cooled substrates, eventually achieving the preparation of ice\ntemplates with hierarchical surface roughness. Figure 4 (a) Schematic illustration\nof two different patterns of the condensation\nprocess. Different substrate temperatures lead to the different structural\nresults of ice templates. (b) Top two rows are the series images of\nice templates taken at different time frames for different substrate\ntemperatures. The bottom two rows are the SEM images of resulting\ncoatings prepared with the substrate temperature below −100\n°C. The chemical durability of the coatings prepared\nwith ice templates\nwas demonstrated through various treatments, including exposure to\nheat, ultraviolet (UV) light, and acid, alkali, and corrosive solutions,\nas shown in Figure 5 a. It was observed that the coatings retained their high contact\nangles even after being exposed to these conditions, indicating the\nexceptional ability to maintain water repellency under a harsh environment.\nThe results were also in accordance with the inherent chemical properties\npossessed by poly( p -xylylenes), including heat and\nUV resistance and chemically inertness, demonstrating that the coatings\nwith the usage of the ice template still retain their inherent chemical\nproperties. 30 The chemical properties of\ncoatings mainly depend upon the outcome of the vapor deposition and\npolymerization process, in which the propagation of the polymer chain\nis of great importance. 31 While ice templates\nundergo sublimation during the process, the produced water vapor basically\ndoes not participate in the propagation reaction of the monomer as\na result of the fact that water is commonly used as an solvent in\nthe radical polymerization process. 32 Therefore,\nthe existence of an ice template does not interfere with the polymerization\nthroughout the whole process and eventually renders the formed polymer\nits inherent chemical properties. Furthermore, the versatility of\ncoatings with superior superhydrophobicity was demonstrated through\ntheir wide applicability to various substrate materials, including\nsilicone dioxide ceramic, polycarbonate plastic, aluminum metal, and\ncellulose filter paper. As shown in Figure 5 b, the coated substrates consistently displayed\nthe WCA ranging from 156° to 164°, which was higher than\nthat observed on uncoated substrates, irrespective of the substrate\nmaterial. Meanwhile, it was also demonstrated that the superhydrophobicity\nof prepared coatings is independent of the surface roughness from\ndifferent substrate materials. Moreover, droplets from commonly available\nsolutions, including deionized water (denoted as water), whole milk\n(denoted as milk), black coffee (denoted as coffee), apple juice (denoted\nas juice), Coca-Cola (denoted as cola), and soy sauce (denoted as\nsauce), were tested on the proposed superhydrophobic coatings, and\nthe WCA was measured to be within the range of 153° and 161°,\nas shown in Figure 5 c, indicating that the superhydrophobic properties of the prepared\ncoatings are independent, regardless of the solution compositions.\nFurthermore, with acknowledging the distinctive and widely adopted\nfeatures of the vapor deposition coatings being their coating fidelity\nand precision on intricate surfaces and devices, i.e., complex 3D\nor curved configurations and geometries, we finally confirmed the\nhypothesis of rendering the superhydrophoibic coating technology with\nthe proposed vapor sublimation and deposition process onto complex\nconditions of substrates, and the results in Figure 5 d showed that the superhydrophobic coatings\nwere successfully applied on complex device surfaces, such as an intraocular\nlens and a cardiovascular stent, using the proposed fabrication process.\nA high contact angle of approximately 155.3° was measured on\nthese surfaces with consistency. In parallel to its sibling of conventional\nvapor deposition coating being a versatile and precise treatment for\ndelicate and complex materials and devices, the present process in\nthe study agreed with the same processing versatility, additionally\nfacilitated the control of the important surface hydrophobicity property,\nand achieved the level of superhydrophobicity on a variety of encountered\nsubstrate conditions. Figure 5 Demonstration of the properties of superhydrophobic coatings.\n(a)\nDurability tests of coatings under different environmental conditions.\n(b) Dyes of water droplets on different substrates (silicone dioxide\nceramic, polycarbonate plastic, aluminum metal, and cellulose filter\npaper, separately) before and after superhydrophobic coating. (c)\nDifferent types of liquid droplets on a coated substrate. (d) Application\nof superhydrophobic coating on a medical device with intricate surfaces."
} | 6,683 |
32695749 | PMC7338834 | pmc | 37 | {
"abstract": "Silkworm silk is mainly known as a luxurious textile. Spider silk is an alternative to silkworm silk fibers and has much more outstanding properties. Silk diversity ensures variation in its application in nature and industry. This review aims to provide a critical summary of up-to-date fabrication methods of spider silk-based organic-inorganic hybrid materials. This paper focuses on the relationship between the molecular structure of spider silk and its mechanical properties. Such knowledge is essential for understanding the innate properties of spider silk as it provides insight into the sophisticated assembly processes of silk proteins into the distinct polymers as a basis for novel products. In this context, we describe the development of spider silk-based hybrids using both natural and bioengineered spider silk proteins blended with inorganic nanoparticles. The following topics are also covered: the diversity of spider silk, its composition and architecture, the differences between silkworm silk and spider silk, and the biosynthesis of natural silk. Referencing biochemical data and processes, this paper outlines the existing challenges and future outcomes.",
"conclusion": "Conclusion The range of spider silk applications is extremely broad due to its unsurpassed biophysicochemical properties and high degree of adaptability. Spider silk provides a good basis for the formation of hybrid functional materials with many uses—an option already being explored. However, it is clear that this field is only at its first stages of development based on the presented approaches for the synthesis of spider silk-based organic-inorganic hybrid materials. Moreover, large-scale industrial production of these materials is currently challenging due to some unavoidable difficulties related to natural spider silk fabrication. Commercially available artificial fibers are still far from natural as the structure and properties of natural spider silk are difficult for accurate reproduction. Furthermore, in comparison with more available silkworm silk, the nanostructure and macro-properties of spider silk vary significantly. The architecture and properties of natural spider silk fibers are still not fully researched. Therefore, further understanding of spider silk structure at both the molecular and supra-molecular levels, as well as its formation process is crucial for the development of more successful material modification and manipulation protocols. Based on these ideas, more accessible and durable advanced fibrous materials with tunable mechanical and biological properties can be generated. New evidence of hierarchical architecture within spider silk contributes to a better understanding of the possible integration of inorganic components within silk fibers to produce biopolymer hybrids with considerably improved functional properties.",
"introduction": "Introduction Nature is rife with nanocomposites, which exhibit high toughness and are found in various tissues—from abalone shells (Smith et al., 1999 ) to human bones (Ji and Gao, 2004 ). Spider silk, one of the most incredible natural hierarchically ordered materials, possesses outstanding material properties, namely high toughness (about three times higher than Kevlar toughness), high extensibility (30% elongation to fracture) equivalent to rubber elongation, and biocompatibility (Vollrath, 2000 ; Allmeling et al., 2006 ; Porter et al., 2013 ). Generally, spider silk is a highly ordered protein fiber spun by spiders. In-depth understanding of natural structuring and synthesis highlights the importance of hierarchical structures in terms of real world functionality (Buehler, 2013 ). This denotes thrilling prospects concerning the idea of converting the properties of hierarchically ordered structures to novel material functions. This approach generates opportunities for innovative material applications in the fields of energy and sustainability, medicine, and nanobiomedical technology (Zhang, 2003 ; Barthelat, 2007 ; Aizenberg, 2010 ; Hauser and Zhang, 2010 ; Salgado et al., 2010 ). It is therefore unsurprising that spider silk is considered one of the most promising materials for industrial applications. Silk is also attractive in optics and photonics (Huby et al., 2013 ), and tissue regeneration (Bandyopadhyay et al., 2019 ). Additionally, the mild conditions of its biosynthesis imply that the fabrication of innovative functional silk-based smart materials would be an eco-friendly process with minimal negative ecological impact. Within the last few years, there has been a dramatic increase in the use of natural fibers to create new hybrid materials (Lau and Cheung, 2017 ). Recent advances in natural fiber development, composite science, and genetic engineering have presented remarkable opportunities for novel high-performance functional materials (Wang F. et al., 2014 ). There is a growing interest in high-performance spider silk-based functional materials, while silkworm silk is widely accepted and used. In this context, the paper summarizes recent progress in spider silk-based organic-inorganic hybrid material synthesis. To provide a more comprehensive understanding of spider silk, the structure–property–function relationships of spider silk fibers are discussed. Considering recent findings, the transformation of silk proteins into highly efficient fibers is also explored. Information regarding spider silk architecture is also extended to advance the targeted design of ultra-performance functional materials based on fibrous proteins. This paper focuses solely on spider silk-based hybrids. Therefore, the structure and properties of silkworm silk, as well as those hybrids based on it are not the subject of this review. Differences between structure and material performance regarding silkworm and spider silk are explored, however, to provide a deeper understanding of silk backbone performance in the fabrication of silk-based functional materials. Namely, their organization at the molecular level, interactions to form secondary structures, and various mechanical properties are highlighted. Challenges and approaches to the large-scale production of spider silk-based materials for numerous applications are also reviewed. Additionally, the high complexity of spider silk organization, and the tunability of its properties are revealed. Recent advances and emerging strategies concerning the fabrication and applications of natural or bioengineered spider silk–inorganic nanoparticle hybrid materials are described. Lastly, this paper concludes with the prospects of hybrid spider silk-based materials."
} | 1,656 |
30373122 | PMC6266336 | pmc | 38 | {
"abstract": "Synaptic devices with bipolar analog resistive switching behavior are the building blocks for memristor-based neuromorphic computing. In this work, a fully complementary metal-oxide semiconductor (CMOS)-compatible, forming-free, and non-filamentary memristive device (Pd/Al 2 O 3 /TaO x /Ta) with bipolar analog switching behavior is reported as an artificial synapse for neuromorphic computing. Synaptic functions, including long-term potentiation/depression, paired-pulse facilitation (PPF), and spike-timing-dependent plasticity (STDP), are implemented based on this device; the switching energy is around 50 pJ per spike. Furthermore, for applications in artificial neural networks (ANN), determined target conductance states with little deviation (<1%) can be obtained with random initial states. However, the device shows non-linear conductance change characteristics, and a nearly linear conductance change behavior is obtained by optimizing the training scheme. Based on these results, the device is a promising emulator for biology synapses, which could be of great benefit to memristor-based neuromorphic computing.",
"conclusion": "4. Conclusions In this paper, a Ta/TaO x /Al 2 O 3 /Pd memristor is fabricated, to be used as artificial synapse. The device shows bipolar analog-resistive switching behavior. Moreover, multilevel conductance states with a satisfying retention time (>1000 s) can be obtained by modulating voltages or compliance currents under DC sweeping mode. Based on the bipolar analog switching, synaptic functions, including long-term potentiation/depression, paired-pulse facilitation, and spiking time dependent plasticity are successfully mimicked. For ANN applications, the determined target conductance, the linearity, and the writing errors are carefully examined. The results suggest that as an artificial synapse, the Ta/TaO x /Al 2 O 3 /Pd memristor is a promising candidate for neuromorphic computing.",
"introduction": "1. Introduction Over the last decades, rapid advances in digital computing system based on complementary metal-oxide semiconductor (CMOS) integrated circuit technology have substantially changed society. However, due to the limitations of classical von-Neumann computers (the von-Neumann bottleneck) in speed, power efficiency, and parallel processing, there are urgent demands for novel computing structures and systems [ 1 ]. The human brain is likely to be the most efficient computing system, because the operating frequency of our brain is in the range of 1–10 Hz, and it consumes only around 1–10 W of power, which means the energy consumption per synaptic event is only approximately 1–100 fJ [ 2 ]. Therefore, the novel computing system—neuromorphic computing, inspired by the brain—has attracted scientists’ attention in recent years for its advantages, such as being massively parallel and fault-tolerant. The weight modulation ability of synapses is known as synaptic plasticity, which is believed to be the primary reason for learning and memory in the brain. In order to implement neuromorphic computing, such as artificial neural networks (ANN), an electronic synaptic device is necessary. Recently, the implementation of artificial synapses with memristors has been proposed. Memristors are two compact terminal devices that change their resistances when subjected to electrical stimulation [ 3 , 4 , 5 , 6 ]. Several memristors, ranging from resistive random access memory (RRAM) [ 7 , 8 , 9 , 10 , 11 ], to phase change memory (PCM) [ 12 ], to ferroelectric RAM [ 13 , 14 , 15 ], have been proposed for neuromorphic computing applications as artificial synapses. Several memristors based on new materials [ 16 , 17 ] have been proposed for neuromorphic computing. However, when memristors are employed in neuromorphic computing systems (e.g., artificial neuron networks), binary memristors with only two resistance states (i.e., high resistance state (HRS) and low resistance state (LRS)) have been proven to be effective only in some specific applications [ 18 , 19 ]. In some neuromorphic computing systems designed for complex applications, such as image recognition, the use of only two states as synaptic weights presents disadvantages in performances [ 20 , 21 ]—for example, low accuracy or area-efficiency. On the other hand, in biology neuromorphic systems, synaptic weights are continuously tunable in depression and potentiation; thus, memristors with gradually changing conductance in bipolarity could be more like the biology synapse, and can therefore emulate brain functions better than binary memristors. As artificial synapses, memristors with tunable conductance have attracted growing attention for being promising candidates for weight storage in neuromorphic computing systems, owing to the advantages in accuracy and area-efficiency. Several methods have been discussed to implement analog-resistive switching behavior, including using multiple memristors to construct one synapse [ 22 ], utilizing a unipolar analog behavior in some metal oxide-based filamentary memristors [ 11 , 23 , 24 ], optimizing programming schemes [ 25 , 26 ], adding heat enhancement layers [ 27 ], or using non-filamentary memristors [ 28 , 29 , 30 ]. Compared with the filamentary memristors, non-filamentary memristors can implement multilevel states more easily, but usually have poorer retention [ 31 , 32 , 33 ] However, realizing bipolar analog conductance change in both SET (transition from HRS to LRS) and RESET (transition from LRS to HRS) processes with satisfying retention time remains an open challenge. In this paper, a fully CMOS-compatible, forming-free, and non-filamentary memristor device based on Ta/TaO x /Al 2 O 3 /Pd, with analog SET and RESET processes, is proposed for neuromorphic computing as an artificial synapse. The direct current (DC) sweeping results demonstrate that the device has bipolar analog resistance switching behavior, and the multilevel conductance states can be obtained with satisfying retention time. Synaptic plasticity, including long-term potentiation/depression (LTP/LTD), paired-pulse facilitation (PPF), and spiking-time-dependent plasticity (STDP), can be mimicked by our devices. For the applications in ANN, determined target conductance states and the linearity of conductance change are carefully examined.",
"discussion": "3. Results and Discussions The resistive switching characteristics of the device were evaluated under DC programming conditions. The typical current–voltage ( I–V ) characteristic of the Ta/TaO x /Al 2 O 3 /Pd device under DC sweep mode from −6 V to 6 V is shown in Figure 1 d. The device is forming-free, and no abrupt change of current in both SET and RESET switching processes is observed, indicating a bipolar analog resistive switching feature. Within 6 V and −6 V stop voltages on SET and RESET processes, a ~10 3 ratio between HRS and LRS can be obtained (read voltage is 1 V), which is larger than our recent work of similar TaO x /Al 2 O 3 stack device (~10 2 ratio, Ti/AlO x /TaO x /Pt) [ 34 ]. To further demonstrate the analog characteristics, the DC sweep with different working voltages and without compliance currents (SET voltage: 2.5 V, 3 V to 5.5 V; and RESET voltage: −2 V, −2.5 to −6 V) and the DC sweep with different compliance currents during SET process are shown in Figure 2 . The initial resistance of the device is ~10 11 Ω (read at 1 V). When the positive sweeping voltage is applied to the device, the resistance of the device is retained until the voltage reaches 2.5 V, then the resistance gradually decreases. During the consecutive SET process, as shown in an inset of Figure 2 b, the responding currents (read at 1 V) can gradually increase with the increment of the stop voltages, indicating that different conductance states can be obtained in the SET process. Various conductance states can also be obtained by setting different compliance currents during the SET process. With compliance currents from 500 nA to 2.2 mA, the corresponding I–V curves and the 60 different resulting conductance states are shown in Figure 2 c and the inset, respectively. The RESET process can be implemented by applying a negative DC sweeping voltage to the device. As shown in Figure 2 a, eight consecutive negative DC sweeps with various stop voltages are applied to the device. As the voltages decrease from −2 to −6 V with a −0.5 V step, the device is switched to a higher resistance state after each step. Moreover, the multilevel resistance states can be preserved within satisfying retention time, as shown in Figure 2 d. The multilevel resistance states are obtained by consecutive positive voltage sweepings (2 to 6 V with a 0.25 V step). After each sweeping, the device resistance states are monitored by a series of 1 V reading pulses at 0.5 s intervals. As it is shown in Figure 2 d, though with slightly decay, nine different states can be clearly distinguished after 1000 s. The characteristics of the bipolar analog-resistive switching in pulse mode are investigated via positive (0 to 4.5 V) and negative (0 to −5 V) triangle pulses, as shown in Figure 3 a,b, respectively. The curves of current and voltage versus time for the SET and RESET processes are shown in the insets of Figure 3 a,b, respectively. These results further confirm the analog resistive switching characteristics under both positive and negative pulses. The results reveal that in both the SET and RESET processes, gradual tuning of the multilevel conductance states can be obtained. Bipolar analog resistive switching characteristics are fully analogous to the biology synapse; thus, the devices have the potential to mimic synaptic functions in neuromorphic computing system. Long-term potentiation/depression (LTP/LTD) is when the synaptic weight can be changed gradually under spiking signals and the changed weight can be maintained from several minutes to years [ 35 ]. To evaluate the long-term potentiation/depression of a device, 50 consecutive pulses with different pulse amplitudes and widths are applied to the device, as shown in Figure 4 . All the conductance of the device is monitored by 1 V reading voltage. The change of conductance can be modulated by different amplitudes and widths. As shown in Figure 4 a,b, the amplitude here was fixed at 5.5 V during potentiation and −5.5 V during depression, with different widths (1 μs, 10 μs, and 100 μs). In addition, Figure 4 c,d show the potentiation and depression with a fixed 100 μs width and different amplitudes (potentiation: from 4.5 to 5.5 V; depression: from −4.5 to −5.5 V). With a higher amplitude or larger width, the change of the conductance is increased in both potentiation and depression. For our device, when the pulse amplitude (write voltage) is ~±4.5 V and the pulse width is 1 μs, the write current is around ~10 −5 A; thus, the switching energy is 50 pJ per spike. To conclude, the device conductance is continuously increased by positive pulses, which can mimic long-term potentiation. In addition, the device conductance is continuously decreased by negative pulses, which can mimic long-term depression. Moreover, the device can emulate other synaptic features, such as paired-pulse facilitation (PPF) and spiking-time-dependent plasticity (STDP), as shown in Figure 5 . Most research on artificial synapses focuses on the long-term plasticity, because long-term changes provide a physiological substrate for learning and memory. However, short-term plasticity is also significant, since it supports a variety of computations, such as synaptic filtering, adaptation, and enhancement of transients, decorrelation, burst detection, and sound localization [ 36 ]. PPF is an important kind of short-term plasticity. In biological synapses, PPF functions can be described as follows: the second post-synaptic response current becomes larger than the first under two successive spike stimuli, with the interval time of spikes less than recovery time [ 8 ]. The experimental demonstration of PPF functions in our device is shown in Figure 5 a. When a pair of pulses is applied to the device, the conductance gradually increases during the positive pulses, and the maximum responding current of the second pulse is clearly larger than the first, and a decay phenomenon can be observed during the pulse interval, which is similar to the PPF in the biological system. In biological systems, synaptic weight can be modulated by the temporal relationship of the activity between the pre- and post-synaptic neurons, which is called spiking-time-dependent plasticity (STDP). According to STDP, the change of synaptic weight (ΔW) is a function of the time difference between pre- and post-synaptic activity (Δt). To emulate the STDP function in the device, a pair of pulses acting as the spiking signals with different time intervals is applied to the device. Individual pre-synaptic or post-synaptic spiking signals are designed as a pair of pulses (−2.5 V, 10 μs pulse and a 2.5 V triangle pulse) applied to the top and bottom electrode, respectively, as shown in Figure 5 b. It should be noted that an individual positive signal or an individual negative signal is not strong enough to modulate the resistance of the device. As shown in Figure 5 b, the effective signal to the device is the pre-synaptic signal minus the post-synaptic signal. When the pre-spike appears before the post-spike (Δt > 0), the conductance (synaptic weight) of the device is enhanced (potentiation), and the change in weight decreases with the increase of Δt. On the contrary, when the pre-spike appears after the post-spike, the conductance of the device depresses and the change of the weight decreases with the increase of Δt. The measurement result shows that the Ta/TaO x /Al 2 O 3 /Pd device can emulate the STDP learning rules successfully, which has potential to be used in the spiking neuron network (SNN). To fully explore bipolar conductance tuning characteristics and demonstrate the potential application of the device in some specific neuromorphic computing systems like ANN, determined target conductance states with different initial states have been tested. As shown in Figure 6 a, the initial state is 2.41 nS, after two tuning processes: 5.7 V positive pulses with 10 μs width for rough-tuning, and −5 V negative pulses with 10 μs width for fine-tuning. The target conductance state of 5.5 nS can be obtained with little deviation (<1%). The same target conductance state can also be obtained when the initial conductance state is 13.5 nS, by −5.7 V negative pulses with 10 μs width for rough-tuning and 5 V positive pulses with 10 μs width for fine-tuning, as shown in Figure 6 b. As shown in Figure 6 c,d, another target conductance state (10 nS), can be obtained with little deviation. It is worth noting that the target conductance states are determined randomly. Based on this result, it can be proven that precision is achieved across a wide dynamic range. Writing error is a standard plot when characterizing resistive switching write noise. The write error of the device has been tested, as shown in Figure 7 . A DC sweeping with 100 μA compliance current is used to get nearly the same initial states. Only one programming pulse (4.5 V/10 μs for potentiation and −4.5 V/10 μs for depression) is applied after each DC sweeping. The conductance states (total 10 cycles) are obtained by 1 V reading voltage. As shown in Figure 7 b,d the standard deviation is 0.079 nS after one potentiation pulse, and 0.11 nS after one depression pulse, respectively. The dynamic range is around 20 nS under 4.5 V/10 μs training pulses. As a result, the write error is only around 0.6% of the total dynamic range. The recognition accuracy of the ANN highly depended on the linearity of the synaptic weight change—i.e., the recognition accuracy is low under high non-linearity [ 37 , 38 ]. However, as shown in Figure 4 , the device is highly non-linear. To improve the linearity of the conductance change of the device, a non-identical pulse scheme is adopted, as shown in Figure 8 . The training pulses are fixed at width but with increasing amplitudes. The amplitude range of the potentiation process is from 2 to 6 V with 0.1 V steps, and the range of the depression process is from −2 to −6 V with −0.1 V steps. The weight updates are recorded in four training cycles, as shown in Figure 8 . The non-linearity factor (NL) has been calculated by NL = average ( G − G l i n e a r G l i n e a r ) [ 39 ], so the non-linearity factors of the normal training method are 1.09, 1.427, and 1.332 respectively, based on the data in Figure 4 a. In addition, the non-linearity factors of the incremental training method are −0.62 for long-term potentiation and 0.13 for long-term depression, based on the data in Figure 8 a. The investigation of the switching mechanism of the device is shown in Figure 9 . The conductance of the filamentary memristors mostly depends on the size and morphology of the conductive filament with several nanometers diameter in the device. Thus, the conductance of filamentary memristor does not significantly change with the change of the electrode areas. The I–V curves of the 1st SET process and conductance distribution of 25 different devices at LRS with various electrode areas (from 10 to 100 μm 2 ) are shown in Figure 9 a,b, respectively. In Figure 9 a, the current level after SET shows a positively proportional relationship with the electrode area. In statistical analysis of 25 devices at LRS in Figure 9 b, such a trend can be more clearly seen in the plotting of the conductance with the electrode area. As shown in inset of Figure 9 b, the linear fit result confirms that the device conductance scales linearly with the device area. As a result, the switching occurs across the entire electrode area, but not just within a local filament, suggesting a non-filamentary switching mechanism of Ta/TaO x /Al 2 O 3 /Pd device. The temperature dependencies of the device conductance at LRS and HRS are studied in Figure 9 c,d, respectively. With the increase of temperature, the conductance at both LRS and HRS increases as well, indicating the semiconductor conduction behavior of the device. To explain the switching mechanism of the device, we proposed a simple model [ 40 ], shown in Figure 9 e. The device can be divided into three parts: a barrier layer (Al 2 O 3 ), a switching area (interface of Al 2 O 3 and TaO x ), and a conductive oxidation layer (TaO x ). The switching area is located at the interface of Al 2 O 3 and TaO x . During SET operation, a positive voltage is applied on the top electrode, the oxygen ions in the barrier layer are pulled away from the interface layer, and the materials in the interface are reduced. During RESET operation, a negative voltage is applied on the top electrode, the oxygen ions in the barrier layer are pushed into the interface layer, and the materials in the interface are oxidized. The push-and-pull of the oxygen ions in the surface can change the resistance of the device. To implement neuromorphic computing, the device should be integrated into an array. To operate an array, a half-bias scheme is a common method. However, our device has a low ON/OFF ratio (<100) between the selected voltage and the half-selected voltage, which may cause a sneak path issue during write operation. As a result, it is hard for our device to implement a dense crossbar array without the help of a transistor or selector device. A one transistor one resistor (1T1R) or one selector one resistor (1S1R) structure should be adopted to overcome the sneak path issue during writing operation. The device structure can still be optimized to improve the linearity of the conductance changes and decrease the working voltage. In addition, the detailed non-filamentary switching mechanism in this device needs to be further explored."
} | 4,982 |
26483629 | PMC4591430 | pmc | 40 | {
"abstract": "Synaptic plasticity plays a crucial role in allowing neural networks to learn and adapt to various input environments. Neuromorphic systems need to implement plastic synapses to obtain basic “cognitive” capabilities such as learning. One promising and scalable approach for implementing neuromorphic synapses is to use nano-scale memristors as synaptic elements. In this paper we propose a hybrid CMOS-memristor system comprising CMOS neurons interconnected through TiO 2− x memristors, and spike-based learning circuits that modulate the conductance of the memristive synapse elements according to a spike-based Perceptron plasticity rule. We highlight a number of advantages for using this spike-based plasticity rule as compared to other forms of spike timing dependent plasticity (STDP) rules. We provide experimental proof-of-concept results with two silicon neurons connected through a memristive synapse that show how the CMOS plasticity circuits can induce stable changes in memristor conductances, giving rise to increased synaptic strength after a potentiation episode and to decreased strength after a depression episode.",
"introduction": "1. Introduction Biological networks provide a tantalizing proof of the existence of a physically implementable computing architecture that is distributed, fault-tolerant, adaptive, and that outperforms conventional architectures in many important problems such as visual processing and motor control. This has motivated the development of various neuromorphic computing systems whose architectures reflect the general organizational principles of nervous systems in an effort to partially reproduce the immense efficiency advantage that biological computation exhibits in some problems. These neuromorphic systems are organized as populations of excitatory and inhibitory spiking neurons with configurable synaptic connections (FACETS, 2005–2009 ; Navaridas et al., 2013 ; Benjamin et al., 2014 ; Merolla et al., 2014 ; Ning et al., 2015 ). Synapses outnumber neurons by several orders of magnitude in biological neural networks (Binzegger et al., 2004 ). Reproducing these biological features in neuromorphic electronic circuits presents a scaling problem, as integrating thousands of dedicated synapse circuits per neuron can quickly become infeasible for systems that require a large number of neurons (Schemmel et al., 2007 ). This scaling problem has traditionally been solved by either treating synapses as simple linear elements and time-multiplexing spikes from many pre-synaptic sources onto the same linear circuit (Benjamin et al., 2014 ), or by treating them as basic binary elements that can be set either ′ on ′ or ′ off ′ externally, without learning abilities (Merolla et al., 2014 ). Real synapses, however, exhibit non-linear phenomena like spike timing dependent plasticity (STDP) that modulate the weight of an individual synapse based on the activity of the pre- and post-synaptic neurons (Bi and Poo, 1998 ). The modulation of synaptic weights through plasticity has been shown to greatly increase the range of computations that neural networks can perform (Abbott and Regehr, 2004 ). Capturing the plasticity properties of real synapses in analog neuromorphic hardware requires the use of distinct physical circuits/elements for each synapse. In conventional CMOS, this can lead to restrictions on scalability. Some potential solutions to the scalability issues in pure CMOS technology involve the use of very large integrated structures (e.g., up to a full wafer, Schemmel et al., 2012 ) or the adoption of deep submicron technologies (Noack et al., 2015 ). Scalability restrictions however can be greatly relaxed if one resorts to compact nano-scale circuit elements that can reproduce the plasticity properties of real synapses. One potential candidate for these elements is the “memristor.” Chua ( 1971 ) described the memristor as an element which behaves somewhat like a non-linear resistor with memory . Since HP first linked resistively switching devices with the concept of a memristor (Strukov et al., 2008 ), work on memristive devices has mostly focused on digital storage and logic functions (Linn et al., 2012 ; You et al., 2014 ), but there are also applications as analog/multi-level storage (Moreno et al., 2010 ; Shuai et al., 2013 ) and even memristive encryption (Lin and Wang, 2010 ; Du et al., 2014 ). In the neuromorphic community, memristors are seen as ideal devices for synapse implementations, as they combine three key functions in one device. Memristors can implement biologically realistic synaptic weight updates, i.e., learning (Jo et al., 2010 ), they can carry out long term multi-valued weight storage, and they can also communicate weighted pre-synaptic activity to the postsynaptic side (Saighi et al., 2015 ), significantly relaxing scalability restrictions (Indiveri et al., 2013 ). Typically, plasticity in these memristive synapses is evoked by applying specific waveforms to the two terminals of the memristor, with the waveforms aligned to pre- respectively postsynaptic pulses (Jo et al., 2010 ). The correlation of the waveforms across the memristor in turn implements STDP-like plasticity (Mayr et al., 2012 ), with the form of the STDP curve defined by the applied wave shape (Serrano-Gotarredona et al., 2013 ). Both hardware and software models of plasticity based on the basic STDP mechanism are typically chosen primarily for their simplicity (Mayr and Partzsch, 2010 ). It has been argued however that more elaborate models of plasticity are required to reproduce the experimental evidence obtained from more complex synaptic plasticity experiments in real neural systems, and to implement algorithms that can learn to store and classify correlated patterns (Senn and Fusi, 2005 ; Sjöström et al., 2008 ; Lisman and Spruston, 2010 ). In this work we present a neuromorphic implementation of one of these extended plasticity models that implements a spike-based Perceptron learning algorithm (Brader et al., 2007 ), which makes use of both analog CMOS circuits and TiO 2− x memristive devices. Compared to the more widely used STDP paradigm, the implementation of this learning algorithm on memristors does not employ the postsynaptic spike timing. Instead, it relies on the correlation of presynaptic spikes with signals derived from the postsynaptic neuron, such as its membrane potential and a measurement of its recent spiking activity. These requirements lead to a novel and quite different approach to the CMOS driver circuits which does not require the generation of temporally long waveforms on the pre- or postsynaptic sides. In addition to spike timing, plasticity in biological synapses also depends on the firing rate of the post-synaptic neuron (Sjöström et al., 2001 ), a phenomenon that can not be captured by pair-wise STDP mechanisms (Pfister et al., 2006 ). The spike-based perceptron learning rule explicitly contains a term that reflects the recent firing rate of the neuron and is thus able to realize the rate-dependence of synaptic weight updates. The rule is also able to realize weight updates that depend on pre-post spike timing even though it does not explicitly depend on the post-synaptic spike times. Instead, it uses the membrane potential of the post-synaptic neuron as an indirect estimator of post-synaptic firing times. The rule is thus able to reasonably match the behavior of biological synapses while having a functional form that can be implemented efficiently on pure CMOS or on hybrid CMOS-memristor neuromorphic systems. We introduce the spike-based Perceptron learning model in Section 2.1 and the TiO 2− x memristive devices employed in this implementation in Section 2.2. The adaptation of the learning model to memristors is described in Section 2.3. Considerations for crossbar operation of this paradigm are given in Section 2.4. Section 3.1 shows basic results characterizing operation of the memristors. Characterization of the learning CMOS driver circuits implemented in VLSI are detailed in Section 3.2. Finally, results from implementing the spike-based Perceptron learning with the CMOS driver circuits on the memristors are presented in Section 3.3.",
"discussion": "4. Discussion 4.1. Memristive device characteristics Figure 5 shows the typical operation of a “well-behaved” memristor in response to trains of input voltage pulses. A number of key features are noteworthy: Bipolar operation: Pulses of opposite polarity precipitate resistive state changes in opposite directions. In the case of our devices, a positive voltage applied to the top electrode (bottom electrode grounded) causes potentiation. Bidirectionally gradual switching: Transitions between resistive state floor and ceiling occur over many pulses, not just one. This allows the device to work as a multi-level weight artificial synapse (as opposed to binary). Bidirectionally saturating switching: When a device is bombarded by trains of identical voltage pulses it approaches its operational resistive state floor and ceiling in progressively smaller steps. This implies that the middle of the resistive state range is expected to be most often unoccupied in operando , as it is traversed quickly in either direction under pulsing. The resistive state will be therefore multi-level in nature, but most of the time distinctly high or low. Biasing parameter variation tolerance: The device can remain functional under a relatively wide range of bias voltages. We obtain good switching behavior for voltage pulses in the 0.75–1.2 V range. The device can safely operate with voltage pulses of up to 2 V. This bodes well for operation in tandem with practical electronic systems and for resistive switching behavior tuning. These features allow the memristive devices to exhibit the correct behavior when coupled to the neuromorphic circuits described in Section 2.3, both as binary and as multi-level synapses. Only binary synaptic operation was investigated in the plasticity experiments. 4.2. The spike-based perceptron learning rule in CMOS-memristor architectures The spike-based Perceptron plasticity rule has been implemented in CMOS neuromorphic systems using various types of circuits such as subthreshold circuits (Mitra et al., 2009 ) and switched capacitor circuits (Noack et al., 2015 ). In this paper, we have presented a physical implementation of the first hybrid CMOS-memristor architecture that implements a spike-based Perceptron learning plasticity rule. The physical CMOS-memristor system we presented is a standalone system in which the custom CMOS chip connects directly to the memristive devices. The CMOS chip implements the neuron elements together with dedicated per-neuron circuits that can program (potentiate or depress) the memristive synaptic elements as well as sense their conductances/weights to generate proportional Excitatory Post-Synaptic Currents (EPSCs) in the post-synaptic neuron in response to pre-synaptic spikes. We have presented direct measurements that illustrate the behavior of this physical CMOS-memristor system. This is the first standalone neuromorphic system that combines custom neuron circuits with memristor programming and sensing circuits acting on physical memristive devices. Many highly accurate and biologically grounded, i.e., non-empirical, synaptic plasticity rules make use of several auxiliary variables beyond spike times in the pre- and post-synaptic neurons to control synaptic weight updates (Pfister et al., 2006 ; Brader et al., 2007 ; Clopath and Gerstner, 2010 ; Mayr and Partzsch, 2010 ; Graupner and Brunel, 2012 ). These auxiliary variables may include low-pass filtered versions of the membrane potential (Clopath and Gerstner, 2010 ) or a low-pass filtered version of the neuron's spike train (Brader et al., 2007 ). Interestingly, the time difference between pre- and post-synaptic spikes does not figure explicitly in these models. This presents a problem for current neuromorphic memristive architectures that mainly depend on this time difference (through the overlap between pre- and post-synaptic spike-triggered waveforms) to induce weight updates. These architectures will not be able to handle weight updates that are triggered on single pre- or post- synaptic spikes. The architecture we presented triggers weight updates on single pre-synaptic spikes. This has a significant advantage: at the time of a pre-synaptic spike, the neuromorphic synapse can be immediately potentiated or depressed based on the current state of the post-synaptic neuron; the neuromorphic system does not have to wait for a post-synaptic spike to know the outcome of the plasticity event. Implementations of classical pair-wise STDP rules using memristors typically trigger long waveforms on the pre- and post-synaptic sides of the memristor in response to pre- and post-synaptic spikes respectively. When these waveforms overlap, the potential difference across the memristor exceeds a threshold and changes in memristor conductance occur. The duration of these waveforms dictate the STDP window. The overlapping waveforms paradigm is problematic in the high spike rate regime as multiple spikes can occur within the STDP window, thereby corrupting the synaptic weight update. By contrast this problem is completely avoided in the case of the spike-based Perceptron learning rule. In the original learning rule (Brader et al., 2007 ) the weights were bistable, i.e., they gradually drifted to one of two stable states. This had the effect of consolidating synaptic changes and making it more difficult for a synaptic pattern to be corrupted by spurious spikes. Our architecture does not implement such continuous (non event-driven) weight drift. This indicates that synaptic rule features that simplify pure CMOS implementations like bistable weights do not necessarily translate to simpler CMOS-memristor implementations. 4.3. Outlook The architecture we describe represents a first step toward hybrid CMOS-memristor implementations of more elaborate plasticity rules that go beyond standard STDP. Further developments will have to address the problem of plastic crossbar operation as well as mechanisms that allow continuous or non event-driven weight updates. Conflict of interest statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest."
} | 3,625 |
23962644 | null | s2 | 41 | {
"abstract": "Spider silks have been a focus of research for almost two decades due to their outstanding mechanical and biophysical properties. Recent advances in genetic engineering have led to the synthesis of recombinant spider silks, thus helping to unravel a fundamental understanding of structure-function-property relationships. The relationships between molecular composition, secondary structures and mechanical properties found in different types of spider silks are described, along with a discussion of artificial spinning of these proteins and their bioapplications, including the role of silks in biomineralization and fabrication of biomaterials with controlled properties."
} | 168 |
33574397 | PMC7878756 | pmc | 42 | {
"abstract": "In this work, we proposed a facile approach to fabricate a superhydrophobic surface for anti-icing performance in terms of adhesive strength and freezing time. A hierarchical structure was generated on as-received Al plates using a wet etching method and followed with a low energy chemical compound coating. Surfaces after treatment exhibited the great water repellent properties with a high contact angle and extremely low sliding angle. An anti-icing investigation was carried out by using a custom-built apparatus and demonstrated the expected low adhesion and freezing time for icephobic applications. In addition, we proposed a model for calculating the freezing time. The experimented results were compared with theoretical calculation and demonstrated the good agreement, illustrating the importance of theoretical contribution in design icephobic surfaces. Therefore, this study provides a guideline for the understanding of icing phenomena and designing of icephobic surfaces.",
"introduction": "Introduction Icing problems present many challenges as the diversity of ice formation. In natural environments, ice accumulation can be found on a wide range of temperatures and humidity owing to the different scenarios, including freezing rain, snow, and frost formation. Specifically, ice accretion on the wings of aircraft by freezing rain or fog icing may cause a sudden loss of control owing to the weight overloading and lack of lifting force. Moreover, ice bulks form on the fuselage may be ingested into the engines causing a partial or total loss of thrust 1 , 2 . Furthermore, ice accretion on power transmission systems, vehicles, or offshore platforms might lead to massive damage and potentially endangering people 3 – 5 . Generally, anti-icing strategies might be separated into active and passive methods. The current active strategies for combatting icing problems primarily involve the heating systems, chemical deicing fluids, and mechanical removal 6 – 11 . On the opposite side, it would be advantageous if surfaces can passively hinder the ice formation and ensure the ease removal process without any external energy 12 – 19 . These processes are more efficient, environmentally favorable compared to industrial active methods and can be achieved using the physicochemical process based on texturing structure incorporates with a low-surface-energy compound. Superhydrophobic surfaces, which inspired by the Lotus leaf concept, are believed as a promising strategy for anti-icing materials owing to their water repellent 20 – 23 . Many types of research have reported the efficient passive anti-icing methods using a superhydrophobic phenomenon for reducing adhesion force 24 – 32 or delaying freezing time 33 – 35 . In this study, we critically examined the facile strategy for attaining icephobicity on the superhydrophobic surface and compared it with a wide range of wettability to point out the effect of surface energy in anti-icing performance. The micro-nano hierarchical structure was generated on as-received Al plate through wet etching and followed by a low surface energy material coating to enable a perfect water repellent surface and nucleation inhibitor as well. Experimental results were compared with ongoing researches to figure out the contribution of surface wettability on anti-icing performance. Also, we proposed a calculation model to determine the freezing time. Freezing time experiments carried out on all samples were compared with theoretical calculation and revealed a good agreement, demonstrating the appropriate model for designing icephobic surfaces. The originality of this work is the experimental demonstration of the anti-icing performance on the superhydrophobic surface and proposing comprehensive insight into icing phenomena for icephobic applications.",
"discussion": "Results and discussion The ability to imitate the lotus leaf micro-nano structure enables the manufacturing of a superhydrophobic surface with extraordinary water repellence. The superhydrophobic concept, therefore, has been widely developed for icephobic applications owing to its beneficial properties including drag reduction and self-cleaning ability. However, recent studies on textured superhydrophobic surfaces have revealed that the performance largely limited by environmental constraints while the system cannot prevent the ice nucleation or obstruct the frost accumulation on surface textures. Besides, the voids between surface features might serve as vulnerabilities under extreme humidity conditions, results in the interlocking effect and eventually comparable adhesion strength on superhydrophobic and superhydrophilic substrates. To basically investigate the contribution of wettability in anti-icing performance, a wide range of contact angles was examined against the corresponded adhesive strength. Figure 1 disclosures the linear relation between surface wettability and adhesive strength. Our data were also compared with some ongoing research with the same interest and indicated the relatively same tendency as the higher contact angle sample exhibited the significant low adhesion compared to higher surface energy samples. The lowest value belongs to the superhydrophobic sample with about 135 kPa, which is % lower than 145° sample and about 9 times lower when compared with the superhydrophilic surface. The reason can be attributed to the contact area between ice and surface as the adhesion strength is attributed to the electrostatic force between molecules at the interface. Hence, the lower the contact area we can achieve, the lower the adhesion we can have. The measured adhesive strength gradually increases as the decrease of water-surface apparent contact angle. Figure 1 The original version of Figure 8. The advantages of the textured structure also can be emphasized through the reduction ratio (Fig. 2 ). The measurement on treated samples was compared to the one on as-received Al plate and essentially demonstrates the necessity of surface functionalization process for passive anti-icing applications. Almost textured samples propose a high reduction ratio even maintain a high contact area than the reference sample. The higher the contact angle we can yield, the higher the reduction ratio we can achieve. The higher surface energy i.e. lower contact angle corresponds to the higher affinity for water, finally results in the spreading form of a water droplet on the surface instead of forming a like-spherical droplet. This formation maintains in the whole freezing process so it eventually leads to higher ice-surface contact area. These results once again reinforced the importance of contact area in optimizing anti-icing effectiveness for icephobic applications. It should be noted here that the contact area parameter only viewed as an index factor when we consider surfaces with the same coating material. Figure 2 The reduction ratio of adhesion strength to as-received Al. Ice often accumulates on a surface when water comes in contact with the surface that is at temperatures below the freezing point. This process consumes the energy, hence it is worth considering the work of the adhesion parameter, which refers to the work that must be done to separate two adjacent phases. Figure 3 shows the distribution of adhesive strength against the work of adhesion calculated from our data and relevant research. The tendency also demonstrates the linear correlation between two investigated terms and once essentially proves the guiding role of the work of adhesion parameters in anti-icing applications. It should be noted here that the discussed results were collected from different experiments working on disparate types of materials including polymer, aluminum, copper but surprisingly illustrated the same tendency. Figure 3 The work of adhesion in correlation with adhesive strength. It is well known that the freezing process occurs by heat transfer from water volume to the cold substrate. However, to our best knowledge, there was no theoretical model to specifically predict the freezing time for an anti-icing experiment. In this work, we proposed a heat transfer model that aimed to investigate the whole freezing process and determine the freezing time. For ease understanding, we assume that water droplet forming on a sub-temperature surface in like-spherical shape in Fig. 4 . We first begin by considering freezing a drop on a flat surface. Figure 4 The theoretical model of a water droplet with a contact angle higher than 90° (left), and lower than 90° (right). When the drop is placed on a flat and sub-cooled surface as shown in Fig. 4 , the water drop starts to freeze from the contact region and the ice-water interface propagates until the water above it changes to ice. Here, the heat transfer rate q from the interface to the flat surface is 1 \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$q = \\mathop \\int \\limits_{0}^{h} k\\Delta T\\frac{A}{dh} = k\\Delta T\\left( {\\mathop \\int \\limits_{0}^{h} \\frac{1}{A}{\\text{d}}y} \\right)^{ - 1}$$\\end{document} q = ∫ 0 h k Δ T A dh = k Δ T ∫ 0 h 1 A d y - 1 where \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\Delta T = T_{i} - T_{s}$$\\end{document} Δ T = T i - T s , k is the conductive coefficient of ice, and A is the interface area at the dt examined moment. They can be defined as \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$A = \\pi r^{2} = \\pi R^{2} \\sin^{2} \\varphi ;\\quad dy = R\\sin \\varphi d\\varphi ;\\quad R = \\frac{d}{\\sin \\theta }$$\\end{document} A = π r 2 = π R 2 sin 2 φ ; d y = R sin φ d φ ; R = d sin θ Because of the heat transfer from the air to water during freezing is so small due to low heat transfer coefficient of air, heat transfer rate q will induce the phase change of the water above the interface to solid-phase: 2 \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$q = \\rho AL\\frac{{{\\text{d}}h}}{{{\\text{d}}t}}$$\\end{document} q = ρ A L d h d t With \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$dh = \\tan \\varphi dr$$\\end{document} d h = tan φ d r , \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\rho$$\\end{document} ρ is the density of ice, L is the coefficient of latent heat of fusion, and h is the position of the water–ice interface from the substrate base. Combining Eq. ( 1 ) and ( 2 ) leads to the freezing time of a drop on a surface: 3 \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$dt = \\frac{\\rho AL}{{k\\Delta T}}\\mathop \\int \\limits_{0}^{h} \\frac{1}{A}dydh$$\\end{document} d t = ρ A L k Δ T ∫ 0 h 1 A d y d h Then we have 4 \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$t\\frac{k\\Delta T}{{\\rho Ld}} = \\mathop \\int \\limits_{\\pi - \\theta }^{\\pi } \\frac{{d^{2} \\sin^{2} \\varphi }}{{\\sin^{2} \\theta }}\\left( {\\mathop \\int \\limits_{{R\\sin \\left( {\\pi - \\theta } \\right)}}^{R\\sin \\varphi } \\frac{1}{{\\left. {r^{2} } \\right|_{y = h} }}dr} \\right)d\\varphi$$\\end{document} t k Δ T ρ L d = ∫ π - θ π d 2 sin 2 φ sin 2 θ ∫ R sin π - θ R sin φ 1 r 2 y = h d r d φ \nfor the water droplet with the apparent contact angle higher than 90°, and 5 \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$t\\frac{k\\Delta T}{{\\rho Ld}} = \\mathop \\int \\limits_{\\theta }^{0} \\frac{{d^{2} \\sin^{2} \\varphi }}{{\\sin^{2} \\theta }}\\left( {\\mathop \\int \\limits_{R\\sin \\theta }^{R\\sin \\varphi } \\frac{1}{{\\left. {r^{2} } \\right|_{y = h} }}dr} \\right)d\\varphi$$\\end{document} t k Δ T ρ L d = ∫ θ 0 d 2 sin 2 φ sin 2 θ ∫ R sin θ R sin φ 1 r 2 y = h d r d φ \nfor the water droplet with an apparent contact angle lower than 90°. Where d is the diameter of the contact area between drop and substrate. The left-hand side of the equation is the non-dimensional freezing time in terms of the material properties of ice ( k , \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$\\rho ,L$$\\end{document} ρ , L ) and measured freezing time t , and the right-hand side represents the term that can be solely determined by the final shape of the freezing drop and initial equilibrium contact angle. Figure 5 describes the theoretical calculation of the non-dimensional freezing time fitting against the experiment results and demonstrates the good agreement. The higher contact angle ensures a higher non-freezing time in a logarithm correlation. Samples in hydrophilic and hydrophobic ranges were both investigated and indicated the strong dependence on surface wettability. This can be explained qualitatively by the project contact area between the water droplet and the cold surface. The large contact area induces a significant high heat transfer rate and described quantitatively using our approach. Of course, the freezing time will be zero when liquid completely spreads out the surface and reaches infinity when the contact angle is 180 degrees ( d = 0 ). It should be noted here that we neglected convective effect inside the water droplet and the radiation between the water droplet and ambient air due to a short time experiment and main heat attributed to the conductive heat. Our calculation proposed the appropriate approach for calculating the freezing time of a sole water droplet on a sub-temperature substrate in a wide range of surface wettability and contributed to the designing of an icephobic surface. Figure 5 Theoretical calculation of the non-dimensional freezing time and the experiment results. In this conclusion, we proposed a facile method to prepare the superhydrophobic surface on Al plates for anti-icing purposes. The treated sample exhibited extremely high anti-icing performance in terms of adhesion strength and freezing time. Our results were compared with ongoing research works and demonstrated the relative agreement when high water contact angle and small work of adhesion ensured the low ice-surface adhesion. Furthermore, we presented a theoretical method to calculate the freezing time from the heat transfer approach. Experimental results were compared with theoretical prediction and described the good agreement, illustrating the correctness when considering the freezing time. This insight should lead to an understanding of icing phenomena and the design of icephobic surfaces."
} | 4,075 |
38438350 | PMC10912231 | pmc | 43 | {
"abstract": "Artificial Intelligence (AI) is currently experiencing a bloom driven by deep learning (DL) techniques, which rely on networks of connected simple computing units operating in parallel. The low communication bandwidth between memory and processing units in conventional von Neumann machines does not support the requirements of emerging applications that rely extensively on large sets of data. More recent computing paradigms, such as high parallelization and near-memory computing, help alleviate the data communication bottleneck to some extent, but paradigm- shifting concepts are required. Memristors, a novel beyond-complementary metal-oxide-semiconductor (CMOS) technology, are a promising choice for memory devices due to their unique intrinsic device-level properties, enabling both storing and computing with a small, massively-parallel footprint at low power. Theoretically, this directly translates to a major boost in energy efficiency and computational throughput, but various practical challenges remain. In this work we review the latest efforts for achieving hardware-based memristive artificial neural networks (ANNs), describing with detail the working principia of each block and the different design alternatives with their own advantages and disadvantages, as well as the tools required for accurate estimation of performance metrics. Ultimately, we aim to provide a comprehensive protocol of the materials and methods involved in memristive neural networks to those aiming to start working in this field and the experts looking for a holistic approach.",
"introduction": "Introduction The development of sophisticated artificial neural networks (ANNs) has become one of the highest priorities of technological companies and governments of wealthy countries, as they can boost the fabrication of artificial intelligence (AI) systems that generate economic and social benefits in multiple fields (e.g., logistics, commerce, health care, national security, etc.) 1 . ANNs are able to compute and store the huge amount of electronic data produced (either by humans or machines), and to execute complex operations with them. Examples of electronic products that contain ANNs with which we interact in our daily lives are those that identify biometric patterns (e.g., face, fingerprint) for access control in smartphones 2 or online banking apps 3 , and those that identify objects in images from social networks 4 and security/traffic cameras 5 . Beyond image recognition, other examples are the engines that convert speech to text in computers and smartphones 6 , natural language processing as for example the novel automated chat system chat-GPT 7 , and those that provide accurate recommendations for online shopping based on previous behaviours from ourselves and/or people in our network 8 . ANNs can be understood as the implementation of a sequence of mathematical operations. The structure of ANNs consists of multiple nodes (called neurons) interconnected to each other (by synapses), and the learning is implemented by adjusting the strength (weight) of such connections. Modern ANNs are implemented via software in general-purpose computing systems based on a central processing unit (CPU) and a memory —the so-called Von Neumann architecture 9 . However, in this architecture a large amount of the energy consumption and computing time is related to continuous data exchange between both units, which is not efficient. The computing time can be accelerated by using graphics processing units (GPUs) to implement the ANNs (see Fig. 1a ), as these can perform multiple operations in parallel 10 – 12 . However, this approach consumes even more energy, which requires large computing systems and thereby cannot be integrated in mobile devices. Another option is to use field programable gate arrays (FPGAs), which consume much less energy than GPUs while providing an intermediate computing efficiency between CPUs and GPUs 13 – 17 . A survey carried out by Guo et al. 18 on the existing hardware solutions for ANN implementation and their performance is condensed in Fig. 1b . Fig. 1 Computing power demand increase and platform transition from Von-Neumann towards highly parallelized architectures. a The increase in computing power demands over the past four decades expressed in petaFLOPS per days. Until 2012, computing power demand doubled every 24 months; recently this has shortened to approximately every 2 months. The colour legend indicates different application domains 10 . Mehonic, A., Kenyon, A.J. Brain-inspired computing needs a master plan. Nature 604, 255–260 (2022), reproduced with permission from SNCSC. b A comparison of neural network accelerators for FPGA, ASIC, and GPU devices in terms of speed and power consumption. GOP/s giga operations per second, TOP/s tera operations per second. In the past few years, some companies and universities have presented application specific integrated circuits (ASICs) based on the complementary metal oxide semiconductor (CMOS) technology that are capable to compute and store information in the same unit. This allow such ASICs to perform multiple operations in parallel very fast, making them capable of mimicking, directly in the hardware, the behaviour of the neurons and synapses in the ANN. A comprehensive list of these ASICs comprising those such as the Google TPU 19 , Amazon inferentia 20 , Tesla NPU 21 , etc., are summarized in ref. 22 . Such integrated circuits can be grouped in two categories. On one hand, dataflow processors are custom-designed processors for neural network inference and training. Since neural network training and inference computations can be entirely deterministically laid out, they are amenable to dataflow processing in which computations, memory accesses, and inter-ALU communications actions are explicitly/statically programmed or placed-and-routed onto the computational hardware. On the other hand, processor in memory (PIM) accelerators integrate processing elements with memory technology. Among such PIM accelerators are those based on an analogue computing technology that augments flash memory circuits with in-place analogue multiply-add capabilities. Please refer to the references for the Mythic 23 and Gyrfalcon 24 accelerators for more details on this innovative technology. Previously mentioned ANNs and those reported in detail in the survey presented in ref. 22 belongs to the subgroup of so-called deep neural networks (DNNs). In a DNN the information is represented with values that are continuous in time and can achieve high data recognition accuracy by using at least two layers of nonlinear neurons interconnected by adjustable synaptic weights 25 . Conversely, there is an alternative information codification which gave birth to another type of ANNs, the Spiking Neural Networks (SNN). In SNNs the information is coded with time-dependent spikes, which remarkably reduces the power consumption compared to DNNs 26 . Moreover, the functioning of SNNs is more similar to the actual functioning of biological neural networks, and it can help to understand complex mammal’s neural systems. Intel probably has the most extensive research program for evaluating the commercial viability of SNN accelerators with their Loihi technology 27 , 28 , and Intel Neuromorphic Development Community 29 . Among the applications that have been explored with Loihi are target classification in synthetic aperture radar and optical imagery 30 , automotive scene analysis 31 , and spectrogram encoder 27 . Further, one company, Innatera, has announced a commercial SNN processor 32 . Also, the platforms developed by IBM (TrueNorth 33 ), and Tsingshua 34 are well known examples of the research effort of both the industry and the academia in this field. However, fully-CMOS implementations of ANNs require tens of devices to simulate each synapse, which threatens energy and area efficiency, and thereby renders large-scale systems impractical. As a result, the performance of CMOS-based ANNs is still very far from that of biological neural networks. To emulate the complexity and ultra-low power consumption of biological neural networks, hardware platforms for ANNs must achieve an ultra-high integration density (>1 Terabyte per cm 2 ) and low energy consumption (<10 fJ per operation) 35 . Recent studies have proposed that the use of memristive devices to emulate the synapses may accelerate ANN computational tasks while reducing the overall power consumption and footprint 36 – 42 . Memristive devices are materials systems whose electrical resistance can be adjusted to two or more stable (i.e., non-volatile) states by applying electrical stresses 43 . Memristive devices that exhibit two non-volatile states are already being commercialized as standalone memory 44 , 45 , although their global market is still small (~621 million USD by 2020, i.e., ~0.5% of the 127-billion-worth standalone memory market 46 ). However, memristive devices can also exhibit three disruptive attributes particularly suitable for the hardware implementation of ANNs: i) the possibility to program multiple non-volatile states (up to ~100 47 , 48 , and even ~1000 49 ), ii) a low-energy consumption for switching (~10fJ per state transition with zero-static consumption when idle 50 ), and iii) a scalable structure appropriate for matrix integration (often referred to as crossbar 51 ) and even 3D stacking 52 . Moreover, the switching time can be as short as 85 ps 42 . So far, several groups and companies have claimed the realization of hybrid CMOS/memristor implementations of ANNs 53 – 61 , —from now on, memristive ANNs— with performance that is superior to that of fully-CMOS counterparts. However, most of those studies in fact only measured the figures-of-merit of one/few devices and simulated the accuracy of an ANN via software 62 – 67 in such type of studies the connection between the memristors fabricated and the ANN is relatively weak. Few studies went beyond that and built/characterized crossbar arrays of memristive devices 48 , 68 – 70 , but that are still very far from real full-hardware implementations of all the mathematical operations required by the ANN. The most advanced studies in this field have reported fully integrated memristor-based compute-in-memory systems 48 , 53 – 55 , 58 , 59 , 61 , 71 – 73 , but a systematic description of essential details on the device structure or circuit architecture are generally lacking in these reports. In this article we provide a comprehensive step-by-step description of the hardware implementation of memristive ANNs for image classification —the most studied application often used to benchmark performance, describing all the necessary building blocks and the information processing flow. For clarity, we consider relatively simple networks, being the multilayer perceptron the most complex case. We take into account the challenges arising at both the device and circuit levels and discuss a SPICE-based approach for their study in the design stage, as well as the required circuital topologies for the fabrication of a memristive ANN."
} | 2,785 |
29727441 | PMC5935395 | pmc | 44 | {
"abstract": "Fish, birds, insects and robots frequently swim or fly in groups. During their three dimensional collective motion, these agents do not stop, they avoid collisions by strong short-range repulsion, and achieve group cohesion by weak long-range attraction. In a minimal model that is isotropic, and continuous in both space and time, we demonstrate that (i) adjusting speed to a preferred value, combined with (ii) radial repulsion and an (iii) effective long-range attraction are sufficient for the stable ordering of autonomously moving agents in space. Our results imply that beyond these three rules ordering in space requires no further rules, for example, explicit velocity alignment, anisotropy of the interactions or the frequent reversal of the direction of motion, friction, elastic interactions, sticky surfaces, a viscous medium, or vertical separation that prefers interactions within horizontal layers. Noise and delays are inherent to the communication and decisions of all moving agents. Thus, next we investigate their effects on ordering in the model. First, we find that the amount of noise necessary for preventing the ordering of agents is not sufficient for destroying order. In other words, for realistic noise amplitudes the transition between order and disorder is rapid. Second, we demonstrate that ordering is more sensitive to displacements caused by delayed interactions than to uncorrelated noise (random errors). Third, we find that with changing interaction delays the ordered state disappears at roughly the same rate, whereas it emerges with different rates. In summary, we find that the model discussed here is simple enough to allow a fair understanding of the modeled phenomena, yet sufficiently detailed for the description and management of large flocks with noisy and delayed interactions. Our code is available at http://github.com/fij/floc .",
"introduction": "Introduction: Collective motion in 2 and 3 dimensions In all fields of life recent technological developments have lead to a surge in data acquisition. However, usually the obtained data can be put to practical use only with improved analytic and predictive methods. For collective motion (swarming, active matter), some of the recent major experimental advances have been the systematic measurements of fish trajectories in small shoals [ 1 ], tracking the individual coordinates of up to 2700 birds in flocks [ 2 ], and obtaining GPS track logs of homing pigeons flying together [ 3 ]. Initially, experimental and modeling efforts were focused on planar (2 dimensional) motion. Due to these efforts it is now well known that in 2 dimensions bacteria, insects, horses, and also humans display collective motion patterns [ 4 – 7 ]. Compared to planar motion, an agent moving in space can be kept aligned by a higher number of nearest neighbor interactors. At the same time, it has also more directions to turn away from the consensus of those nearest neighbors. The most straightforward local rule that can describe the alignment of moving agents with their neighbors is to set each agent’s direction of motion explicitly to the average direction of its neighbors [ 8 , 9 ]. Turning continuously toward the average direction of the neighbors is also possible [ 10 , 11 ]. More detailed mechanisms of the alignment include anisotropic interactions caused by elongated shapes [ 12 – 15 ], also combined with a frequent reversal of the direction of motion [ 16 ], the preference for movements in the horizontal plane (as opposed to vertical movements) [ 17 ], a viscous medium [ 18 ], friction among the agents and inelastic collisions [ 19 , 20 ], and sticking together [ 21 ]. While these rules can set the direction of motion for the agents, collision avoidance and cohesion (staying together) are also necessary for flock formation. To avoid the collisions of moving and interacting agents the simplest solution is to let all agents (in the model) have zero size [ 8 ]. A more realistic solution is a strong short-range repulsive interaction, in which the magnitude of the repulsion force becomes very high when two agents come too close. Finally, for keeping the group together two commonly applied modeling tools are the spatial confinement of the group (e.g., periodic boundaries) and a weak attraction that is turned on when distances between the agents grow. The model that we discuss here focuses on controlling the speed of the agents individually . The speed of an agent is adjusted to the preferred speed with a rate that is proportional to the difference from the preferred value (see Fig 1 ). This modeling approach is realistic, because—according to recent experimental and modeling evidence—individual speed control plays a key role in the formation and the stability of bird flocks and fish shoals [ 1 , 22 , 23 ]. Also, experiments and models for vibrated self-propelled hard disks have shown that the binary collisions caused by maintaining speed can align velocity vectors first locally, and then also across the entire system [ 24 ]. 10.1371/journal.pone.0191745.g001 Fig 1 The spatial flocking model used in the paper. Image from thedroneinfo.com. We investigate the effect of time delay and noise, too. Time delay is a common phenomenon caused by latent communication between agents, information processing cost, and inertial reasons [ 25 , 26 ]. Noise at all levels is also inherent to the communication and decisions of all moving agents in a dynamic system and can lead to transitions between behavioral patterns [ 26 – 29 ]. Regarding the combination of time delay and noise, simulation results in [ 30 ] showed that a system with noise and delay displays bistability of several coherent patterns. Here we investigate both aspects and show their fundamental dissimilarities.",
"discussion": "Discussion and outlook In this paper, we investigated a minimal continuous model of 3-dimensional collective motion. The model contains continuous adjustment of particle speed to a preferred value, pairwise radial repulsion for collision avoidance, and an effective weak attraction (periodic boundaries). We found that the combination of these three model components is sufficient for stable spatial ordering, and beyond these three no further model components are necessary. After investigating the model on the microscopic level, we found that for the majority of symmetric two-agent encounters the total momentum of the two agents increases. Regarding the macroscopic level, we found that the transition from disorder to order is fast for both small and large system sizes, which is in good agreement with previous results [ 27 ]. In the minimal continuous model that we investigated we found also that if the noise intensity is above a threshold value, then the system cannot reach the ordered state, similarly to results reported in [ 36 ]."
} | 1,708 |
35792877 | PMC9871526 | pmc | 45 | {
"abstract": "Abstract Microbial production of biopolymers derived from renewable substrates and waste streams reduces our heavy reliance on petrochemical plastics. One of the most important biodegradable polymers is the family of polyhydroxyalkanoates (PHAs), naturally occurring intracellular polyoxoesters produced for decades by bacterial fermentation of sugars and fatty acids at the industrial scale. Despite the advances, PHA production still suffers from heavy costs associated with carbon substrates and downstream processing to recover the intracellular product, thus restricting market positioning. In recent years, model‐aided metabolic engineering and novel synthetic biology approaches have spurred our understanding of carbon flux partitioning through competing pathways and cellular resource allocation during PHA synthesis, enabling the rational design of superior biopolymer producers and programmable cellular lytic systems. This review describes these attempts to rationally engineering the cellular operation of several microbes to elevate PHA production on specific substrates and waste products. We also delve into genome reduction, morphology, and redox cofactor engineering to boost PHA biosynthesis. Besides, we critically evaluate engineered bacterial strains in various fermentation modes in terms of PHA productivity and the period required for product recovery.",
"conclusion": "CONCLUSIONS AND FUTURE DIRECTIONS Over the past decades, genetic interventions in natural PHA producers have resulted in superior cell factories with varying degrees of success (Tables 1 & 2 ). Most of the studies in the literature present the increased PHA synthesis as the weight percentage of the cell dry weight (wt%). However, enhancing the product titre, yield, and PHA‐specific volumetric productivity (TYP) is imperative and a cornerstone in industrial biotechnology. In this sense, high PHA productivities have been achieved by engineered C. necator , P. putida , and Halomonas spp. employing glucose, sugar hydrolysates, crude glycerol, and waste vegetable oils (Orita et al., 2012 ; Poblete‐Castro et al., 2014a ; Ye et al., 2018 ). Most of these strategies relied on the overexpression of the native PHA biosynthetic genes, replacement of more efficient PHA polymerases, and the blockage of carbon fluxes towards competing pathways (Figure 3 ). Less successful has been the bioconversion of CO 2 , lignin hydrolysates, PET derivatives into PHA and co‐production in microbial consortia, which reached biopolymer titers lower than 2 (g L −1 ) in several days given the toxic effect of some feedstocks or the slow uptake capacity of the host strains to metabolize inexpensive substrates. Overall, engineered microbes are still behind PHA productivities attained by natural PHA producers such as C. necator (3.1 g L −1 h −1 ) (Shang et al., 2003 ), P. putida (2.3 g L −1 h −1 ) (Maclean et al., 2008 ), and P. megaterium (1.7 g L −1 h −1 ) (Kanjanachumpol et al., 2013 ). Among strategies, adaptive laboratory evolution appears as a powerful tool to generate evolved lineages of cells that overcome growth arrest and low biomass yields (Portnoy et al., 2011 ). Resequencing more resistant bacterial strains might enable unveiling the genetic changes that confer such features and contribute the most to the desired phenotype. This reverse engineering approach combined with genome‐scale metabolic modelling (GSMM) might deliver novel targets to optimize PHA synthesis (Sandberg et al., 2019 ). The elucidation of carbon partitioning through multiple metabolic pathways and intracellular modulators that preclude PHA accumulation in the cell can also benefit from using GSMM and regulatory models. For example, during genome reduction of PHA‐producing bacteria many regulatory elements of unknown role in the PHA synthesis process were eliminated (Fan et al., 2020 ; Liang et al., 2020 ). Also, cofactor requirement is a crucial element when producing PHAs from waste streams as NADPH is involved in both fighting oxidative stress during replication and as a cofactor in many PHA biosynthetic pathways. Concerning programmable cell lysis systems, lysozyme‐based and phage holin‐endolysin enzymes have been implemented in bacterial strains for cell disruption and biopolymer recovery. Unfortunately, some of these genetic circuits arrest cell growth affecting the overall PHA formation. So far, the created genetic circuits exhibit a certain level of leakage. And activating the programable circuits in nitrogen‐limiting conditions compromises protein synthesis resulting in delayed cell disruption. Thus, further developing tightly regulated genetic circuits (on/off mode of action) that can be integrated into the chromosome of the biocatalyst might avoid antibiotic supplementation eluding unwanted cell growth impairment. Moreover, activating the cell lytic systems in high cell density cultures remain a challenge and it is an open question whether these genetic constructs function under such conditions. Finally, another important aspect poorly covered in studies aiming to engineer PHA producers is the characterization and comparison of the biosynthesized PHA in terms of thermal and mechanical features. The polymerization process, catalysed by the PHA polymerases accompanied by the PHA phasin protein, defines in part the molecular weight of the formed biopolymer (Shen et al., 2019 ; Tian et al., 2005 ; Tsuge, 2016 ). Besides, fermentation conditions also influence the physical properties of the intracellular polyoxoester (McAdam et al., 2020 ). There is an interconnection between genetic manipulations and bioreactor settings that modulate the PHA polymerization process and possible applications. How the alteration of the genetic repertory, metabolic operation, or regulatory elements affect the physical properties of the resulting PHA is still elusive. More than ever, there is a need to establish more efficient bioconversion processes of renewable materials into biopolymers for tackling the global plastic problem where metabolic engineering, synthetic biology, and mathematical modelling of cellular processes play a pivotal role in this endeavour.",
"introduction": "INTRODUCTION Petrochemical plastics are increasingly accumulating in soil and aquatic environments, with recent studies demonstrating that nearly 80% of the ever‐produced plastics are intact across the planet (Geyer et al., 2017 ). A sustainable alternative to deriving polymers from renewable carbon sources is the use of microbial cell factories (Choi et al., 2020 ). Among them, polyhydroxyalkanoates (PHAs) are polyoxoesters that accumulate intracellularly as carbon and energy form in bacteria and archaea (Wilkinson, 1963 ) when they encounter famine and stress conditions (Mason‐Jones et al., 2021 ). They also display similar mechanical and physical properties to various synthetic plastic like polypropylene and polystyrene (Van de Velde & Kiekens, 2002 ) and proven biodegradable in landfills, soil, and water ecosystems (Meereboer et al., 2020 ). Some microbes can amass more than 90% on a cell mass basis as PHA, and the hydrophobic granules are prone to hydrolysis to satisfy cell's metabolic demands (Madison & Huisman, 1999 ). Initially described as a nongrowth associated intracellular product where the limitation of an inorganic nutrient is necessary to trigger PHA accumulation (Sudesh et al., 2000 ), yet several studies showed that in some environmental microbes, PHAs are intrinsic parts of the cell components during replication (Godard et al., 2020 ; Poblete‐Castro et al., 2012a ; Shrivastav et al., 2010 ). PHA polymer molecular structures display desired technological features such as water, oxygen and flavour barriers (Follain et al., 2014 ) along with suitable melting points and molecular weights for manufacturing containers and packaging materials (Israni & Shivakumar, 2019 ). Despite these advantages, the global polymer market still relies on materials originated from petrochemical sources, accounting for 99% of the worldwide production volume (Chen et al., 2020 ). In fact, these polyoxoesters share nearly 1% of the biopolymer market. In this review, we underline the principles and rationale behind metabolic and genetic engineering strategies to enhance PHA production in natural biopolymer producers belonging to different bacterial genus and archaea. The construction of novel pathways to generate not naturally occurring metabolites of these biocatalysts' metabolic networks is also highlighted. We also delve into designing strategies to program cell autolysis and discuss the lytic system's efficiency during PHA accumulation."
} | 2,168 |
30966026 | PMC6418807 | pmc | 46 | {
"abstract": "It is not unusual for humans to be inspired by natural phenomena to develop new advanced materials; such materials are called bio-inspired materials. Interest in bio-inspired polymeric superhydrophilic, superhydrophobic, and superoleophobic materials has substantially increased over the last few decades, as has improvement in the related technologies. This review reports the latest developments in bio-inspired polymeric structures with desired wettability that have occurred by mimicking the structures of lotus leaf, rose petals, and the wings and shells of various creatures. The intrinsic role of surface chemistry and structure on delivering superhydrophilicity, superhydrophobicity, and superoleophobicity has been extensively explored. Typical polymers, commonly used structures, and techniques involved in developing bio-inspired surfaces with desired wettability are discussed. Additionally, the latest applications of bio-inspired structures with desired wettability in human activities are also introduced.",
"conclusion": "5. Conclusions and Perspective This review article summarized the recent development of bio-inspired superhydrophilic, superhydrophobic, and superoleophobic structures inspired by plant leaves (e.g., the lotus leaf and the rose petal) and creatures (e.g., butterfly wings, water strider legs, and shark skin) that have employed a variety of polymers and techniques. By reviewing the theory of wettability, a series of studies that investigate the effects of surface energy (low/high surface energy) and surface structures (nano-/micro-structure, hierarchical structure) on delivering desired wettability has been launched. The techniques used for developing surfaces with desired wettability are similar, but the requirements for surface energy to delivery superhydrophilicity, superhydrophobicity and superoleophobicity are different. Oil has a much lower surface tension and can easily wet superhydrophobic surfaces. Superhydrophilic surfaces require an extremely high surface energy and a proper structure, while superhydrophobic and superoleophobic surfaces require a low surface energy. Repelling oils for superoleophobic surfaces especially requires a critically low surface energy. Bio-inspired surfaces with special wettability are applied widely in oil/water separation, anti-wear, anti-fouling, self-cleaning, friction reduction, drug release, adhesion reduction, etc. With the rapid progress in technology, bio-inspired super-wettable surfaces with advanced properties are in high demand. However, the current development on constructing bio-inspired super-wettable materials has disadvantages. Most polymeric bio-inspired materials lack sufficient mechanical properties and advanced multifunctional properties; contaminates on surfaces cause a loss of function and lifetime; the ease of damage and defects in current construction techniques as well as short lifetimes increase its production costs. In the near future, it is crucial to develop bio-inspired super-wettable materials with high mechanical properties and multifunctional performances with simple techniques to enhance their lifetime and durable wettability.",
"introduction": "1. Introduction Wetting phenomena can normally be found in nature and are commonly seen in our daily life. Wetting on solid surfaces can be induced either by water (hydrophobicity) or oils (oleophobicity). One classic case of wetting phenomena is defined as superhydrophobicity [ 1 , 2 , 3 , 4 ], where extremely high water repellency is found on the surface. Rice leaves illustrate the wetting phenomenon in nature, where water droplets bead up rest on the surface without actually wetting it ( Figure 1 a) [ 1 ]. Another classic wetting phenomena is defined as superhydrophilicity, where water wets and spreads over the surface quickly [ 1 ]. Clothes illustrate superhydrophilicity, where fabrics are completely wetted by water, leaving water stains on the surface ( Figure 1 b) [ 5 ]. This is defined as superhydrophilicity [ 4 ]. In real applications, superhydrophobic surfaces can be seen everywhere, e.g., water droplets ball up on transparent and superhydrophobic glass slides ( Figure 1 c) [ 6 ] and roll down, carrying away particles and leaving clean traces on glass (red rectangles) ( Figure 1 d) [ 6 ]. For technical applications, wettability is an essential property in painting [ 7 , 8 ], printing [ 9 , 10 , 11 ], anti-fogging [ 12 , 13 , 14 ], anti-fouling [ 15 , 16 , 17 , 18 ], transportation [ 19 , 20 , 21 , 22 ], waterproof products [ 23 , 24 , 25 ], oil recovery [ 26 , 27 , 28 ], anti-corrosion [ 29 , 30 , 31 ], water recycling, etc. [ 4 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ] On the micro-/nanoscale, wettability affects micromachining, such as microfluidic channels [ 40 , 41 , 42 ], nanoprinting [ 43 , 44 ], and lab-on-a-chip systems [ 45 , 46 ]. Inspired by the wetting phenomena in nature, a variety of bio-inspired materials with superhydrophilic/superhydrophobic properties have been developed. An important bio-inspired superhydrophobic structure with high water repellence is inspired by the lotus leaf [ 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ], which has been found to play critical roles in nature, human daily life, and industry. It has been well accepted that the hierarchical micro-/nanostructure and the low surface energy are responsible for the superhydrophobicity and excellent self-cleaning properties [ 2 , 47 ]. To date, several classical theories have been devised to comprehensively understand the intrinsic principles behind wetting phenomena, including the Wenzel theory and the Cassie–Baxter theory [ 2 , 47 ]. So far, a series of bio-inspired superhydrophobic structures [ 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 ] have been developed for self-cleaning [ 48 , 49 , 50 , 51 ], anti-biofouling [ 15 , 16 , 17 , 18 , 52 , 53 ], anti-icing [ 54 , 55 , 56 ], anti-corrosion [ 31 , 57 ], and adhesion reduction [ 58 , 59 ], etc. Meanwhile, enormous bio-inspired superhydrophilic structures have been fabricated and are similar to bio-inspired superhydrophobic structures but have a relatively high surface energy [ 1 , 2 , 4 , 14 , 35 , 43 ]. Nevertheless, liquid in real cases (such as oil transportation, oil pumps, and wastewater treatment) usually contains oily components, which easily wet the superhydrophobic surfaces [ 60 , 61 ]. Thus, the fabrication of superhydrophobic surfaces with both water and oil repellency (hydrooleophobicity) will meet the increasing demand of bio-nanotechnology and will largely expand their potential applications. The design of oleophobic surfaces follows the same principles for fabricating superhydrophobic surfaces, but with a particular emphasis on the reduction of surface energy in combination with surface structures. Fluoropolymers with high amounts of function groups, such as –CF 3 and –CF 2 [ 35 , 36 , 37 ], have the lowest surface energy, thus fluoropolymers have been commonly applied to reduce the surface energy of solid surfaces. Additionally, the trapped air pockets in the patterned structures prevent the penetration of liquid, which contributes to superoleophobicity [ 36 , 37 ]. Thus far, fluoropolymers with modified superoleophobic surfaces (nano-/micro-/hierarchical structure [ 37 , 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 , 70 , 71 ], micropillars [ 60 , 66 , 72 , 73 , 74 , 75 ], etc.) have been developed. The typical surface modification approaches in the fabrication of superoleophobic materials are chemical vapor deposition [ 74 , 76 ], electrospinning [ 70 ], dip-coating [ 77 , 78 , 79 , 80 , 81 , 82 , 83 ], and plasma treatment to graft the fluoropolymer to the surface [ 71 , 84 , 85 ]. Over the past few decades, a variety of superhydrophobic, superhydrophilic, and superoleophobic structures have been designed to meet performance requirements using copper [ 86 ], aluminum [ 87 , 88 ], stainless steel [ 89 , 90 ], silica [ 91 ], silicon [ 92 , 93 ], and so on. Nevertheless, metals are difficult for constructing nano-/hierarchical structures since they can easily suffer from defects and corrosion [ 94 ]. Inorganic materials are normally lack of sufficient mechanical properties and the reduction of the surface energy is challenging [ 95 , 96 , 97 , 98 ]. Therefore, using simple methods to fabricate flexible polymers with desired wettability and properties are regarded as an optimum option. This review mainly focuses on the development of bio-inspired polymeric superhydrophilic, superhydrophobic, and superoleophobic micro-/nanostructures, the techniques involved, and applications. The development of materials with desired wettability and multifunctional properties are also discussed."
} | 2,168 |
34986278 | PMC9303884 | pmc | 47 | {
"abstract": "Abstract Biotechnological production is a powerful tool to design materials with customized properties. The aim of this work was to apply designed spider silk proteins to produce Janus fibers with two different functional sides. First, functionalization was established through a cysteine‐modified silk protein, ntag Cys eADF4(κ16). After fiber spinning, gold nanoparticles (AuNPs) were coupled via thiol‐ene click chemistry. Significantly reduced electrical resistivity indicated sufficient loading density of AuNPs on such fiber surfaces. Then, Janus fibers were electrospun in a side‐by‐side arrangement, with “non‐functional” eADF4(C16) on the one and “functional” ntag Cys eADF4(κ16) on the other side. Post‐treatment was established to render silk fibers insoluble in water. Subsequent AuNP binding was highly selective on the ntag Cys eADF4(κ16) side demonstrating the potential of such silk‐based systems to realize complex bifunctional structures with spatial resolutions in the nano scale."
} | 249 |
34389536 | PMC8363143 | pmc | 48 | {
"abstract": "Probiotics mitigate post-heat stress disorder, preventing coral mortality.",
"introduction": "INTRODUCTION Coral reefs have been undergoing unprecedented mass coral bleaching events in recent decades, fueled by ocean warming ( 1 ), heightening the need to devise effective countermeasures to mitigate further declines ( 2 , 3 ). Increasing sea surface temperatures trigger the disruption of the symbiotic relationship between the coral host and its endosymbiotic algae of the family Symbiodiniaceae ( 4 ), resulting in the physical whitening of coral colonies known as “bleaching.” Photosynthetic products from the endosymbiont algae provide more than 90% of the host’s nutritional demands ( 5 ). Thus, prolonged periods of heat stress and bleaching lead to coral mortality ( 6 ). Besides endosymbiotic algae, corals are associated with a suite of other organisms (bacteria, protists, fungi, viruses, etc.), collectively referred to as the coral holobiont or metaorganism ( 7 – 10 ). In particular, bacteria are assumed to contribute to coral holobiont biology, notably stress tolerance and adaptation to disparate environments ( 10 – 15 ). The importance of bacteria led to the proposal of the coral probiotic hypothesis ( 16 ), which states that microbes support coral biology through selection of the most advantageous holobiont configuration in a given environment. This was later refined by the microbiome flexibility hypothesis to include the notion that the potential or propensity for microbiome change differs among host species ( 15 ). The proposal to use these concepts to select and manipulate specific microbes to aid the stress tolerance and resilience of the coral holobiont was dubbed “beneficial microorganisms for corals” (BMCs) ( 10 ). Beneficial microorganisms putatively support nitrogen fixation, sulfur cycling, scavenging reactive oxygen species (ROS), and production of antibiotics to thwart pathogens, for example ( 10 , 11 , 17 ). The proof of concept that manipulating coral microbes improves coral stress tolerance was recently demonstrated in the first experiments to identify the beneficial nature of a selected BMC consortium in ameliorating coral bleaching ( 18 ). Nevertheless, exactly “how” these BMCs were associated with functional changes in the host remained unknown. Notably, BMCs do not necessarily need to exert their effect on the coral host directly. Hence, the measured holobiont response does not need to be a perfect reflection of the BMC consortium added. Rather, the BMC consortium may benefit the host indirectly, by means of niche occupation, microbial succession, or the prevention of dysbiosis through pathogen deterrence ( 10 , 11 , 18 ). Furthermore, although the ability of BMCs to ameliorate coral bleaching has been demonstrated ( 18 ), it is unknown whether they have the capacity to help corals evade mortality, e.g., through the provisioning of alternate metabolites to compensate for the loss of Symbiodiniaceae. Despite the diversity of the coral microbiome, which makes it challenging to decipher the contribution of associated microbes to coral holobiont biology, the dynamic nature of the coral microbiome, which can often change markedly—e.g., across sites, species, age, and under stress—further hampers the ability to conduct such studies in the natural environment ( 15 , 19 – 21 ). For this reason, manipulation of BMCs in controlled experimental setups, such as mesocosms ( 22 – 24 ), provides an avenue to identify important microbial players and study holobiont responses (and putative underlying mechanisms), while maintaining a quasi-reef environment, to improve and inform the development of biotechnological solutions to promote coral reef resilience. Here, we used coral mesocosms in combination with multiomics evaluation to assess responses and potentially decipher the mechanisms that underlie the increased stress tolerance and coral mortality evasion, offered by the provisioning of probiotics. In a large-scale effort, fragments of the coral Mussismilia hispida were exposed to thermal stress in a 75-day mesocosm experiment and inoculated with either a M. hispida –tailored BMC consortium or a saline solution placebo. Coral health (measured via F v /F m rates and survivorship) ( 25 ), microbial activity, and functional responses were assessed through a multiomics approach. Our analysis shows that increased stress tolerance and survivorship of coral holobionts exposed to a BMC consortium coincided with holobiont restructuring and a defined reprogramming of the coral host’s gene expression, targeting cellular reconstruction, immune response, and stress protection during a post-heat stress recovery period.",
"discussion": "DISCUSSION The promise of coral probiotics to increase the stress tolerance of corals has been very recently shown ( 11 , 18 , 30 ), although the effect that BMCs exert on the holobiont or whether BMCs can increase survivability of corals under stress remained elusive. Here, we show that the inoculation of coral fragments with a native BMC consortium instigated holobiont changes at the level of the microbiome, host gene expression, and metabolism, which coincide with an increase in coral survival rates ( Fig. 5 ). Hence, our results provide a first insight into the putative mechanistic underpinnings of how the coral (host) responds to BMC inoculation, although the detailed functional changes that cause the altered phenotype await further elucidation. Our results argue for PHSD recovery improvement of the metaorganism by the BMC consortium, as indicated by changes at the coral host, Symbiodiniaceae, and bacterial compartment level. From the results obtained, a number of key findings emerge that we discuss in the following. Fig. 5 Probiotics-mediated mitigation of coral PHSD. Summary of the overall differential recovery mechanisms observed at the end of the 75-day mesocosm experiment, comparing the process in BMC-treated ( A ) and placebo-treated ( B ) M. hispida fragments. We observed major changes in microbial community structure observed during heat stress ( 12 , 31 , 32 ) in conjunction with the dynamic microbiome restructuring following the recovery period, indicating that M. hispida exhibits microbiome adaptation. Thus, it may fit into the “microbiome conformer” type previously suggested ( 14 , 15 ) and observed for this coral species regarding other impacts ( 33 – 35 ). Following this notion, the level of microbiome flexibility may be considered as a factor to identify corals with high(er) manipulative potential. Corals that naturally alter their microbial composition and potentially uptake microbes from the environment are more likely to “accept” inoculants ( 14 , 15 , 36 ). Notably, shifts in metaorganism microbial composition are, potentially, rapid and versatile means of adaptation to environmental change ( 12 – 15 ). It is important to consider that the host’s ability to take up microorganisms from the environment is hypothesized to increase when under stress, a conclusion based on the finding that many host microbiomes appear less ordered when stressed ( 14 , 21 , 37 ). Inoculation with high numbers of different BMC cells (i.e., a consortium) may therefore ensure (and improve) uptake of at least some microorganisms exhibiting beneficial characteristics, which may, at the same time, preclude colonization by pathogens considering that “space is limited.” The use of bacterial consortia provides a combination of beneficial mechanisms to increase stress tolerance, even if not all members of the BMC successfully associate with the coral holobiont ( 18 , 31 , 38 – 40 ). Here, we show that the use of a bacterial consortium resulted in incorporation of some of the selected BMCs, which were found in the microbiome of BMC-treated corals during the thermal stress, i.e., at T1 and T2 (see Fig. 2A ). Notably, members of the BMC consortium were not detected after the 15-day period (T3; i.e., recovery). This suggests three things: first, a dynamic restructuring of the microbiome can happen on a relatively small time scale ( 12 , 14 , 41 ); second, incorporation of BMCs might be facilitated under stress (in this experiment, during the peak of heat stress) because coral defense is compromised or selection for beneficial microbes is supported; and third, it is currently unclear how long the beneficial effect of BMCs is lasting. From our results, it appears that BMC members colonized coral fragments during stress and instigated significant changes in the coral holobiont but reverted to the original microbiome structure after ceasing (or the absence) of stress [sensu Ziegler et al . ( 14 ) who used the term “microbiome recovery”]. Accordingly, the duration of the presence of the stressor might determine the longevity of the BMC effect, which suggests that repeated addition of BMCs might be needed to ensure a long-lasting effect under natural conditions ( 11 ). The early and detectable incorporation of some of the BMC consortium members into the coral microbiome and the subsequent microbial restructuring were correlated with significant improvements in coral recovery after thermal stress, as most convincingly demonstrated by mortality evasion. Heat stress–driven mortality and/or decrease in F v /F m rates observed in fragments that were not treated with BMCs suggest damage to the temperature-related photosystem II electron transport of the Symbiodiniaceae through chronic photoinhibition ( 42 ), which ultimately leads to a breakdown in symbiosis and results in loss/expelling of the Symbiodiniaceae, i.e., bleaching ( 43 ). Notably, bleaching is a symptomatic phenotype, i.e., loss of Symbiodiniaceae can occur through multiple processes, including host cellular apoptosis ( 44 ) or necrosis, and eventually death from starvation ( 6 , 45 ), which was corroborated by the up-regulation of different kinases directly involved in triggering apoptosis in placebo-treated corals ( 46 , 47 ). Our transcriptome results indicate that BMCs did not buffer the immediate heat stress response in M. hispida but exerted its effect during recovery, supported by the gene expression patterns and coral physiology. Most notably, we observed low F v /F m rates for both BMC and placebo treatments at T2 (during heat stress), but only the BMC treatment promoted recovery at T3, as indicated by the “return-to-normal” F v /F m rates. Our interpretation is that BMCs exert their effect through mitigation of the effects from what we term PHSD. The molecular evidence for this condition includes not only apoptosis activity, which may be triggering inflammatory responses, but also membrane and cellular reconstruction due to tissue loss caused by recent-past heat stress. In this regard, the remaining placebo-treated surviving corals seemed to be still struggling from the effects of recent heat stress, even 19 days after the end of the heat stress period, while all BMC-treated corals seem to have recovered. As an analogy to posttraumatic stress disorder ( 48 ), coral PHSD is characterized by the contrast of the coral response and its attempts to recover from a heat stress event while still fading due to the cellular, immune, and metabolic consequences of such stress. The significant up-regulation of numerous kinases and receptors, as well as signaling molecules, by the remaining placebo-treated survivor corals at T3 suggests ongoing apoptosis ( 47 ). In addition, the oxidative stress increased by thermal photodamage to the photosynthetic apparatus of Symbiodiniaceae might be further contributing to trigger inflammatory responses ( 49 ). Previous studies have also found expression of immune-related and apoptosis genes in corals affected by heat stress for extended periods of time ( 50 , 51 ), suggesting a persistent bleaching effect on the coral transcriptome of susceptible corals ( 52 ). We hypothesize that down-regulation of orthogroups involved in apoptosis and the concomitant up-regulation of thermal stress protection proteins, such as chaperones, promoted by BMC inoculation, protected corals from tissue damage and Symbiodiniaceae loss, with consequences for coral survival. Such prominent recovery promoted by coral probiotics indicates that if the selection of BMCs based on the hypothetical framework proposed by Peixoto et al . ( 10 , 11 ) already provides measurable benefit and survivorship improvement, more careful selection of BMCs could result in even larger improvements. It is worthwhile and interesting to also highlight that such reprogramming was also observed when no stress was applied (at 26°C), but only at an early stage. While it seems that inoculation of BMCs rapidly trigger change of response norms from the host, long-term BMC reprogramming and exerted effects are only manifested under and subsequent to stress. It is tempting to speculate that the increased host survivorship observed in this study is a direct consequence of the transcriptomic changes discussed above, which arguably will result in altered metabolic profiles. For instance, the observed changes in the metabolomic profile of corals treated with BMC supports the hypothesis that the selected microbes play a direct role in increasing coral stress tolerance as evidenced by correspondence between selected traits of BMC bacteria (i.e., DMSP degradation) and observed metabolic changes. Shifts in BMC-treated metabolomic profiles were signified by a decrease in the DMSP concentration and lipidic reservoir maintenance. This connects directly to the presence of M24 in the 16 S rDNA data: M24 was found exclusively in BMC-treated samples at T1 and T2, indicating its incorporation into the coral microbiome, and was selected because of its ability to degrade DMSP (table S1). Notably, DMSP is mostly produced by algae (Symbiodiniaceae), and its degradation generates antimicrobial compounds, helping to control pathogens ( 53 – 55 ). Peixoto et al . ( 10 , 11 ) suggested that this is a desirable BMC trait. In parallel to the DMSP degradation as one of the direct mechanisms provided by the BMC consortium to ameliorate heat stress, the BMC treatment may have also indirectly influenced DMSP metabolism, through the enrichment of bacteria able to assimilate DMSP, such as Ruegeria ( 56 ), the most abundant genus found in BMC-treated coral samples during the course of the experiment. This genus has been previously observed to inhibit and control the growth or pathogenicity of V. coralliilyticus ( 57 ). These observed traits may have triggered the molecular responses observed by the host. These results highlight the importance of microbiome restructuring to coral resilience ( 12 , 15 ) and the additional potential role of the M. hispida BMC consortium in modulating the microbial colonization and succession of inoculated coral fragments. This parallel colonization/succession/enrichment of beneficial microbes has also been observed in other hosts, including humans, as a result of the use of pre- or probiotics ( 58 , 59 ). The increasing frequency and severity of ocean warming events has caused coral die-offs worldwide in the last few years ( 1 , 60 – 62 ). The development and better understanding of novel interventions to mitigate large-scale coral mortality is one of the climate priorities for the coming decades ( 63 , 64 ). The results of this study provide three completely novel insights that can aid the development of tools to promote human-accelerated environmental adaptation of corals: (i) BMC treatment and heat stress are both necessary conditions to trigger a long-term BMC thermal protection effect, whereas neither on its own is sufficient; (ii) the BMC thermal protection effect manifests after the heat stress and affects recovery; and (iii) such BMC-promoted protection mitigates coral PHSD, preventing mortality. Our results support the potential of microbiome restructuring to aid in the environmental adaptation of the coral metaorganism to global change ( 15 ) and identify a suite of microbial-mediated host responses underlying coral survival and recovery to thermal bleaching provided through BMCs. This is most prominently highlighted by the marked increase of 40% in coral survival rates following thermal stress and prior BMC treatment. This was accompanied by overall shifts in the coral microbiome that suggest a dynamic restructuring of the microbiome, due partially to the incorporation of BMC members and the relative increase of other bacteria. We further show that such microbiome restructuring directly affects the host, exerting beneficial effects and PHSD mitigation, as evidenced by transcriptional reprogramming (i.e., down-regulating apoptosis and inflammatory triggering molecules and up-regulating thermal stress protection proteins). In this light, our results reinforce the promise and potential of coral probiotics as an effective tool to rehabilitate coral reefs, particularly because the ability to “recover” is what eventually makes the difference in the real world, i.e., not only the difference in responding to heat stress but also in surviving the heat stress. In this regard, our data also suggest that prophylactic inoculation of BMCs, a few weeks before thermal events, can be advantageous for corals to sustain heat stress as it supposedly allows them to more rapidly and readily recover from thermal stress."
} | 4,349 |
36258070 | null | s2 | 49 | {
"abstract": "Pseudomonas aeruginosa, like many bacteria, uses chemical signals to communicate between cells in a process called quorum sensing (QS). QS allows groups of bacteria to sense population density and, in response to changing cell densities, to coordinate behaviors. The P. aeruginosa QS system consists of two complete circuits that involve acyl-homoserine lactone signals and a third system that uses quinolone signals. Together, these three QS circuits regulate the expression of hundreds of genes, many of which code for virulence factors. P. aeruginosa has become a model for studying the molecular biology of QS and the ecology and evolution of group behaviors in bacteria. In this chapter, we recount the history of discovery of QS systems in P. aeruginosa, discuss how QS relates to virulence and the ecology of this bacterium, and explore strategies to inhibit QS. Finally, we discuss future directions for research in P. aeruginosa QS."
} | 235 |
40234391 | PMC12000578 | pmc | 50 | {
"abstract": "Neuromorphic technologies adapt biological neural principles to synthesise high-efficiency computational devices, characterised by continuous real-time operation and sparse event-based communication. After several false starts, a confluence of advances now promises widespread commercial adoption. Gradient-based training of deep spiking neural networks is now an off-the-shelf technique for building general-purpose Neuromorphic applications, with open-source tools underwritten by theoretical results. Analog and mixed-signal Neuromorphic circuit designs are being replaced by digital equivalents in newer devices, simplifying application deployment while maintaining computational benefits. Designs for in-memory computing are also approaching commercial maturity. Solving two key problems—how to program general Neuromorphic applications; and how to deploy them at scale—clears the way to commercial success of Neuromorphic processors. Ultra-low-power Neuromorphic technology will find a home in battery-powered systems, local compute for internet-of-things devices, and consumer wearables. Inspiration from uptake of tensor processors and GPUs can help the field overcome remaining hurdles.",
"introduction": "Introduction Two main processing architectures currently dominate commercial computing: von Neumann architectures, comprising the majority of computing devices and now spread like dust in every corner of the planet; and tensor processors such as GPUs and TPUs, which have seen a rapid rise in use for computation since 2010. Inspiration for computing systems came early from human and animal nervous systems. McCullough-Pitts 1 proposed simplified logical units that communicate with binary values, inspired directly by activity in the nervous system. Subsequent to his state-machine model of computation which underlies modern procedural programming, Turing proposed self-modifying neurally-inspired computing systems 2 . Von Neumann himself was fascinated by information processing in the brain 3 , and embarked on a research program to define principles for self-constructing computational machines (automata) 4 , predating the concept of DNA as a computational substrate. Von Neumann highlighted the need to program a device in terms of the fundamental computational functions it can perform 3 . In the case of tensor processors these are matrix multiplications; in the case of CPUs these are primarily basic arithmetic and branching operations. In the case of Neuromorphic (NM) technologies, the basic computational elements are directly inspired by biological nervous systems—temporal integration and dynamics; binary or low-bit-depth multiplication; and thresholding and binary communication between elements. These technologies comprise a third computational architecture, in addition to CPUs and tensor processors, and several ongoing commercial and research projects are taking steps to bring NM technologies to widespread use. How can we learn from the wild success stories of tensor processors (particularly NVIDIA’s GPUs) and older CPUs, as they rose to dominate the computing landscape? And how can we adapt their success to NM processors? While the workings of our own brains are fundamentally parallel, convoluted, and remain obscure, we often use simple linear sequences of instructions when explaining a task to others. Consequently, Turing/Jacquard inspired procedural programming models came to dominate digital computing, bringing with them von Neumann fetch-and-execute processing architectures (Fig. 1 a). These systems provide a straightforward recipe for general-purpose computing—if you can describe step-by-step what you want a computer to do, then you can code and deploy it. This success came at the expense of analog computers, which had been highly important for scientific and war-time computations until around the middle of the 20th Century. Alternative massively-parallel multi-processor architectures fell out of favour in the 1990s, partially due to the difficulty in developing applications for these systems 5 , 6 . Fig. 1 Programming models for von Neumann, tensor processor, and Neuromorphic hardware architectures. a von Neumann processors such as CPUs and MCUs use a number of programming models, all of which compile to a predominately serial, fetch-and-execute approach (small-scale multi-processing and optimised pre-fetch notwithstanding). Programming these systems requires decomposing a desired task into a set of explicit procedures for a CPU to execute. b Tensor processors such as GPUs, TPUs and NPUs have achieved great success for general-purpose applications, by adopting an example-driven, supervised machine-learning (ML) programming model based on gradient-based parameter optimisation (green box). This is supported by APIs (yellow box) which efficiently implement the ML programming model on tensor processors. c Neuromorphic architectures are increasingly adopting a similar programming model, based on methods developed for deep learning (green arrow). At the same time, new software APIs for Neuromorphic application development are making efficient use of the SW tools for deep learning (yellow arrow), which permit rapid parallelised training of NM architectures on commodity tensor processors (dashed arrow). In the late 1990s and early 2000s the consumer graphics hardware market emerged primarily to satisfy the computational needs of home gaming. But computational scientists soon realised that the parallel vector processing cores of GPUs could be applied efficiently to data-parallel problems. Neural Network (NN) researchers found that GPUs could dramatically accelerate NN computations, permitting larger and more complex architectures to be trained faster than ever previously. Since 2010 the rise of deep neural networks has driven and been driven by a concomitant rise in tensor processor architectures, with enormous commercial success in particular for the largest GPU manufacturer NVIDIA. This shift back to parallel processing was directly enabled by the success of deep learning in providing a programming model for tensor computations — the ability to map almost any arbitrary use case to tensor computing, via a data-driven machine-learning approach 7 (Fig. 1 b). Similarly, the commercial success of NVIDIA was driven by their foresight in providing a software Application Programming Interface (API; i.e. CUDA) for their GPUs 8 , which allowed this new programming model to be mapped efficiently to their hardware. In a beneficial feedback process, availability of GPUs and increasing demand for neural network processing have driven development of both. Formerly designed to accelerate graphical loads, more recent high-end GPUs are now explicitly designed for tensor-based computational tasks. Coupled with the provision of simple, high-level APIs embodying the deep learning programming model (i.e. Tensorflow/Keras, Pytorch), tensor processors are accessible to developers in a way that earlier multi-processing architectures were not. In this perspective we outline the early forays of Neuromorphic hardware towards commercial use; describe the new steps the field has made in the last few years; compare and contrast the current design alternatives in nascent commercial Neuromorphic hardware; and sketch a path that we believe will lead to widespread consumer adoption of Neuromorphic technology. We focus primarily on processors making use of sparse event-driven communication with limited precision, with or without analog computing elements, and incorporating temporal dynamics as an integral aspect of computation. We include some discussion of compute-in-memory architectures, where these overlap with our focus. This covers digital implementations of spiking neurons and some analog compute-in-memory arrays, along with traditional sub-threshold silicon Neuromorphic processors."
} | 1,969 |
35378031 | PMC9092336 | pmc | 51 | {
"abstract": "The tiny spider makes\ndragline silk fibers with unbeatable toughness,\nall under the most innocuous conditions. Scientists have persistently\ntried to emulate its natural silk spinning process using recombinant\nproteins with a view toward creating a new wave of smart materials,\nyet most efforts have fallen short of attaining the native fiber’s\nexcellent mechanical properties. One reason for these shortcomings\nmay be that artificial spider silk systems tend to be overly simplified\nand may not sufficiently take into account the true complexity of\nthe underlying protein sequences and of the multidimensional aspects\nof the natural self-assembly process that give rise to the hierarchically\nstructured fibers. Here, we discuss recent findings regarding the\nmaterial constituents of spider dragline silk, including novel spidroin\nsubtypes, nonspidroin proteins, and possible involvement of post-translational\nmodifications, which together suggest a complexity that transcends\nthe two-component MaSp1/MaSp2 system. We subsequently consider insights\ninto the spidroin domain functions, structures, and overall mechanisms\nfor the rapid transition from disordered soluble protein into a highly\norganized fiber, including the possibility of viewing spider silk\nself-assembly through a framework relevant to biomolecular condensates.\nFinally, we consider the concept of “biomimetics” as\nit applies to artificial spider silk production with a focus on key\npractical aspects of design and evaluation that may hopefully inform\nefforts to more closely reproduce the remarkable structure and function\nof the native silk fiber using artificial methods.",
"conclusion": "6 Conclusion As recent reports suggest and as we hope this review illustrates,\nwe are still far from having a complete grasp of the true complexity\nof spider dragline silk. Future studies should continue to shed light\non the details concerning the interactions among the different biomolecular\ncomponents of spider silk and in doing so allow scientists to better\nmimic the unique properties of the natural fiber using artificial\nmethods. Such developments would pave the way for the creation of\nnew and exciting materials with high-performance capabilities while\nminimizing the negative environmental impact.",
"introduction": "1 Introduction From antiquity to the\npresent era, spiders’ ability to spin\nbeautiful and functional silken structures, unmatched in nature, has\nevoked a deep fascination in mankind. 1 From\nthe standpoint of science, too, spider silk has inspired generations\nof researchers, mainly for the remarkable properties of the fiber,\nwhose mechanical performance and hierarchical organization are still\nunmatched by the most sophisticated artificial materials. This\nPerspective is not intended to be a comprehensive review of\nthe topic of spider silks, excellent examples of which are thankfully\navailable. 2 − 5 Rather, we wish to take this opportunity to highlight some recent\ndevelopments in the field and then consider what these could mean\nin terms of future research directions. The discussion is focused\nmainly on dragline (major ampullate) silk, which is by far the most\nstudied type of spider silk. The first part will explore the complexity\nof spider dragline silk in terms of its physical constituents. In\nrecent years, reports from the next-generation sequencing front have\nbegun to hint at a more nuanced picture of dragline silk composition\nthan previously anticipated, findings that might require a reassessment\nof the conventional and relatively simple two-component model based\non MaSp1 and MaSp2. Next, we turn to the functions of the individual\nspidroin domains and discuss the different frameworks that can be\nused to understand and further explore the silk self-assembly mechanisms\nwith an emphasis on more recent findings. We subsequently tackle the\ntopic of biomimetics , where we ask the question what\nconstitutes a true biomimetic approach for making spider silk? We\nhighlight several studies that have claimed to apply biomimetic principles\nin making artificial spider silks and provide suggestions for possible\nfuture directions."
} | 1,024 |
19686078 | null | s2 | 52 | {
"abstract": "Quorum sensing is a cell-cell communication process in which bacteria use the production and detection of extracellular chemicals called autoinducers to monitor cell population density. Quorum sensing allows bacteria to synchronize the gene expression of the group, and thus act in unison. Here, we review the mechanisms involved in quorum sensing with a focus on the Vibrio harveyi and Vibrio cholerae quorum-sensing systems. We discuss the differences between these two quorum-sensing systems and the differences between them and other paradigmatic bacterial signal transduction systems. We argue that the Vibrio quorum-sensing systems are optimally designed to precisely translate extracellular autoinducer information into internal changes in gene expression. We describe how studies of the V. harveyi and V. cholerae quorum-sensing systems have revealed some of the fundamental mechanisms underpinning the evolution of collective behaviors."
} | 236 |
35694505 | PMC9178770 | pmc | 53 | {
"abstract": "Triboelectric nanogenerators\n(TENGs) have shown huge application\npotential in the fields of micro–nano energy harvesting and\nmultifunctional sensing. However, the damage of triboelectric material\nis one of the challenges for their practical applications. Herein,\nwe fabricated a flexible TENG employing self-healing hydrogel and\nfluorinated ethylene propylene film as triboelectric materials for\nmechanical energy harvesting and pressure monitoring. The prepared\nhydrogel not only has excellent flexibility, transparency, and self-healing\nproperty but also exhibits good mechanical property without plastic\ndeformation and damage under a large stretchable strain of 200%. The\noutput electric signals of TENGs are as high as 33.0 V and 3 μA\nunder a contact frequency of 0.40 Hz and a pressure of 2.9 N, respectively,\nwhich can charge a capacitor of 0.22 μF to 24.3 V within 300\ns. Note that the voltage retention rate of TENGs after self-healing\nis up to 88.0%. Moreover, hydrogel-based TENGs can act as a wearable\npressure sensor for monitoring human motion, exhibiting a high sensitivity\nof 105.9 mV/N or 1.73 nA/N under a contact frequency of 0.40 Hz. This\nresearch provides a reference roadmap for designing TENGs and self-powered\npressure sensors with flexibility, self-healing, and robustness.",
"conclusion": "4 Conclusions In summary, this work presents a novel\nand simple self-healing\nhydrogel-based TENG and sensor for efficient mechanical harvesting\nand motion monitoring. Through structure design and regulation, a\nflexible and transparent hydrogel was prepared with excellent self-healing\nproperty. Moreover, the prepared self-healing hydrogel shows good\nmechanical properties without plastic deformation even at a large\nstretchable strain of 200%. Under a contact frequency of 0.40 Hz and\na pressure of 2.9 N, the fabricated TENG generates the output electrical\nsignals of 33.0 V and 3 μA, respectively, which can be used\nto charge capacitors. Comparing the output performance changes of\nthe hydrogel-based TENG and the healed hydrogel-based TENG, it was\nfound that the latter does not decrease significantly. As a wearable\narray pressure sensor based on several individual TENGs, a high sensitivity\nof 105.9 mV/N can be realized. Moreover, the output signals of the\nsensor are different under different motion states of the human body.\nThe study demonstrates the potential application of self-healing hydrogels\nas triboelectric layers for TENGs and wearable triboelectric pressure\nsensors.",
"introduction": "1 Introduction As technology continues to advance, flexible wearable electronics\ngradually show huge application potential in electronic skin (E-skin), 1 , 2 soft robotics, 3 , 4 health monitoring, 5 , 6 and other aspects in human daily lives. 7 However, as traditional power supply methods, batteries and capacitors\nrequire frequent charging and maintenance due to limited capacitance,\nwhich affect the continuous operation and stability of wearable electronics,\nespecially in harsh environments. 8 , 9 At the same\ntime, discarded batteries/capacitors also cause serious pollution\nto the environment. 10 − 14 Therefore, developing high-performance sustainable energy technologies\nbecomes one of the important research topics. In 2012, Wang’s\ngroup first invented the triboelectric nanogenerator (TENG) that can\nconvert various kinds of mechanical energies into electrical energy\nfor self-powered electronics, 15 such as\nhuman motion energy, 16 vibration energy, 17 , 18 wind energy, 19 , 20 rain drop energy, 21 water wave energy, 22 , 23 and sound energy. 24 , 25 With continuous research and\nprogress, the output performance of TENGs have been greatly improved, 26 − 34 which facilitates their practical process in daily life. It plays\na very good supplementary role to traditional energy and is of great\nsignificance to the realization of carbon neutrality goals. As a new intrinsically conductive material, polymeric hydrogels\nhave adjustable conductivity, excellent self-healing performance,\nand good biocompatibility, thus showing great potential applications\nin soft robots, biomimetic prostheses, health monitoring, and wearable\nelectronics. 35 − 38 The hydrogel-based TENGs have already attracted great attention\nand obtained satisfactory achievement by hydrogel structure design\nand optimization. 39 − 42 Pu et al. fabricated a flexible sandwich-structured TENG based on\na polyacrylamide hydrogel as an electrode for harvesting biomechanical\nenergy and acting as an E-skin. 41 However,\nthis TENG does not possesses self-healing property and cannot work\nnormally after damage. Sun et al. developed a polyacrylamide/gelatin/PEDOT:PSS\ncomposite hydrogel that has good flexibility, stretchability, and\nsensitivity to stress. 42 As the electrode\nof a sandwich-structured TENG, only the hydrogel has good self-healing\nproperty, and TENG cannot work normally when the charged layer is\ndamaged. In addition, a linear silicone-modified polyurethane coating\nand a temperature-responsive polycaprolactone film as self-healing\nfriction layers have been used to fabricate TENGs, which is of great\nsignificance to prolong the service life of TENGs. 43 , 44 Thus far, there are no reports on the use of self-healing hydrogels\nas triboelectric materials to fabricate TENGs. Here, we report\na novel TENG based on a flexible and transparent\nhydrogel with excellent self-healing property directly as a triboelectric\nmaterial, which shows great potential for a broad range of applications\nin mechanical energy harvesting and pressure monitoring. After being\nstretched to 200% strain, the hydrogel still shows good mechanical\nperformance without plastic deformation. After a complete self-healing\nprocess, a cut hydrogel as the friction material can still give TENG\nhigh output performance with a large voltage retention rate of 88%.\nA hydrogel-based wearable array sensor exhibits a high sensitivity\nof 105.9 mV/N or 1.73 nA/N, showing great application potential in\nself-powered wearable sensing systems.",
"discussion": "3 Results and Discussion Figure 1 a shows\nthe preparation process of the self-healing hydrogel. Using APS as\nan initiator, AM and DAAM as monomers were free-radically copolymerized\nto form long chains of PAM- co -DAAM. The ketone group\nin the long chain reacts with the hydrazide of ADH to form an acylhydrazone\nbond as the first cross-linking point and can improve the toughness\nand stretchability of the hydrogel. 45 Moreover,\nPVP can synergize with the CONH 2 functional groups of PAM- co -DAAM to generate hydrogen bonds to form a second cross-link. 46 The FTIR spectrum of the hydrogel shows that\nthe stretching bands at 3435, 1630, and 1097 cm –1 correspond to the characteristic absorption bands of O–H,\nC=O, and C–N in ADH, respectively ( Figure S1 ). The self-healing principle of the hydrogel is\ndisplayed in Figure 1 b. When the hydrogel is cut, owing to the hydrogen bonds between\nPVP chains and the CONH 2 functional groups of PAM- co -DAAM, the hydrogel can be repaired spontaneously without\nexternal interference. After self-healing for 12 h, the two hydrogels\ncan self-heal well under the dynamic cleavage and reconstruction of\nhydrogen bonds and the rearrangement of polymer segments. The process\nof forming hydrogen bonds is shown in Figure 1 c. Figure 1 (a) Schematic diagram for the preparation process\nof the self-healing\nhydrogel. (b,c) Self-healing principle diagram of the hydrogel. (d)\nDemonstration of tensile property of the self-healing hydrogel. (e)\nDemonstration of tensile property of the self-healing hydrogel after\nself-healing. (f) Stretched length and corresponding tensile force\ncurves of self-healing hydrogel from the first to fifth stretches.\n(g) Demonstration of flexibility of self-healing hydrogel: bending\nand twisting. Note that the obtained hydrogel\nhas a typical porous network morphology\nof the gel matrix after freeze-drying, indicating the formation of\nthe hydrogel ( Figure S2a,b ). Also, the\nLSCM images showed that the surface of the hydrogel before freeze-drying\nwas smooth ( Figure S2c,d ). To examine the\ntensile properties of the hydrogels, the prepared hydrogels with a\nthickness of 1.0 mm were dyed red and green, respectively, and then\ncut into strips of 1.0 × 6.0 cm ( Figure S3a ). The two ends were clamped for 0.5 cm and then slowly stretched.\nAs can be seen from Figures 1 d and S4 , the clamped hydrogels\ncan be stretched from 5.0 to 15.0 cm with a stretchable strain of\n200%. Note that when the two hydrogels were cut from the middle ( Figure S3b ), self-healing process could be completed\nwithin 12 h and the original tensile properties could be maintained\n( Figure 1 e). To explore\nthe mechanical properties of the hydrogel, the relationship between\nthe stretched length of the hydrogel and corresponding tensile force\nwas measured. As shown in Figure 1 f, the hydrogel with a width of 1.0 cm was gradually\nelongated from 5.0 to 15 cm, and the tensile force was also gradually\nincreased from 0 to 0.71 N. During the recovery process, the pulling\nforce was gradually reduced from 0.71 to 0 N, showing excellent stability\nfor 5 cycles. This indicates that the hydrogel has good elastic deformation\nproperties without irreversible deformation during the stretching-recovery\nprocess from 5.0 to 15.0 cm. In addition, the photographs of the prepared\nself-healing hydrogels in the bending and twisting states are displayed\nin Figure 1 g, showing\ngood transparency and excellent flexibility. To further explore\nthe application value of the self-healing hydrogel\nin energy-harvesting and self-powered sensing system, we designed\na TENG and a wearable pressure sensor based on the hydrogel, respectively. Figure 2 a shows the schematic\ndiagram of the TENG with a contact-separation mode fixed on a linear\nmotor. The self-healing hydrogel and FEP film were used as triboelectric\nmaterials with a size of 5.0 × 5.0 cm, Al foils were used as\nelectrodes, and acrylic sheets were used as support materials. Figure 2 b,c shows the structure\ndiagram and optical image of the wearable pressure sensor, which contains\nfour TENGs individually with the size of 2.0 × 2.0 cm. Figure 2 d is a schematic\ndiagram of the working principle of the TENG. Its work process can\nbe divided into four steps: (I) When pressing the TENG, the FEP film\ncontacts with the hydrogel film, and the equal and opposite charges\nare generated on the surface of FEP (negative charge) and hydrogel\n(positive charge) films due to the triboelectrification. (II) When\nremoving the external force, the FEP film separates from the hydrogel\nfilm. Owing to electrostatic induction, the electron flows from the\nupper Al electrode to the bottom Al electrode, and the opposite charges\nare generated on two Al electrodes, while negative charges are generated\nby the FEP film. In this process, a current was produced from the\nbottom to upper electrodes. (III) When the FEP film recovers to its\noriginal state, there is no electron flow between the two electrodes\nowing to the electrostatic balance. (IV) when pressing the TENG again,\nthe electrostatic balance is broken, and the electron flows from the\nbottom to the upper Al electrode, corresponding to an opposite current\ndirection compared with step II. The potential difference and potential\ndistribution of the two electrodes of the TENG were theoretically\nsimulated using COMSOL modeling, and the results are shown in Figure 2 e. Figure 2 (a,b) Schematic diagram\nof the self-healing hydrogel-based TENG\n(a) and pressure array sensor (b). (c) Photograph of the pressure\narray sensor. (d) Working principle of TENG. (e) Simulation calculations\nof the electric potential distribution of TENG between contacting\ninterfaces by COMSOL software. To measure the output performance of the hydrogel-based TENG with\na size of 5.0 × 5.0 cm ( Figure 2 a) under different conditions, we systematically studied\nthe output voltages and currents of the TENG under different working\nfrequencies and forces. First, we measured the effect of different\nfrequencies on the output performance of the device at 1.7 N. As shown\nin Figure 3 a,b, as\nthe contact frequencies gradually increased from 0.10 to 0.40 Hz,\nthe open-circuit voltages/short-circuit currents increased from 2.0\nV/0.3 μA to 8.0 V/0.9 μA, respectively. The improvement\nof the output performance is attributed to the increase of the electrostatic\ninduction rate induced by the faster contact frequency, thus increasing\nthe charges migration rate. Then, we further studied the output performance\nof the TENG under different pressure conditions and a constant contact\nfrequency of 0.40 Hz. With the increase of the forces from 1.7 to\n2.9 N, the corresponding output voltages/currents increase from 7.0\nV/0.9 μA to 33.0 V/3.0 μA, respectively ( Figure 3 c,d). This is due to the increase\nof the contact area and degree between the hydrogel and the FEP film,\nwhich promotes the charge generation, thereby enhancing the output\nperformance. The above research illustrates that the hydrogel-based\nTENG is sensitive to contact frequency and pressure, showing huge\napplication potential in frequency and pressure monitoring. To determine\nthe authenticity of the output signals of TENG, the circuit is connected\nthrough forward and reverse connections. The output performance of\nthe TENG with the two connection modes above were measured under a\npressure of 2.9 N and a contact frequency of 0.40 Hz ( Figure 3 e,f). The output voltages/currents\nare approximately 34.0 V/3.5 μA and −34.0 V/–3.5\nμA, respectively. Despite the change in the circuit’s\nforward and reverse connections, the absolute values of output electrical\nsignals were unchanged, which confirms the authenticity of the output\nsignals. It is worth noting that there is no obvious wear phenomenon\non the hydrogel surface when the TENG runs continuously for 6 h ( Figure S5 ), indicating that the device has good\nstability. In addition, in order to test the output performance change\nof the hydrogel-based TENG before cutting and after self-healing,\nthe TENGs were constructed using the hydrogel before cutting and after\nself-healing as the triboelectric materials ( Figure S6 ), respectively. Figure S7a,b shows\nthe output voltage and current signals of the TENGs under a pressure\nof 2.9 N and a frequency of 0.40 Hz, the open-circuit voltage is reduced\nfrom 34.0 V before cutting to 30.0 V after self-healing, and the corresponding\nshort-circuit current is reduced from 3.0 to 2.5 μA. It can\nbe seen from the calculations that the output voltage and current\nretention rates of the TENG fabricated with the self-healed hydrogel\ncurrent are as high as 88 and 83%, respectively, indicating excellent\nstability of the TENG. Note that both the output voltages and currents\nof the TENG increased with the increase of healing time, and when\nthe hydrogel was completely healed, the output performance of TENG\nalmost reached the level before cutting ( Figure S8a,b ). Figure 3 g displays the measured output currents and calculated powers of\nthe TENG under different external loading resistances at a pressure\nof 2.9 N and a contact frequency of 0.40 Hz, showing that the output\ncurrent decreases with the increase in the external loading resistance.\nAccording to the formula P = I 2 R , the output power of the TENG under different\nexternal loading resistances could be calculated. As the picture shows,\nthe corresponding output power first increases and then decreases\nrapidly. Moreover, the optimum output power of the TENG is about 383\nμW under a loading resistance of 70 MΩ. In order to facilitate\nthe use for tiny electronic devices, the produced electric energy\nfrom the TENG by harvesting mechanical energy is usually stored in\ncapacitors or batteries. We studied the charging performance of the\ncapacitors with different capacities. Note that a rectifier is required\nto connect to the TENG to convert the alternating current generated\nby the TENG into direct current for charging purposes. As illustrated\nin Figure 3 h, the smaller\nthe capacitor capacity, the faster the increase in voltage in the\ncharging process. The 0.22 μF capacitor voltage can be charged\nfrom 0 to 24.3 V in approximately 295 s, while the 0.33, 10, and 22\nμF capacitor voltages can be charged to 11.7, 2.2, and 1.2 V,\nrespectively. The charging results show that the electrical energy\ngenerated by the TENG can be successfully stored in the capacitor,\nwhich provides the possibility for the continuous operation of the\nmicroelectronic devices. Figure 3 (a,b) Measured output voltage (a) and current\n(b) signals of TENG\nunder different frequencies (0.10–0.40 Hz) at 1.7 N. (c,d)\nMeasured output voltage (c) and current (d) signals of TENG under\ndifferent pressures (1.7–2.9 N) at 0.40 Hz. (e,f) Measured\noutput voltage (e) and current (f) signals of TENG under forward connection\nand reversed connection at 2.9 N and 0.40 Hz. (g) Measured output\ncurrents and calculated output powers of TENG under different loads\nat 2.9 N and 0.40 Hz. (h) Charging curves of capacitors with different\ncapacitances driven by TENG at 2.9 N and 0.40 Hz. To demonstrate the application potential of self-healing hydrogel-based\nTENG in wearable pressure sensors, a 2 × 2 sensor array consisting\nof four small-scale TENGs was constructed. It is important to analyze\nthe response of the output performance to contact frequency and pressure. Figure 4 a,b shows the output\nvoltage and current signals of the four channels of the sensor under\nconditions of constant pressure (15 N) and different frequencies,\nrespectively. It can be seen that when the contact frequency is 0.20\nHz, the output voltage of the channel 1 is 0.9 V, and the corresponding\ncurrent is approximately 45 nA. As the frequency increases, the output\nelectrical signals also increase accordingly. At 0.33 and 0.40 Hz,\nthe output voltage/current signals are 1.3 V/56 nA and 1.8 V/68 nA,\nrespectively. The output signals of channel 2, 3, and 4 have the same\ntrend and similar values. Figure 4 c,d further presents the output variation of the four\nchannels at different pressures at a constant contact frequency of\n0.40 Hz. Similar to the effect of contact frequency, with the increase\nof applied pressure from 15 to 28 N, the generated electrical signals\nof four channels all gradually increase from ∼1.9 V/67 nA to\n∼3.4 V/94 nA. To measure the sensitivity of the pressure sensor,\nwe have fitted the linear relationship between pressure and output\nvoltage/current peaks at different contact frequencies. As shown in Figure 4 e,f, for the pressure\nsensor, both the output voltage and current signals show an excellent\nlinear relationship with applied pressures. The slope of the fitted\ncurves represents the sensitivity of the sensor to pressure. At 0.20\nHz, the calculated sensitivities of the pressure sensor from voltage\nand current are 91.0 mV/N and 2.315 nA/N, respectively. When the contact\nfrequencies are 0.33 and 0.4 Hz, the corresponding sensitivities are\n97.2 mV/N/2.16 nA/N and 105.9 mV/N/1.73 nA/N, which shows good sensitivity\nof the pressure sensor applied under different contact frequencies. Figure 4 (a,b)\nMeasured output voltage (a) and current (b) signals of the\nfour channels of the sensor under different frequencies at 15 N. (c,d)\nMeasured output voltage (c) and current (d) signals of the four channels\nof the sensor under different pressures at 0.40 Hz. (e) Linear fitting\ncurves between the output voltages and pressures of the sensor under\ndifferent frequencies. (f) Linear fitting curves between the output\ncurrents and pressures of the sensor under different frequencies. The fabricated 2 × 2 array pressure sensor\nabove can act as\na wearable pressure sensor for motion monitoring. First, we examined\nthe practical effect of the motion sensor, as shown in Figure 5 a. A finger presses the four\nindividual sensor units from channel 1 to 4 in sequence, then from\nchannel 4 to 1; each channel produces the corresponding induction\nsignals in turn. Note that no electrical signal output is detected\nin the unit where no force is applied. Due to the different pressures\nof the fingers, the generated signals also vary from 0.33 to 0.98\nV. The corresponding test photos of pressing the four channels with\nfingers are shown in Figure 5 b–e. Figure 5 f is the sensor output voltage signals measured as the volunteer\nperiodically straightens and bends his elbow; the corresponding test\nphotos are displayed in Figure 5 g,h, respectively. When the elbow is straightened, the sensor\nhas no voltage signal output, but when the elbow is first bent and\nthen straightened, the four channels of the sensor can simultaneously\ngenerate four similar electrical signals. Due to the different positions\nof each channel, the magnitude of the signal induced by it is also\ndifferent, ranging from 1.69 to 2.36 V. Moreover, when the elbow is\nbent at different angles, each sensor unit detects different output\nvoltage signals, showing huge application potential in monitoring\nmotion. Figure 5 (a) Measured output voltage signals of the sensor when the right\nindex finger presses the four channels in turn. (b–e) Photographs\nof the right index finger pressing sensor channel 1 (b), channel 2\n(c), channel 3 (d), and channel 4 (e). (f) Measured output voltage\nsignals of the four channels of the sensor when the elbow is flexed.\n(g,h) Photographs of the person wearing a sensor, straightening (g)\nand flexing (h) the elbow."
} | 5,361 |
29955057 | PMC6023896 | pmc | 54 | {
"abstract": "Neuromorphic computing has emerged as a promising avenue towards building the next generation of intelligent computing systems. It has been proposed that memristive devices, which exhibit history-dependent conductivity modulation, could efficiently represent the synaptic weights in artificial neural networks. However, precise modulation of the device conductance over a wide dynamic range, necessary to maintain high network accuracy, is proving to be challenging. To address this, we present a multi-memristive synaptic architecture with an efficient global counter-based arbitration scheme. We focus on phase change memory devices, develop a comprehensive model and demonstrate via simulations the effectiveness of the concept for both spiking and non-spiking neural networks. Moreover, we present experimental results involving over a million phase change memory devices for unsupervised learning of temporal correlations using a spiking neural network. The work presents a significant step towards the realization of large-scale and energy-efficient neuromorphic computing systems.",
"introduction": "Introduction The human brain with less than 20 W of power consumption offers a processing capability that exceeds the petaflops mark, and thus outperforms state-of-the-art supercomputers by several orders of magnitude in terms of energy efficiency and volume. Building ultra-low-power cognitive computing systems inspired by the operating principles of the brain is a promising avenue towards achieving such efficiency. Recently, deep learning has revolutionized the field of machine learning by providing human-like performance in areas, such as computer vision, speech recognition, and complex strategic games 1 . However, current hardware implementations of deep neural networks are still far from competing with biological neural systems in terms of real-time information-processing capabilities with comparable energy consumption. One of the reasons for this inefficiency is that most neural networks are implemented on computing systems based on the conventional von Neumann architecture with separate memory and processing units. There are a few attempts to build custom neuromorphic hardware that is optimized to implement neural algorithms 2 – 5 . However, as these custom systems are typically based on conventional silicon complementary metal oxide semiconductor (CMOS) circuitry, the area efficiency of such hardware implementations will remain relatively low, especially if in situ learning and non-volatile synaptic behavior have to be incorporated. Recently, a new class of nanoscale devices has shown promise for realizing the synaptic dynamics in a compact and power-efficient manner. These memristive devices store information in their resistance/conductance states and exhibit conductivity modulation based on the programming history 6 – 9 . The central idea in building cognitive hardware based on memristive devices is to store the synaptic weights as their conductance states and to perform the associated computational tasks in place. The two essential synaptic attributes that need to be emulated by memristive devices are the synaptic efficacy and plasticity. Synaptic efficacy refers to the generation of a synaptic output based on the incoming neuronal activation. In conventional non-spiking artificial neural networks (ANN), the synaptic output is obtained by multiplying the real-valued neuronal activation with the synaptic weight. In a spiking neural network (SNN), the synaptic output is generated when the presynaptic neuron fires and typically is a signal that is proportional to the synaptic conductance. Using memristive devices, synaptic efficacy can be realized using Ohm’s law by measuring the current that flows through the device when an appropriate read voltage signal is applied. Synaptic plasticity, in contrast, is the ability of the synapse to change its weight, typically during the execution of a learning algorithm. An increase in the synaptic weight is referred to as potentiation and a decrease as depression. In an ANN, the weights are usually changed based on the backpropagation algorithm 10 , whereas in an SNN, local learning rules, such as spike-timing-dependent plasticity (STDP) 11 or supervised learning algorithms, such as NormAD 12 could be used. The implementation of synaptic plasticity in memristive devices is achieved by applying appropriate electrical pulses that change the conductance of these devices through various physical mechanisms 13 – 15 , such as ionic drift 16 – 20 , ionic diffusion 21 , ferroelectric effects 22 , spintronic effects 23 , 24 , and phase transitions 25 , 26 . Demonstrations that combine memristive synapses with digital or analog CMOS neuronal circuitry are indicative of the potential to realize highly efficient neuromorphic systems 27 – 33 . However, to incorporate such devices into large-scale neuromorphic systems without compromising the network performance, significant improvements in the characteristics of the memristive devices are needed 34 . Some of the device characteristics that limit the system performance include the limited conductance range, asymmetric conductance response (differences in the manner in which the conductance changes between potentiation and depression), nonlinear conductance response (nonlinear conductance evolution with respect to the number of pulses), stochasticity associated with conductance changes, and variability between devices. Clearly, advances in materials science and device technology could play a key role in addressing some of these challenges 35 , 36 , but equally important are innovations in the synaptic architectures. One example is the differential synaptic architecture 37 , in which two memristive devices are used in a differential configuration such that one device is used for potentiation and the other for depression. This was proposed for synapses implemented using phase change memory (PCM) devices, which exhibit strong asymmetry in their conductance response. However, the device mismatch within the differential pair of devices, as well as the need to refresh the device conductance frequently to avoid conductance saturation could potentially limit the applicability of this approach 34 . In another approach proposed recently 38 , several binary memristive devices are programmed and read in parallel to implement a synaptic element, exploiting the probabilistic switching exhibited by certain types of memristive devices. However, it may be challenging to achieve fine-tuned probabilistic switching reliably across a large number of devices. Alternatively, pseudo-random number generators could be used to implement this probabilistic update scheme with deterministic memristive devices 39 , albeit with the associated costs of increased circuit complexity. In this article, we propose a multi-memristive synaptic architecture that addresses the main drawbacks of the above-mentioned schemes, and experimentally demonstrate an implementation using nanoscale PCM devices. First, we present the concept of multi-memristive synapses with a counter-based arbitration scheme. Next, we illustrate the challenges posed by memristive devices for neuromorphic computing by studying the operating characteristics of PCM fabricated in the 90 nm technology node and show how multi-memristive synapses can address some of these challenges. Using comprehensive and accurate PCM models, we demonstrate the potential of the multi-memristive synaptic concept in training ANNs and SNNs for the exemplary benchmark task of handwritten digit classification. Finally, we present a large-scale experimental implementation of training an SNN with multi-memristive synapses using more than one million PCM devices to detect temporal correlations in event-based data streams.",
"discussion": "Discussion The proposed synaptic architecture bears similarities to several aspects of neural connectivity in biology, as biological neural connections also comprise multiple sub-units. For instance, in the central nervous system, a presynaptic neuron may form multiple synaptic terminals (so-called boutons) to connect to a single postsynaptic neuron 52 . Moreover, each biological synapse contains a plurality of presynaptic release sites 53 and postsynaptic ion channels 54 . Furthermore, our implementation of plasticity through changes in the individual memristors is analogous to individual plasticity of the synaptic connections between a pair of biological neurons 55 , which is also true for the individual ion channels of a synaptic connection 55 , 56 . The involvement of progressively larger numbers of memristive devices during potentiation is analogous to the development of new ion channels in a potentiated synapse 53 , 54 . A significant advantage of the proposed multi-memristive synapse is its crossbar compatibility. In memristive crossbar arrays, matrix–vector multiplications associated with the synaptic efficacy can be implemented with a read operation achieving O (1) complexity. Note that memristive devices can be read with low energy (10–100 fJ for our devices), which leads to a substantially lower energy consumption than in conventional von Neumann systems 57 – 59 . Furthermore, the synaptic plasticity is realized in place without having to read back the synaptic weights. Even though, the power dissipation associated with programming the memristive devices is at least 10 times higher than that required for the read operation, as only one device of the multi-memristive synapse is programmed at each instance of synaptic update, our scheme does not introduce a significant energy overhead. Memristive crossbars can also be fabricated with very small areal footprint 27 , 29 , 60 . The neuron circuitry of the crossbar array, which typically consumes a larger area than the crossbar array, only increases marginally owing to the additional circuitry needed for arbitration. Finally, because even a single global counter can be used for arbitrating a whole array, the additional area/power overhead is expected to be minimal. The proposed architecture also offers several advantages in terms of reliability. The other constituent devices of a synapse could compensate for the occasional device failure. In addition, each device in a synapse gets programmed less frequently than if a single device were used, which effectively increases the overall lifetime of a multi-memristive synapse. The potentiation and depression counters reduce the effective number of programming operations of a synapse, further improving endurance-related issues. Device selection in the multi-memristive synapse is performed based on the arbitration module alone, without knowledge of the conductance values of the individual devices, thus there is a non-zero probability that a potentiation (depression) pulse will not result in an actual potentiation (depression) of the synapse. This would effectively translate into a weight-dependent plasticity whereby the probability to potentiate reduces with increasing synaptic weight and the probability to depress reduces with decreasing synaptic weight (see Supplementary Notes 7 , 8 ). This attribute could affect the overall performance of a neural network. For example, weight-dependent plasticity has been shown to impact the classification accuracy negatively in an ANN 61 . In contrast, a study suggests that it can stabilize an SNN intended to detect temporal correlations 49 . The ANN and SNN simulations in section “Simulation results on handwritted digit classification” with the PCM model perform worse, even in the presence of multi-memristive synapses with N > 10, than the simulations with double-precision floating-point weights. We show that this behavior does not arise from the weight-dependent plasticity of the multi-memristive synapse scheme, but from the nonlinear PCM conductance response (see Supplementary Note 9 ). Using a uni-directional linear device model where the conductance change is modeled as a Gaussian random number with mean (granularity) and standard deviation (stochasticity) of 0.5 μS, accuracies exceeding 96.7% are possible in ANN with only 1% performance loss compared with double-precision floating-point weights. Similarly, the network can classify more than 77% of the digits correctly in the SNN using the linear device model, reaching the accuracy of the double-precision floating-point weights. Note also that the drift in conductance states, which is unique to PCM technology, does not appear to have a significant impact on our studies. As described recently 62 , as long as the drift exponent is small enough (<0.1; in our devices it is on average 0.05, see Supplementary Note 4 ), it is not very detrimental for neural network applications. Our own experimental results on SNNs presented in section “Experimental results on temporal correlation detection” point in this direction, as the network seems to maintain the classification accuracy despite drift. Although conductance drift is not intended to be countered using the multi-memristive concept, there are attempts to overcome it via advanced device-level ideas 35 , which could be used in conjunction with a multi-memristive synapse. In the presence of significant nonlinear conductance response and drift, one could envisage an alternate multi-memristive synaptic architecture in which multiple devices are used to store the weights, but with varying significance. For instance, if N -bit synaptic resolution is required, N memory devices could be used, with each device programmed to the maximum (fully potentiated) or minimum (fully depressed) conductance states to represent a number in binary format. In such a binary system, for synaptic efficacy, each device needs to be read independently, which could be accomplished by reading each of the N bits one by one, or alternatively, N amplifiers could be used to read the N bits in parallel. For synaptic plasticity, the desired weight update should be done with prior knowledge of the stored weight. Otherwise, a blind update could have a large detrimental effect, especially if the error is associated with devices representing the most significant bits. However, a direct comparison between these alternate architectures and our proposed scheme requires a detailed system-level investigation, which is beyond the scope of this paper. In summary, we propose a novel synaptic architecture comprising multiple memristive devices with non-ideal characteristics to efficiently implement learning in neural networks. This architecture is shown to overcome several significant challenges that are characteristic to nanoscale memristive devices proposed for synaptic implementation, such as the asymmetric conductance response, limitations in resolution and dynamic range, as well as device-level variability. The architecture is applicable to a wide range of neural networks and memristive technologies and is crossbar-compatible. The high potential of the concept is demonstrated experimentally in a large-scale SNN performing unsupervised learning. The proposed architecture and its experimental demonstration are a significant step towards the realization of highly efficient, large-scale neural networks based on memristive devices with typical, experimentally observed non-ideal characteristics."
} | 3,836 |
39472959 | PMC11520598 | pmc | 55 | {
"abstract": "Background Extensive research on the diversity and functional roles of the microorganisms associated with reef-building corals has been promoted as a consequence of the rapid global decline of coral reefs attributed to climate change. Several studies have highlighted the importance of coral‐associated algae ( Symbiodinium ) and bacteria and their potential roles in promoting coral host fitness and survival. However, the complex coral holobiont extends beyond these components to encompass other entities such as protists, fungi, and viruses. While each constituent has been individually investigated in corals, a comprehensive understanding of their collective roles is imperative for a holistic comprehension of coral health and resilience. Results The metagenomic analysis of the microbiome of the coral Oculina patagonica has revealed that fungi of the genera Aspergillus , Fusarium , and Rhizofagus together with the prokaryotic genera Streptomyces , Pseudomonas , and Bacillus were abundant members of the coral holobiont . This study also assessed changes in microeukaryotic, prokaryotic, and viral communities under three stress conditions: aquaria confinement, heat stress, and Vibrio infections. In general, stress conditions led to an increase in Rhodobacteraceae, Flavobacteraceae, and Vibrionaceae families, accompanied by a decrease in Streptomycetaceae. Concurrently, there was a significant decline in both the abundance and richness of microeukaryotic species and a reduction in genes associated with antimicrobial compound production by the coral itself, as well as by Symbiodinium and fungi. Conclusion Our findings suggest that the interplay between microeukaryotic and prokaryotic components of the coral holobiont may be disrupted by stress conditions, such as confinement, increase of seawater temperature, or Vibrio infection, leading to a dysbiosis in the global microbial community that may increase coral susceptibility to diseases. Further, microeukaryotic community seems to exert influence on the prokaryotic community dynamics, possibly through predation or the production of secondary metabolites with anti-bacterial activity. \n Video Abstract Supplementary Information The online version contains supplementary material available at 10.1186/s40168-024-01921-x.",
"conclusion": "Conclusions Taxonomic characterization of the O. patagonica holobiont showed a highly diverse microeukaryotic community. This, in addition to the well-known Symbiodinium , includes fungi from the genera Aspergillus , Fusarium , and Rhizophagus . Among prokaryotes, the most abundant genera found were the archaea Nitrosopumilus together with the bacteria Streptomyces , Pseudomonas , and Bacillus . Our aquaria experiments were designed to address changes on coral microbiome related to different stressors. Importantly, many of these microeukaryotic symbionts suffered a significant decline (below detectable levels) in O. patagonica , particularly under heat stress. Concurrently, a notable increase in the abundance of the prokaryotic community was observed, shedding light on the complex interplay between these two communities. Our data suggest that the microeukaryotic community potentially exerts influence on the prokaryotic community dynamics, possibly through predation or the production of secondary metabolites with anti-bacterial activity. The increase in the contribution of the prokaryotic community to the holobiont appears to coincide with an increase of CAZymes and genes responsible for DMSP catabolism. This correlation suggests the existence of a potential recycling mechanism for organic products generated by the holobiont under stress conditions. Thus, the alteration of conditions (such as temperature, confinement, and pathogen presence) within the coral holobiont appears to enhance the production of enzymes dedicated to carbohydrate degradation. This discovery underscores the dynamic and adaptive nature of microbial communities within coral ecosystems and offers exciting prospects for further research into the exploitation of these enzymatic resources in various industrial and environmental contexts.",
"introduction": "Introduction Corals form a dynamic meta-organism known as the coral holobiont, which involves a multipartite relationship between the cnidarian host, its endosymbiotic dinoflagellate algae (family Symbiodiniaceae; [ 1 ]), a diverse array of prokaryotes (Archaea and Bacteria), viruses, and eukaryotes (fungi and non-Symbiodiniaceae protists) [ 2 – 4 ]. Symbiodiniaceae, the primary photosymbiont in corals, has long been recognized for their crucial role in fulfilling most of the host energy needs, facilitating effective calcification and the formation of modern reefs [ 5 ]. Prokaryotes also play a pivotal role in the health and fitness of coral holobionts, and, although most of their specific roles and functions remain unknown, they are involved in nutrient acquisition, production of bioactive secondary metabolites, and probiotic mechanisms such as competitive exclusion of pathogenic bacteria [ 6 ].\n All these microorganisms coexist and interact with their host, maintaining the coral’s health and facilitating its ability to adapt to climate change. When environmental stressors, such as rising temperatures, occur, coral species often lose Symbiodiniaceae algae, leading to coral bleaching due to an energy deficit [ 7 – 9 ]. Additional changes can range from the replacement of mutualistic species by commensalism or parasitic species, as well as an increase in potential pathogens such as Vibrio species [ 10 – 12 ]. The simultaneous occurrence of these changes makes it difficult to determine whether the variations in microbial composition in thermally stressed corals are a consequence or cause of bleaching. The variability observed in coral microbial composition across seasons and spatial scales poses a significant challenge in unraveling the specific roles of microbial species within the coral holobiont under natural conditions. Consequently, many studies aiming to reduce complexity conduct experiments under controlled conditions, such as responses to pH gradients, nutrient fluxes, or thermal stress [ 13 – 16 ]. However, the applicability of data obtained from these controlled settings to natural conditions is still not well-understood. For example, in coral maintained in aquaria, some taxa rapidly decline, or even disappear, while other microorganisms remain unaffected for extended periods [ 17 , 18 ]. This phenomenon has also been observed in incubations of seawater samples, where a reduction in microbial diversity occurs, accompanied by the replacement of autotrophic microbes with heterotrophs [ 19 – 21 ]. Therefore, within confined corals, additional factors, not well-understood and beyond nutrient imbalances or temperature variations, may contribute to the observed variations. Omics tools have revolutionized coral research, allowing for a deeper understanding of the various partnerships within the coral holobiont. However, metagenomic approaches aiming to describe the entire holobiont face technical challenges, particularly due to the limited availability of genomes for individual holobiont components, especially microeukaryotes. In fact, only a few coral metagenomes have been sequenced, and most of these studies have analyzed specific partnerships separately, such as focusing on only the prokaryotic [ 22 , 23 ] or the viral community [ 24 ]. In this work, we conducted a metagenomic analysis of the entire Oculina patagonica microbiome. This coral, known as an opportunistic colonizer, is currently expanding throughout the Mediterranean Sea coast [ 25 – 27 ]. Its expansion is attributed in part to its broad tolerance for varying light levels and trophic conditions [ 28 , 29 ], as well as its heightened thermal resilience compared to other scleractinian corals [ 27 , 30 ]. However, O. patagonica is still susceptible to high temperature, which can lead to changes in the coral microbiome, coral bleaching, and, in some cases, irreversible mortality [ 28 , 31 – 33 ]. In the Western Mediterranean Sea, these coral bleaching episodes have been linked to the presence of two Vibrio species ( Vibrio coralliilyticus and Vibrio mediterranei ), whose pathogenicity increases when both bacteria infect the coral simultaneously [ 28 ]. The main objective of this study is to provide the first characterization of the O. patagonica microbiome and to investigate its changes in response to three different types of stress (confinement, thermal stress, and the presence of two Vibrio pathogens). For this purpose, we first described the O. patagonica holobiont in its natural environment and then compared it with that of corals maintained under controlled aquarium conditions. Our study sheds light on how alterations in the microeukaryotic community affect the prokaryotic community and whether these microbiome shifts result in a dysbiosis that may increase disease susceptibility in corals.",
"discussion": "Results and discussion To characterize the coral holobiont community, encompassing eukaryotes, prokaryotes, and viruses, a metagenome of approximately 70 Gb of sequence from a sample of O. patagonica collected during the summer of 2016 in Tabarca (Alicante, Spain) was sequenced. This is referred to as the Ocu-2016 dataset. Additionally, to investigate alterations in these communities in response to heat stress or the presence of Vibrio coral pathogens, we conducted complementary analyses using published metagenomes from Rubio-Portillo et al. [ 34 ]. The published metagenomes came from the same coral sample that had been split into several experimental regimes, though the published analysis primarily focused on characterizing the Vibrio assemblages and did not encompass the complete holobiont. For all nine metagenomes, Nonpareil analysis indicated a coverage of approximately 94% for Ocu-2016 and between 50 and 68% for the metagenomes from the aquaria (Supplementary Table 1), indicating a relatively good representation of the biodiversity contained within them. Oculina patagonica holobiont description First, to obtain a general overview, a comparison of all O. patagonica metagenomes (including the natural and the stressed samples) was conducted against other published coral and Mediterranean seawater datasets. As expected, samples from seawater clustered separately from the coral samples, which formed a cluster comprising metagenomes from O. patagonica and other corals (Supplementary Fig. 1). The reads annotated within the Ocu-2016 coral metagenome were distributed as follows: 10.86% belonged to Bacteria, 3.85% to Eukarya, 0.43% to Archaea, and 0.07% to Viruses, with 84.78% remaining unclassified (Supplementary Table 2). To assess whether this pattern was specific to our sample, a similar analysis was conducted using a dataset from the coral Porites lutea (SRR9182857) [ 23 ]. The results were comparable, with only 1.4% of the reads identified as eukaryotes and 7.7% classified as prokaryotes (Supplementary Table 2). The low percentage of eukaryotes found suggests that, due to the relatively lower availability of genomes from eukaryotes compared to prokaryotes, some of the unclassified reads may belong to the coral and other microeukaryotic organisms. O. patagonica associated eukaryotes Among coral symbionts, Symbiodoniaceae were the first and most important to be recognized, and the mutual transport of nutrients between both taxa has been well described [ 1 ]. O. patagonica harbors Symbiodinium species belonging to three different clades (A, B, and C) from which representative genomes are available (Supplementary Fig. 2). Among them, the genome of Breviolum minutum Mf1.05b (also known as Symbiodinium minutum ), which belongs to clade B, exhibited the highest read recruitment rates. This result was consistent with previous findings that identified Symbiodinium type B1 as the primary clade in O. patagonica [ 25 ]. It is also consistent with prior investigations in other coral species, demonstrating the concurrent association of corals with multiple Symbiodiniaceae, with a prevailing clone [ 66 ]. Besides Symbiodinium , other protists were also abundant in the O. patagonica holobiont, with the Evosea, Euglenozoa, and Apicomplexa phyla being particularly prevalent (Fig. 1 A). Previous studies have already documented the presence of other protists in the coral holobiont, suggesting that they may play a role in assisting coral hosts in obtaining sufficient nutrients or serve as an additional food source during recovery from stress, such as tissue loss and bleaching events [ 67 ]. Among the detected protists were two Choanoflagellate species, Monosiga brevicollis and Salpingoeca rosetta , which were not previously known to be associated with corals. The unexpected occurrence of reads matching the parasitic protist Plasmodium may have been attributed to the presence of the “ap1icoplast,” an organelle in Plasmodium and other apicomplexans [ 68 ] that is derived from endosymbiotic cyanobacteria. In other coral reef samples, sequences from apicomplexan-related plastids from Chromera and Vitrella have been consistently detected [ 69 , 70 ], but given the low recruitment of these genera in Ocu-2016, it is possible that many of the reads matching Plasmodium belong to a yet unknown microalgae plastid(s) present in O. patagonica . Fig. 1 Microbial taxonomic classification of O. patagonica microbiome inferred from metagenomic reads using Kaiju. Taxonomic classification at the phylum ( A ), family ( B ), and genera ( C ) levels Numerous non-photosynthetic microeukaryotes associated with corals have been identified, exhibiting diverse roles within the holobiont, ranging from beneficial symbiosis to parasitic relationships, and even acting as primary pathogens [ 71 – 73 ]. In this study, most of the non-photosynthetic microeukaryotic reads were classified as Fungi (Fig. 1 A). Among them, the most prominently represented phyla were Ascomycota, Basidiomycota, and Mucoromycota. Aspergillaceae and Glomeraceae were the two major families (Fig. 1 B), in good agreement with previous results reporting Ascomycetes and Basidiomycetes as the two major groups of fungi associated with corals [ 72 , 74 ]. Within the Ascomycota, the genera Aspergillus and Fusarium were the most abundant (Fig. 1 C). A significant abundance of reads assigned to the genus Rhizophagus (Glomeraceae family) was also detected. This is a beneficial mycorrhizal fungus commonly used as a soil inoculant in agriculture and forest ecosystems to enhance phosphorus uptake [ 75 ]. Recruitment analysis of Ocu-2016 reads against the reference genome of Rhizophagus irregularis (DAOM 181602) revealed the presence of reads with similarity to their ribosomal operons and housekeeping genes. Similar results were also obtained using the microbial metagenomes of the corals P. lutea and Acropora palmata (results not shown). These findings suggest that although Rhizophagus species have not been previously described in aquatic environments, related fungi may contribute beneficial traits to corals. O. patagonica associated prokaryotes Among the prokaryotic community, bacterial reads greatly outnumbered those assigned to archaea, maintaining a ratio of approximately 15:1 (Supplementary Table 2). The archaeal community (Fig. 1 A) was composed of Euryarchaeota (0.2% of the total reads) and Thaumarchaeota (0.1%). The genus Nitrosopumilus exhibited the highest abundance within this family (Fig. 1 C). This genus has been previously linked to the crucial ammonium oxidation process occurring within the coral mucus layer [ 76 ]. Notably, Nitrosopumilus has been previously detected in O. patagonica , and its presence has been exclusively associated with healthy colonies [ 77 ]. Also in accordance with previous findings obtained by 16S rRNA gene analyses [ 77 ], the bacterial reads were primarily composed of Proteobacteria (4.3%), Actinobacteria (1.6%), Firmicutes (1.2%), and Bacteroidetes (1.1%), with Streptomyces , Pseudomonas , and Bacillus as the most abundant genera (Fig. 1 ). Binning of the assembled contigs from the Ocu-2016 metagenome resulted in five metagenome-assembled genomes (MAGs), which belonged to Desulfobacterales (MAG1-Ocu2016), Flavobacteriales (MAG2-Ocu2016), Holosporales (MAG3-Ocu2016), Parvularculales (MAG4-Ocu2016), and Rhizobiales (MAG5-Ocu2016) (Supplementary Table 3). These MAGs recruited less than 0.001% of the metagenomics reads, with ANIr values over 99% (Supplementary Table 4), suggesting that only MAGs from low-abundance bacteria displaying a very low intra-population diversity could be retrieved. Excluding MAG1, the average genome size for the other four MAGs from the Ocu-2016 dataset was 1.7 ± 0.5 Mb, a size small enough to be consistent with a host-associated lifestyle, reflecting the loss of non-essential genes [ 78 ]. For example, MAG3-Ocu2016 ( Candidatus Hepatobacter penai; 87.3% complete) lacks key biosynthetic pathways for essential amino acids and shows incomplete synthesis of purines and pyrimidines (Supplementary Table 5). However, it harbors genes for biotin (vitamin B7) synthesis. This could be crucial for the coral holobiont, as both corals and Symbiodinium are suggested to be auxotrophs for different B vitamins, potentially obtaining them from associated bacteria [ 79 , 80 ]. Furthermore, within the MAGs, genes encoding mechanisms for stable symbiosis with the host, such as ankyrin repeats proteins (ARPs), were identified (Supplementary Table 3). ARPs are common protein interaction motifs that modulate intracellular processes, promoting stable symbiotic or pathogenic associations [ 81 ]. Previous analysis of ARP distribution in microbial genomes showed that species dedicating more than 0.2% of their protein-coding genes to ARPs are typically obligate intracellular or facultative host-associated species [ 82 ]. In this case, MAG3-Ocu2016 and MAG4-Ocu2016, both from the Alphaproteobacteria class, surpass this percentage. Moreover, in the case of MAG4-Ocu2016 (unknown Parvularculales), these proteins were affiliated with other corals and sponges, suggesting its potential as a symbiont for marine invertebrates. MAG1-Ocu2016 represents a putatively diazotrophic dissimilatory sulfate-reducing bacterium (SRB) belonging to the Desulfobacter genus (Supplementary Table 5). The identification of a N 2 -fixing SRB, not previously described in corals, carries substantial implications for coral communities that commonly inhabit nutrient-depleted environments. These bacteria may enhance the coral’s ability to efficiently convert gaseous nitrogen into a usable form such as ammonia, a vital process for sustaining a steady nitrogen supply for Symbiodinium -based primary production within corals [ 83 ]. O. patagonica associated virus A total of 275 different viral operational taxonomic units (vOTUs) were retrieved from Ocu-2016 metagenome, which accounted for 0.08% of the metagenomics reads. Only 14% of these vOTUs exhibited similarity to known reference genomes, and most of them (86%) were classified as dsDNA phages belonging to the class Caudoviricetes (Supplementary Table 6). A phage phylogenetic network using phage reference genomes (Fig. 2 ) showed that most of the phages identified in the O. patagonica holobiont were new. Others grouped mainly with phages infecting Pseudoalteromonas , Pseudomonas , and Roseobacter , three bacterial genera consistently associated with corals. Three vOTUs presented a partial similarity to Suoliviridae sequences (Crassvirales), with 30 to 63% similarity to a major head protein, DNA polymerase or hypothetical proteins. Although marine crassviruses have been previously described [ 84 ], O. patagonica assembled crassviruses-like clustered separately from them and may thus represent novel lineages. Fig. 2 vConTACT clustering of O. patagonica phages and prophages and related prokaryotic Viral RefSeq genomes (version 201). Color nodes represent the Oculina predicted viral sequences and black ones the RefSeq virus. Edges represent the vConTACT-generated similarity score between each pair of viruses (only similarity scores of ≥ 1 are included in the network). Highly similar viruses are positioned close together. Only reference viruses that are connected to ≥ 1 predicted phage are included in the network Five of the identified vOTUs were assigned as eukaryotic viruses belonging to the Phycodnaviridae family within the Nucleocytoplasmic Large DNA Viruses (NCLDVs) (Supplementary Table 6). The family Phycodnaviridae infects phytoplankton and has been previously detected in both heat-stressed corals and cultures of Symbiodinium spp., suggesting that these viruses may have a role in the destruction of algal symbionts or the dysfunction of symbiont–host mutualism, although the extent of such infections is unknown [ 85 – 87 ]. O. patagonica holobiont changes under stress conditions When the Ocu-2016 dataset (the natural holobiont metagenome) was compared with a collection of the metagenomes derived from the stress experiments, it clustered with the coral samples maintained at 20 °C, and apart from the 28 °C samples (Supplementary Fig. 1). Read-level analysis further supported these findings, with Ocu-2016 sharing 46.9% of reads with C20, only 15% with C28, and lower similarities were observed when compared to the Vibrio infection datasets (Fig. 3 ). Genus-level analysis using a resampling approach showed that coral microbiome composition significantly changed when corals were exposed to the different stress tested in the experiment (Supplementary Table 7). Further, Shannon diversity index was significantly higher in corals maintained in the aquaria compared to the Ocu-2016, mainly in corals under heat stresses and in corals in the presence of vibrios compared to the control ones (pairwise Hutcheson t -tests, all adjusted p -values 0, Supplementary Table 1). All these results highlight significant shifts in the microbial community of O. patagonica under heat stress, particularly in the presence of Vibrio species. These community shifts in the aquaria confinements were mainly enrichments of heterotrophic microbes. This phenomenon, known as the “bottle effect,” has been previously observed when seawater samples or corals are kept under prolonged confinement conditions [ 18 , 88 ]. Fig. 3 Sequences shared between metagenomes determined using an “all versus all” comparison of metagenomic reads. Numbers indicate the percentage of shared sequences among datasets O. patagonica holobiont changes due to aquarium confinement One of the most extensive effects of confinement was the decrease in the proportion of reads assigned to Eukarya within the metagenome from the coral maintained at 20 °C (from 3.85 to 2.2% compared to Ocu-2016), and the increase of bacterial reads (from 10.85 to 17%). A differential abundance test was used to identify genera with different proportions in Ocu-2016 compared to the coral maintained at 20 °C (Supplementary Table 8). This analysis showed that in the aquaria a decline in Symbiodinium , as well as other microeukaryotes and prokaryotes compared to Ocu-2016. Meanwhile, an expansion of the (“natural”) rare biosphere was observed under confinement conditions (Fig. 4 ). Microeukaryotes that proliferated in the aquarium experiments were mainly members of the Saccharomycodes genus, which were not detected in Ocu-2016 (Fig. 4 ). Most of these reads showed similarity to the yeast Saccharomycodes ludwigii , with some matching Hanseniaspora . The presence of these sugar-consuming microorganisms may be attributed to the increased carbohydrate content in the mucus of stressed corals [ 89 ]. Fig. 4 Prokaryotic genera that showed significant different proportions among treatments. C, control corals; I, infected corals; MS, corals in the presence of Vibrio monocultures; MX, corals in the presence of Vibrio co-culture. The numerical annotations 20 and 28 refer to corals maintained at 20 °C and 28 °C, respectively Also under confinement conditions, there was a notable increase in certain prokaryotic genera, including potential human pathogens like Acinetobacter , Bordetella , Neisseria , Klebsiella , and Salmonella , which were not detected in Ocu-2016 (Fig. 4 ). This shift could be linked to changes in the microeukaryotic community associated with the coral. It has been suggested that microeukaryotes play a role in regulating microbial communities within corals through processes like phagocytosis and the production of antimicrobial compounds [ 2 , 70 , 72 ]. Thus, a search for genes encoding the synthesis of secondary metabolites was conducted to investigate the potential antimicrobial activities of the O. patagonica microbiome. Genes encoding polyketide synthases (PKSs) and multienzymatic nonribosomal peptide synthetases (NRPSs) were detected. To determine whether they were predominantly produced by eukaryotes or prokaryotes, we compared the proportions of these genes in both the natural sample and the aquarium samples using the chi-square test. In the natural sample (Ocu-2016), these genes were found almost exclusively in eukaryotes (e.g., Symbiodinium , Scleractinia , Blastocladiomycota , and Dictyosteliales ), with no bacterial PKSs detected (Fig. 5 ). Conversely, in the aquarium-maintained corals, these genes were primarily produced by prokaryotic microorganisms. Statistically significant differences were observed between the proportions of eukaryotic and prokaryotic genes (adjusted- p -value < 0.00005), except for the samples C20 and MS20, where the differences were not statistically significant (adjusted- p -value > 0.05). The absence or reduction of these natural eukaryotic PKS could potentially lead to an increase in the proliferation of fast-growing organisms, as discussed below. This would agree with previous results indicating that the coral-associated fungi that decreased under confinement conditions, such as species from Aspergillus or Fusarium , displayed antimicrobial activities against human pathogenic bacteria [ 90 , 91 ]. Fig. 5 The taxonomic origins of the predicted polyketide synthase (PKS) clusters in O. patagonica . A Eukaryotic clusters and B bacterial clusters detected. C, control corals; I, infected corals; MS, corals in the presence of Vibrio monocultures; MX, corals in the presence of Vibrio co-culture. The numerical annotations 20 and 28 refer to corals maintained at 20 °C and 28 °C, respectively O. patagonica holobiont changes due to thermal stress under experimental conditions The proportion of annotated eukaryal reads in the metagenome from the coral maintained at 28 °C remained similar to that at 20 °C. However, there was a substantial increase in the bacterial reads, from 17 to 40%. The number of detected bacterial families also increased, from 112 at 20 °C to 171 at 28 °C (Supplementary Table 2). The differential abundance approach was used to identify genera with different proportions in corals maintained at 20 °C and 28 °C (Supplementary Table 8). This analysis identified various genera from the Rhodobacteraceae family, including Marivitia , Ruegeria , Loktanella , and Yoonia , as bacterial bloomers under heat-stress conditions (Fig. 4 ). In line with our results, the Rhodobacteraceae family has been recently suggested as an indicator species for thermal stress in corals [ 92 ]. Furthermore, a significant number of Rhodobacteraceae genera detected in our study are known to be involved in the breakdown of organic sulfur compounds like dimethylsulfoniopropionate (DMSP) and dimethylsulfide (DMS). The production of these compounds increases in corals under thermal stress [ 93 – 95 ]. In good agreement, genes responsible for DMSP catabolism ( dmdA , dmdB , dmdC , dddD , dddP , and dddL ) increased in abundance in corals kept in aquaria, particularly in corals maintained at 28 °C (from 0.00001% in Ocu-2016 to 0.0001–0.001% of reads). Furthermore, two MAGs recovered from the aquaria metagenomes (MAG9-MS28 and MAG10-MX28) corresponded to two Rhodobacteraceae genera potentially involved in DMSP metabolism, Pelagibaca and Yoonia , respectively (Supplementary Table 3). In fact, an orthologue gene encoding the DddL enzyme, responsible for DMSP catabolism, was detected in a Yoonia -related MAG, with a more pronounced increase in coral samples maintained at 28 °C (Fig. 6 ). This suggests that sulfur compounds produced in response to stressors, like confinement or thermal stress, may contribute to shaping coral-associated bacterial communities. This supports the hypothesis proposed by Raina et al. [ 96 ] that DMSP and DMS play a pivotal role in structuring coral-associated bacterial communities. Fig. 6 Relative abundances of MAGs. The relative abundance of each MAG was estimated using fragment recruitment analyses carried out by BLASTn comparisons. Only reads that matched with over 95% identity and 70% coverage were considered. The fraction of nucleotides mapping to the respective MAG was normalized by the length of that MAG and size of the metagenome Other genera that increased under confinement conditions and particularly under heat stress were Marinifilum , a member of the Bacteroidetes genus, and Halodesulfovibrio , originally affiliated with Desulfovibrio genus (Fig. 4 ). These genera have previously been observed in coral samples, and their proliferation under aquarium conditions has been documented by [ 97 ]. Marinifilum may play a role in the sulfur and carbon cycle, as well as lipid catabolism. On the other hand, Halodesulfovibrio has been identified as a secondary pathogen responsible for initiating and progressing black band disease in coral hosts, producing sulfide as a product of dissimilatory sulfate reduction [ 98 ]. The increase of bacterial bloomers under heat stress may be partially influenced by the recycling of metabolic waste products within the holobiont [ 99 ] and subsequent increase of available nutrients. Additionally, corals regularly release mucus into the surrounding seawater, which carries elevated concentrations of nutrients like organic carbon [ 100 ]. In heat-stressed corals, an increase in carbohydrate content in mucus has been detected [ 89 ], thus serving as a carbon source for heterotrophic microbes. To explore this idea, variation in the abundance of genes encoding carbohydrate-active enzymes (CAZymes) was examined. Compared to Ocu-2016, there was an increase in the proportion of glycosylhydrolases, glycosyltransferases, and carbohydrate-binding modules in corals maintained under experimental conditions, particularly in the thermally stressed ones (Supplementary Fig. 3). This suggests that mucus released under stressful conditions could provide newly accessible nutrients, serving as sustenance for fast-growing microbes. This process could potentially be enhanced by the decrease of other bacterial taxa and microeukaryotes under stress conditions, which may act as microbial regulators within the coral holobiont, as explained below. Regarding viruses, it is well established that virus-like particles (VLPs) increase in corals during stressful conditions or bleaching events [ 85 , 101 ]. In this case, recruitment analysis showed that one of the viruses, identified as Phycodnaviridae in the Ocu-2016 metagenome, clearly increased in abundance under thermal stress (Supplementary Fig. 4) when Symbiodinium abundances decreased. This fact suggests that this virus could be involved in the destruction of algal symbionts or the dysfunction of symbiont–host mutualism as mentioned above. O. patagonica holobiont changes due to Vibrio infection under experimental conditions All the above-described changes observed in confined or heat stress were enhanced in corals exposed to Vibrio cells ( V. coralliilyticus and V. mediterranei ), either directly or indirectly. Specifically, corals at 28 °C in the presence of Vibrio pathogens experienced a more significant increase in Vibrio species abundance compared to controls. Accordingly, a previous transcriptomic study carried out with these samples [ 34 ] revealed that pathogenic Vibrio release quorum sensing molecules, triggering alterations in coral-associated bacteria and an increase in other potential pathogens already present in the coral sample, thus explaining the increase of Vibrio spp. detected here. Furthermore, the increase in genera related to DMSP metabolism, such as Marivitia , Ruegeria , Loktanella , and Yoonia , was also more pronounced under the presence of Vibrio pathogens, which also suggests vibrios play a key role in coral microbiome modulation (Figs. 4 and 6 ).\n Analysis of the viral contigs in these coral metagenomes revealed a significant decrease in the proportion of genomes classified as prophages in corals experiencing heat stress (0.002–0.0004%) and in those exposed to Vibrio coral pathogens (0.001–0.0005%) compared to the natural sample (0.03%) (Fig. 7 A). To delve into the lysogenic dynamics within the coral holobiont, recruitment analyses were performed using only prophage genomes identified within bacterial contigs; this approach confirmed their integration into their respective bacterial hosts. Our data unveiled the induction of three distinct prophages in corals subjected to Vibrio infection (I20) or heat-induced stress (C28), and two of these prophages were found to be integrated within contigs belonging to the Rhodobacterales order (see Fig. 7 B). These findings align well with recent research suggesting that coral pathogens, such as V. coralliilyticus , produce hydrogen peroxide to initiate the lytic cycle of prophages in their competitors, thereby providing the coral pathogen with an advantage by reducing competition during coral colonization [ 102 ]. This mechanism further contributes to host dysbiosis by shifting the balance from symbionts to pathobionts as observed in the human gut [ 103 ]. Fig. 7 Fragment recruitment plots for metagenome sequence reads on bacterial contigs with prophages (marked in grey in each contig) from each coral. The vertical axis indicates the sequence identity of an alignment between a metagenomic sequence and the reference contig using BLASTn, the identity ranges from 100% (top) to 95% (bottom). *Prophages induced from bacterial host upon heat or Vibrio infection stress"
} | 8,651 |
25354956 | PMC4265203 | pmc | 56 | {
"abstract": "Although rising ocean temperatures threaten scleractinian corals and the reefs they construct, certain reef corals can acclimate to elevated temperatures to which they are rarely exposed in situ . Specimens of the model Indo-Pacific reef coral Pocillopora damicornis collected from upwelling reefs of Southern Taiwan were previously found to have survived a 36-week exposure to 30°C, a temperature they encounter infrequently and one that can elicit the breakdown of the coral–dinoflagellate (genus Symbiodinium ) endosymbiosis in many corals of the Pacific Ocean. To gain insight into the subcellular pathways utilized by both the coral hosts and their mutualistic Symbiodinium populations to acclimate to this temperature, mRNAs from both control (27°C) and high (30°C)-temperature samples were sequenced on an Illumina platform and assembled into a 236 435-contig transcriptome. These P. damicornis specimens were found to be ∼60% anthozoan and 40% microbe ( Symbiodinium , other eukaryotic microbes, and bacteria), from an mRNA-perspective. Furthermore, a significantly higher proportion of genes from the Symbiodinium compartment were differentially expressed after two weeks of exposure. Specifically, at elevated temperatures, Symbiodinium populations residing within the coral gastrodermal tissues were more likely to up-regulate the expression of genes encoding proteins involved in metabolism than their coral hosts. Collectively, these transcriptome-scale data suggest that the two members of this endosymbiosis have distinct strategies for acclimating to elevated temperatures that are expected to characterize many of Earth's coral reefs in the coming decades.",
"conclusion": "Conclusions This represents the first study to simultaneously measure global expression patterns of genes derived from two compartments of an endosymbiotic organism within an experimental framework. From a comprehensive assessment of the DEGs, it is evident that the Symbiodinium compartment responds more strongly at the mRNA level to a 2-week, but not a 36-week, elevated temperature exposure, relative to its host coral, the model reef-building coral P. damicornis . Furthermore, it was found that Symbiodinium were more likely than the host corals in which they resided to up-regulate metabolism-targeted genes at high temperature; this observation highlights the fact that the two members of this endosymbiosis have different sub-cellular strategies for acclimating to high temperatures. Host coral gfp-like cp molecules may have contributed to this acclimation capacity via the light-absorbing capacity of the proteins they encode, while the high expression of ubiqlig by the Symbiodinium populations may have allowed them to rapidly metabolize proteins via the proteasome that had become denatured by high temperatures. Collectively, then, it appears that host control of the light environment, Symbiodinium protein turnover, and both host coral and Symbiodinium metabolism are modulated by elevated temperature, and future work will attempt to decipher which of these processes is most critical for coral acclimatization to temperatures they are likely to face in the coming decades as part of GCC.",
"introduction": "Introduction Coral reefs have been hypothesized to undergo extensive degradation in the coming decades due to both local ( Fabricius 2005 ) and global-scale ( Hoegh-Guldberg et al. 2007 ) anthropogenic impacts. Specifically, increasing seawater temperature and acidity associated with global climate change (GCC) could potentially lead to more frequent episodes of coral bleaching ( Veron 2011 ), whereby the endosymbiotic association between the anthozoan host and its gastrodermal endosymbionts of the genus Symbiodinium breaks down ( Gates 1990 ). However, most GCC models do not account for acclimation, which has recently shown to be notable in reef-building corals exposed to increases in temperature ( e.g., \n Mayfield et al. 2011 , 2012a , 2013a ; Barshis et al. 2013 ; Palumbi et al. 2014 ). For instance, populations of the model reef-building coral Pocillopora damicorni s from upwelling reefs in Southern Taiwan were found to acclimate for nine months to a temperature (30°C) they encounter in situ for only several hours on an annual timescale ( Mayfield et al. 2013b ), although the physiological mechanisms by which such acclimation occurred are still poorly understood. In recent years, molecular biology-driven techniques have advanced the collective knowledge of both the cell biology of coral–dinoflagellate endosymbioses ( e.g., \n Mayfield et al. 2009 , 2010 , 2012b , 2014b ; Peng et al. 2011 ; Chen et al. 2012 ; Wang et al. 2013 ) and their responses to environmental changes ( e.g., \n Mayfield et al. 2013c , d ; Putnam et al. 2013 ). Microarrays ( e.g., \n DeSalvo et al. 2008 ) and next-generation sequencing (NGS; e.g., \n Vidal-Dupiol et al. 2013 ) are particularly promising approaches for addressing an array of functional questions in this globally important mutualism; however, only one work to date ( Shinzato et al. 2014 ) has considered both endosymbiotic compartments in its transcriptomic characterization of a reef coral, and this study was not set within an experimental framework. Indeed, it has been posed that both members of the coral– Symbiodinium endosymbiosis should be acknowledged when attempting to gauge the response of this mutualism to environmental shifts ( Fitt et al. 2009 ). Inhibition of the photosynthetic pathways of Symbiodinium in response to high light and temperature is known to elicit reactive oxygen species (ROS) production, a phenomenon that can ultimately lead to bleaching ( Lesser 1997 ). These findings suggest to many that the Symbiodinium populations represent the more sensitive compartment to environmental change, although the host coral surely plays an important role in the stress (or acclimation) response of the holobiont, as well ( Mayfield & Gates 2007 ). NGS-based mRNA sequencing ( i.e., ‘RNA-Seq’) could not only aid in uncovering the genetic basis of acclimation to 30°C in P. damicornis , but it could also directly test the hypothesis that one compartment of the endosymbiosis is more responsive to temperature change at the subcellular level ( sensu \n Leggat et al. 2011 ). In fact, a series of recent studies have reported that Symbiodinium does not show a marked response at the mRNA level to elevated temperatures ( e.g., \n Boldt et al. 2010 ; Rosic et al. 2011 ; Mayfield et al. 2014a ), although transcriptome-scale approaches may better clarify whether or not this is a ubiquitous property of in hospite Symbiodinium populations. Herein, mRNAs from P. damicornis samples that had been exposed to a control (27°C) or high (30°C) temperature for 2 or 36 weeks ( n = 3 replicates for each treatment–time) were sequenced with an Illumina Tru-Seq™ kit (ver. 2) on an Illumina Genome Analyzer IIx (GAIIx), and the resulting sequences were assembled into a meta-transcriptome against which mRNA expression of the 12 individual samples was assessed. It was hypothesized that the transcriptome-scale data set generated herein could aid in the development of a subcellular mechanism by which P. damicornis specimens from upwelling reefs of Southern Taiwan were able to acclimate to high temperature over a multi-season timescale.",
"discussion": "Discussion What is Pocillopora damicornis ? This Illumina-based P. damicornis transcriptome is larger than a previous, 454-based one ( Traylor-Knowles et al. 2011 ); an additional ∼175 000 contigs were produced herein, and only ∼3000 contigs in the Traylor-Knowles et al . (2011) transcriptome were not identified. This discrepancy is partially due to the previous assembly containing only a small proportion (<5%) of Symbiodinium genes, but could also be due to the sequencing of a large number of splice variants or even non-coding RNAs (despite the employment of a poly-A-selection step) in the present work, of which only ∼30% of the contigs could be confidently assigned a protein identity. The low percentage of Symbiodinium genes in the previous P. damicornis transcriptome ( Traylor-Knowles et al. 2011 ) is in stark contrast to not only the results obtained herein (∼20%), but also those of Shinzato et al . (2014 ; ∼35%). The differences in the host: Symbiodinium contig ratio between this study and Shinzato et al . (2014) , which was conducted with the massive coral Porites australiensis , may highlight the fact that not all corals have a similar biological composition. It should be mentioned, however, that the in vivo biomass ratio may not necessarily approximate the mRNA ratio; it is likely that the former is closer to 50/50% for many anthozoan–dinoflagellate endosymbioses based on microscopic images of Symbiodinium in hospite ( Chen et al. 2012 ; Mayfield et al. 2013b ), in which the majority of the gastrodermal volume is occupied by these dinoflagellates. It is also possible that methodological differences accounted, in part, for the differences in host/ Symbiodinium ratio between this study and that of Shinzato et al . (2014) ; both the host coral and Symbiodinium transcriptomes were assembled de novo herein, and the origin of the assembled contigs was hypothesized based on alignments to published sequences queried across the entire NCBI database. In contrast, Shinzato et al . (2014) aligned their host and Symbiodinium transcriptomes directly to published host coral ( A. digitifera ; Shinzato et al. 2011 ) and Symbiodinium ( Shoguchi et al. 2013 ) genomes, respectively; doing so may have allowed for the assignment of a higher percentage of Symbiodinium contigs. Symbiodinium vs. host coral mRNA-level response to high-temperature exposure The clade C Symbiodinium populations of the samples sequenced herein demonstrated a more pronounced mRNA-level response than their coral hosts after 2 weeks of exposure to 30°C; specifically, a higher number and percentage of DEGs were detected in the Symbiodinium compartment at the 2-week, but not the 36-week, sampling time at a 10 −3 α level. Whether the Symbiodinium transcriptome always responds more strongly to elevated temperature than that of their hosts remains to be determined; for instance, it could be that the host coral transcriptome changed dramatically after several hours of exposure to elevated temperature, with gene expression levels quickly returning to baseline before the first sampling time. Future work should seek to assess transcriptome-wide changes in this coral holobiont over a more fine-tuned timescale to determine whether this is indeed the case. High-temperature effects on coral metabolism and osmoregulation Although genes related to metabolism compromised ∼30% of the host coral and Symbiodinium reference assemblies, ∼45% of the Symbiodinium and host DEGs were involved in metabolism (Fig. 2 ), suggesting that this process may have been altered in corals exposed to high temperatures. Whereas their coral hosts were found to down-regulate the expression of metabolism genes at high temperature, Symbiodinium were more likely to up-regulate genes involved in metabolism (although zifl1-l represents an exception). This could mean that the Symbiodinium populations incubated at high temperatures were exhibiting elevated metabolic rates, a hallmark of many organisms exposed to abnormally elevated temperatures ( Hochachka & Somero 2002 ). Such elevated metabolic rates in the Symbiodinium populations of the high-temperature samples could theoretically have driven the metabolism gene expression changes documented in their hosts. Symbiodinium populations within corals exposed to elevated temperatures typically become photoinhibited ( Jones et al. 2000 ), at which point flux of the food source and compatible osmolyte glycerol may be diminished ( Gates & Edmunds 1999 ). As most cnidarians rely on glycerol and similar small organic compounds to establish their osmotic pressure ( Shick 1991 ), coral respiration of the remaining glycerol pools may cause a collapse of the host cytoskeleton over the Symbiodinium cell(s) in the coral gastroderm ( Mayfield et al. 2010 ). In contrast, the increases in Symbiodinium metabolism inferred herein from the gene expression data may have driven increases in osmolyte flux into the host gastroderm, causing both the down-regulation of many metabolism-targeted genes and the gastrodermal tissue swelling revealed previously by scanning electron microscopy ( Mayfield et al. 2013b ). The role of chromoproteins in coral acclimation to high temperature This increase in gastrodermal thickness was only observed in samples analysed after 36 weeks of high-temperature exposure ( Mayfield et al. 2013b ), which was during the boreal summer. Given the higher light levels experienced by all corals at this time ( Mayfield et al. 2013b ), the thicker gastroderms of the high-temperature samples could also have been necessary to provide a greater space in which GFP-like CPs could absorb high light levels that might otherwise have caused Symbiodinium photoinhibition ( Mayfield et al. 2014c ). As both high light and temperature are typically needed to elicit bleaching ( Hoegh-Guldberg 1999 ), such shading by gastrodermal GFP-like CPs ( sensu \n Smith et al. 2013 ), whose respective gene mRNA was expressed at 90-fold higher levels by high-temperature corals, might have allowed for a greater degree of self-shading that ultimately allowed these corals to acclimate to high temperatures during the summer months. Future work will attempt to verify these chromoprotein expression differences at the protein level, as well as to determine their capacity to buffer the intra-gastrodermal Symbiodinium populations from excessive irradiation. Such studies may also elucidate whether photoprotection or osmoregulatory/metabolic influences were more important in driving the increases in gastrodermal thickness witnessed in these high-temperature samples ( Mayfield et al. 2013b ). Ubiquitin ligase, protein turnover, and coral acclimation to elevated temperatures In addition to the inferred increases in Symbiodinium metabolism and enhanced shading by GFP-like CP under high light levels, ubiquitin ligase proteins may also have aided in the ability of these corals to acclimate to high temperatures given the dramatic increases in expression of the respective gene, ubiqlig , in Symbiodinium populations exposed to 30°C for 2 weeks. Ubiquitin ligases are involved in tagging proteins to be degraded by the proteasome ( Welchman et al. 2005 ), and elevated levels of ubiquitin-conjugated proteins were associated with corals inhabiting thermally extreme backreefs in Samoa ( Barshis et al. 2010 ). High levels of ubiquitin ligase protein expression are likely to be associated with high protein turnover rates, and while Symbiodinium protein turnover was not measured herein, the elevated levels of Symbiodinium ubiglig expression at high temperature may provide evidence for this phenomenon. Gates & Edmunds (1999) suggested that corals characterized by high levels of protein turnover should have a consequently greater capacity for acclimatization, as is the case for a plethora of other organisms ( Hochachka & Somero 2002 ); as such, the high protein turnover rates of the Symbiodinium compartment in corals exposed to high temperatures for 2 weeks might have, for instance, aided in their ability to readily catabolize and process temperature-denatured proteins in an efficient manner and hence allowed for their ultimate acclimation. Although host coral protein turnover was not measured herein, genes involved in protein homeostasis were actually under-represented in the host coral DEG pool; this may mean that protein processing/turnover may be less important in the acclimation response of the coral host. Furthermore, P. damicornis has been hypothesized, but not directly shown, to demonstrate low rates of protein turnover given its high growth rates and low metabolic rates relative to those of massive poritids ( Loya et al. 2001 ), which have repeatedly been found to be among the most resilient corals to environmental change ( Brown 1997 ). Therefore, the low protein turnover rates hypothesized by a combined assessment of conjectures put forth by Gates & Edmunds (1999) and Loya et al . (2001) , in conjunction with the metabolic suppression inferred from the down-regulation of a multitude of metabolism-targeted genes in high-temperature samples herein, may have allowed such high-temperature nubbins to continue to grow at comparable rates to controls ( Mayfield et al. 2013b ) over the duration of the experiment by, for instance, conserving metabolic energy for growth-related processes."
} | 4,220 |
31817956 | PMC6947318 | pmc | 57 | {
"abstract": "Memristor crossbar arrays without selector devices, such as complementary-metal oxide semiconductor (CMOS) devices, are a potential for realizing neuromorphic computing systems. However, wire resistance of metal wires is one of the factors that degrade the performance of memristor crossbar circuits. In this work, we propose a wire resistance modeling method and a parasitic resistance-adapted programming scheme to reduce the impact of wire resistance in a memristor crossbar-based neuromorphic computing system. The equivalent wire resistances for the cells are estimated by analyzing the crossbar circuit using the superposition theorem. For the conventional programming scheme, the connection matrix composed of the target memristance values is used for crossbar array programming. In the proposed parasitic resistance-adapted programming scheme, the connection matrix is updated before it is used for crossbar array programming to compensate the equivalent wire resistance. The updated connection matrix is obtained by subtracting the equivalent connection matrix from the original connection matrix. The circuit simulations are performed to test the proposed wire resistance modeling method and the parasitic resistance-adapted programming scheme. The simulation results showed that the discrepancy of the output voltages of the crossbar between the conventional wire resistance modeling method and the proposed wire resistance modeling method is as low as 2.9% when wire resistance varied from 0.5 to 3.0 Ω. The recognition rate of the memristor crossbar with the conventional programming scheme is 99%, 95%, 81%, and 65% when wire resistance is set to be 1.5, 2.0, 2.5, and 3.0 Ω, respectively. By contrast, the memristor crossbar with the proposed parasitic resistance-adapted programming scheme can maintain the recognition as high as 100% when wire resistance is as high as 3.0 Ω.",
"conclusion": "4. Conclusions Wire resistance is one of the factors that degrade the performance of the crossbar circuits significantly. In this work, we proposed a parasitic resistance-adapted programming scheme to mitigate the impact of wire resistance in memristor crossbar array. Firstly, a wire resistance modeling method using equivalent wire resistance matrix was proposed. The equivalent wire resistance matrix was achieved by analysis the crossbar circuit using the superposition method. The connection matrix was updated before it was used as a target for memristor crossbar programming. The updated connection matrix was obtained by subtracting the proposed equivalent wire resistance matrix from the original connection matrix. The circuit simulations were performed to verify the proposed wire resistance modeling method and the parasitic resistance-adapted programming scheme. The simulation results showed that the discrepancy of the output voltages of the crossbar circuit between the conventional wire resistance modeling method and the proposed wire resistance modeling method was as low as 2.9% when wire resistance varied from 0.5 to 3.0 Ω. The recognition rate of the memristor crossbar with conventional programming scheme was 99%, 95%, 81%, and 65% when wire resistance was set to be 1.5, 2.0, 2.5, and 3.0 Ω, respectively. By contrast, the memristor crossbar with the proposed parasitic resistance-adapted programming scheme could maintain the recognition as high as 100% when wire resistance was as high as 3.0 Ω.",
"introduction": "1. Introduction Neuromorphic computing was investigated by C. Mead in the late 1980s as a hardware-based approach for artificial intelligence [ 1 ]. The word “Neuromorphic” refers to an electronic circuit that is based on digital and analog components to mimic the neurobiological structures in nervous systems. Neuromorphic computing systems can be implemented on various VLSI (very-large scale integration) systems [ 2 , 3 , 4 , 5 , 6 ]. The prevailing VLSI technology today comprises mainly of CMOS (complementary-metal oxide semiconductor) devices. However, CMOS technology is approaching the end of their capabilities because scaling CMOS down faces several fundamental limiting factors stemming from electron thermal energy and quantum-mechanical tunneling [ 7 , 8 ]. The emerging memristive devices, termed memristors, have been considered a promising candidate for realizing the neuromorphic computing systems. Memristor was postulated by L. O. Chua in 1971 as the fourth fundamental passive circuit element and experimentally demonstrated by HP (Hewlett Packard) Labs in 2008 [ 9 , 10 ]. Memristors has been potentially used to implement the neuromorphic computing systems because the nonlinear relationship between magnetic flux and electric charge of memristors is very similar to the plasticity behavior of biological brain [ 11 , 12 ]. In biological brains, synapse is the connection between a presynaptic neuron and a postsynaptic neuron. The strength of a synapse is represented by a synaptic weight. According to the neuron activities including excitatory and inhibitory, synaptic weights can be positive or negative [ 13 , 14 ]. Synapses can be modeled by memristors as shown in Figure 1 [ 11 ]. The synaptic weight is represented by the conductance of memristor, which can increase or decrease according to the current flowing through the device. A memristor crossbar array is a fully connected mesh of perpendicular wires, in which any two crossing wires are connected by a memristor [ 15 ]. Neuromorphic computing systems employing crossbar architecture of memristors have gained more advantages in terms of the flexibility, power consumption, cost, and area [ 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 ]. Miao Hu et al. proposed a crossbar architecture of synaptic array composing of a plus and minus crossbar arrays representing plus- and minus-polarity connection matrices for analog neuromorphic computing [ 20 ]. To reduce the area and power consumption, S. N. Truong proposed a new memristor crossbar architecture, which is composed of a single memristor array and a constant-term circuit [ 21 ]. The proposed architecture can reduce the power consumption by 48% and the area by 50% [ 21 ]. The memristor crossbar has also applied to the applications of speech recognition and image recognition [ 22 , 23 ]. In a memristor crossbar array, some amount of voltage drop can be caused by parasitic resistance, also known as wire resistance along the row and the column lines [ 19 , 24 , 25 , 26 , 27 , 28 ]. Hereinafter “wire resistance” and “parasitic resistance” are used interchangeably. The impact of wire resistance becomes inevitable when the array size increases [ 22 ]. To mitigate the impact of wire resistance, several interesting schemes were proposed [ 25 , 26 , 27 , 28 ]. A design methodology has been proposed to reduce the impact of wire resistance in a one-selector-one resistive device (1S1R) crossbar array [ 27 ]. The proposed design methodology seems to be complicated since the physical specification of the devices must be considered [ 27 ]. Another approach to deal with the wire resistance is to use a dynamic reference scheme [ 25 ]. The read operation is performed with two steps associated with a special reading circuit. [ 25 ]. These proposed schemes are effective when they are applied to a memristor crossbar array, in which memristors are used as binary switches between two distinct high and low resistance states (HRS (High Resistance State) and LRS (Low Resistance State)). These solutions are mainly based on the additional techniques or circuits to compensate the variation of reading voltage caused by wire resistance. To the best of our knowledge, there is a lack of the techniques that can be applied to the programming process of crossbar circuit to lessen the impact of wire resistance in the inference process. In this work, we propose a parasitic resistance-adapted programming scheme for memristor crossbar-based neuromorphic computing systems, in which memristors are used as analog connections. An equivalent wire resistance is proposed for modeling wire resistance in crossbar circuit. The proposed equivalent wire resistance matrix is used to compensate wire resistance during the programming process. As the result, the impact of wire resistance in inference process is reduced significantly."
} | 2,068 |
26175666 | PMC4485154 | pmc | 58 | {
"abstract": "Memristive devices are popular among neuromorphic engineers for their ability to emulate forms of spike-driven synaptic plasticity by applying specific voltage and current waveforms at their two terminals. In this paper, we investigate spike-timing dependent plasticity (STDP) with a single pairing of one presynaptic voltage spike and one post-synaptic voltage spike in a BiFeO 3 memristive device. In most memristive materials the learning window is primarily a function of the material characteristics and not of the applied waveform. In contrast, we show that the analog resistive switching of the developed artificial synapses allows to adjust the learning time constant of the STDP function from 25 ms to 125 μs via the duration of applied voltage spikes. Also, as the induced weight change may degrade, we investigate the remanence of the resistance change for several hours after analog resistive switching, thus emulating the processes expected in biological synapses. As the power consumption is a major constraint in neuromorphic circuits, we show methods to reduce the consumed energy per setting pulse to only 4.5 pJ in the developed artificial synapses.",
"introduction": "Introduction Since the discovery of spike-timing dependent plasticity (STDP) in biological synapses (Bi and Poo, 1998 ; Snider, 2008 ; Di Lorenzo and Victor, 2013 ), scientists have been captivated by the idea of changing the synaptic weight, i.e., the strength between the pre- and post-neuron, in bioinspired electronic systems in a fashion similar to biology (Indiveri et al., 2006 ). However, the circuit-oriented approach is complicated because the “synaptic weight” variable has to be stored typically either as charge in a capacitor (Koickal et al., 2006 ) or even digitally in neuromorphic IC (Schemmel et al., 2012 ; Mayr et al., 2013 ). This adds circuit complexity and increases energy consumption (Indiveri et al., 2006 ; Adee, 2009 ; Ananthanarayanan et al., 2009 ). Therefore, nonvolatile analog resistive switches, namely resistive random-access memory (RRAM) or memristors (Chua, 1971 ; Du et al., 2013 ), responding to well-defined input signals by suitably changing their internal state (“weight”) are currently developed. For example, the emulation of STDP with 60–80 pairings of pre- and post-synaptic spikes has been shown for artificial synapses based on memristive TiO x (Seo et al., 2011 ; Thomas and Kaltschmidt, 2014 ), WO x (Chang et al., 2011 ), HfO x (Yu et al., 2011 ), GST (Kuzum et al., 2012 ), and on the memristive BiFeO 3 (Mayr et al., 2012 ; Cederström et al., 2013 ). Figure 1A shows a memristor between the electrical Integrate & Fire (I&F) neurons. The synaptic weight of the memristor can be controlled by the time delay Δt between pre- and post-spike from the 1st layer I&F neuron (Figure 1A ) (Zamarreño-Ramos et al., 2011 ). The 2nd layer I&F neuron sums up the signals from all incoming neurons and generates voltage spikes transmitted to other neurons (not shown) through memristor-based artificial synapses. The memristive BiFeO 3 (BFO) can serve as an analog resistive switch (Shuai et al., 2011 ) with multiple distinguishable low resistance states (LRSs) (Shuai et al., 2013 ; Jin et al., 2014 ) and with a single detectable high resistance state (HRS). Due to the thermal diffusion of Ti atoms and their substitutional incorporation into the lower part of the BiFeO 3 (BFO) layer during BFO thin film growth on a Pt/Ti bottom electrode, the barrier at the Pt/Ti bottom electrode is flexible. Figure 1 (A) Schematic illustration of the memristor-based synaptic electronics. The artificial synapses are placed between Integrate & Fire neurons (I&F neuron). With a well-defined time delay Δt between the pre- and post-spikes the internal state (“weight”) of the memristor is suitably changed. (B) Hysteretic current-voltage (IV) characteristics of a Au/BiFeO 3 /Pt memristor in LRS and HRS with a top electrode area of 4.5E4 μm 2 under source voltages with maximum sweeping pulse amplitude of 8.5 V and a pulse width of 100 ms. The current in high resistance state I HRS and in low resistance state I LRS is read out at +2.0 V, after having switched the memristor into HRS and LRS, respectively. The long term potentiation current I LTP and the long term depression current I LTD lie below the reading current in LRS (I LRS ) and HRS (I HRS ). Inset shows the structure of a BFO memristor. (C) Schematic demonstration of the distribution of fixed Ti 4+ , fixed Fe 3+ and mobile V + o . Earlier we have shown that STDP and triplet plasticity with learning windows on the millisecond time scale can be faithfully emulated on BFO-based artificial synapses by applying 60–80 pairings of pre- and post-synaptic spikes (Mayr et al., 2012 ; Cederström et al., 2013 ). In this work we investigate a significantly wider range of timescale configurability, ranging from 25 ms to 125 μs. To the best of our knowledge, this kind of timescale configurability has not been shown in memristive synapses before. We also examine the evolution of the induced memristive weight change over time and provide several power consumption figures. By increasing the programming voltage (HRS/LRS writing pulse amplitude), it is possible to decrease the switching pulse width as well as the power consumption during a single STDP writing process on BFO-based artificial synapses. Furthermore, the increased programming voltage also shortens the total pairing spike time, and enables to move from the standard biology-like 60–80 spike pairing STDP experiment to a single pairing STDP experiment that results in the same weight/memristance change. Our work is structured as follows: In Section Materials and Methods, we describe the non-volatile resistive switching of BFO–based artificial synapses and introduce the single pairing STDP pulse sequence. In Section Results, we present the measured learning window, memory consolidation, and energy consumption of the single pairing STDP in BFO-based artificial synapses and discuss configurability, energy consumption, and retention of weight change in Section Discussion. The paper is summarized and an outlook is given in Section Summary and Outlook.",
"discussion": "Discussion Configurability In this work single pairing STDP in BFO-based artificial synapses has been demonstrated for emulating the functionality and the plasticity of biological synapses. The waveform-defined plasticity of BFO memristors in addition to their multilevel memristive programming capability enables easy control of the STDP time windows, as evidenced by the three orders of magnitude timescale configurability shown in this paper. While there has been a lot of simulation work on this topic, the number of devices where STDP or variations have actually been implemented and measured is still fairly small (Jo et al., 2010 ; Alibart et al., 2012 ). Among those, our highly-configurable, finely grained learning curves are unique, other implementations exhibit statistical variations (Jo et al., 2010 ), can only assume a few discrete levels (Alibart et al., 2012 ) or the learning windows are device-inherent, i.e., cannot be adjusted (Ohno et al., 2011 ). We expect that for BFO-based artificial synapses at least 32/64 levels are possible in a power efficient manner. In addition, the wide range of timescales possible in BFO-based synapses enables e.g., a timebase-tunable system that could learn a classification offline in an accelerated manner, while still able to interact with real-time sensors before or after this learning. As mentioned in the introduction, BFO-based artificial synapses can be used for conventional STDP experiments, where only multiple spike pairings exhibit significant weight change, as well as in the mode used in this paper, where a single pairing already induces a significant weight change. By changing the voltage of the pre- and post-synaptic pulses, any point in between these two extremes can also be chosen, again showing the excellent configurability of BFO-based artificial synapses. However, the versatility of BFO memristors comes at the price that in contrast to e.g., phase-change materials, BFO is not easily integrated on top of CMOS (Shuai et al., 2013 ). Energy consumption In Table 2 , we have shown an energy consumption of E = 4.7 pJ in a BFO-based artificial synapse with electrode size of 4.52E4 μm 2 . While this is still three orders of magnitude above the energy consumption of biological synapses, it is one of the lowest reported so far for other artificial synapses. Compared to neuromorphic approaches, all memristive approaches are several orders of magnitude better (Azghadi et al., 2014 ). In terms of absolute area, the BFO memristor is comparable to some neuromorphic implementations (Hasler and Marr, 2013 ; Noack et al., 2015 ), but not competitive with memristor crossbar devices, as we are employing a single device test structure that has a large contact size for reasons of convenience. However, BFO device scaling is well established, thus we can aggressively scale the size of the top electrode to 10 μm 2 and the thickness of the BFO to 100 nm (Jin et al., 2014 ). For BFO with larger electrode area size, the current scales linearly with area size. For smaller electrode area size we would expect that the current scales with the number of BFO crystallites below the electrode. And in the limit case of nanoscale electrodes, the smallest possible current should be the current through single BFO crystallites. Retention of weight change We have investigated the retention of memristance weight change across time. As Figure 5A shows, the basic shape of the STDP curves is preserved across time. Figure 5B illustrates that even after memory consolidation, we retain a graded weight, i.e., a unimodal weight distribution. Our synapse does not collapse in either a potentiated or depressed (bimodal) distribution as predicted in some synaptic models (Fusi et al., 2000 ; Clopath et al., 2008 ). In memristive literature, there is usually no investigation of these phenomena, the weight change is taken at some unspecified time after induction and then assumed to be non-volatile. Only very few articles have investigated the actual non-volatility/weight retention across time and shown that the assumption of a non-volatile change is not necessarily valid (Chang et al., 2011 ). Thus, compared to other reports, this article gives a neuromorphic designer a clear guide on how to utilize the memristive synapses for long-term storage. Interestingly, this investigation of memory consolidation is also somewhat missing in the original biological measurements. Usually, data on the weight evolution ca. 30–60 min after induction is provided, but only on single example pairing experiments. These data points show various behaviors, from unchanged weights after initial weight induction (Froemke and Dan, 2002 ) to increases of weight change across time (Bi and Poo, 1998 ), decreases across time (Markram et al., 1997 ) or slow oscillations around the initial potentiated/depressed weight value (Sjöström et al., 2001 ). However, it is unclear how the overall STDP window consolidates over time. Thus, measuring the evolution of an STDP curve across time after induction at biological synapses similar to our investigation on memristive synapses may actually be a quite interesting scientific question."
} | 2,860 |
39729346 | PMC11759055 | pmc | 59 | {
"abstract": "Conventional artificial intelligence (AI) systems are\nfacing bottlenecks\ndue to the fundamental mismatches between AI models, which rely on\nparallel, in-memory, and dynamic computation, and traditional transistors,\nwhich have been designed and optimized for sequential logic operations.\nThis calls for the development of novel computing units beyond transistors.\nInspired by the high efficiency and adaptability of biological neural\nnetworks, computing systems mimicking the capabilities of biological\nstructures are gaining more attention. Ion-based memristive devices\n(IMDs), owing to the intrinsic functional similarities to their biological\ncounterparts, hold significant promise for implementing emerging neuromorphic\nlearning and computing algorithms. In this article, we review the\nfundamental mechanisms of IMDs based on ion drift and diffusion to\nelucidate the origins of their diverse dynamics. We then examine how\nthese mechanisms operate within different materials to enable IMDs\nwith various types of switching behaviors, leading to a wide range\nof applications, from emulating biological components to realizing\nspecialized computing requirements. Furthermore, we explore the potential\nfor IMDs to be modified and tuned to achieve customized dynamics,\nwhich positions them as one of the most promising hardware candidates\nfor executing bioinspired algorithms with unique specifications. Finally,\nwe identify the challenges currently facing IMDs that hinder their\nwidespread usage and highlight emerging research directions that could\nsignificantly benefit from incorporating IMDs.",
"introduction": "1 Introduction Artificial intelligence\n(AI) has experienced unprecedented growth\nin the past decade. The latest AI models demonstrate their remarkable\ncapabilities across various domains, including conversational agents\n(ChatGPT), gaming (AlphaGo Zero), research (AlphaFold), and art (Sora,\nSuno, and Midjourney), even surpassing human performance in some aspects. 1 − 6 However, these impressive achievements come at a significant cost\nin terms of money, energy, and time during the training process. The\ndemand for computing resources to train state-of-the-art AI models\nis doubling every two months, a rate that far outpaces the historical\nrate of Moore’s law. 7 This unsustainable\ntrend necessitates the exploration of more efficient devices, systems,\nand algorithms for future AI. Biological systems, in contrast, consume\nsignificantly less energy and occupy much less space than current\nAI systems, while exhibiting more comprehensive learning capabilities\nand more robust operational performance. Inspired by this efficiency,\nresearchers have been striving to emulate the behaviors of biological\nneural networks (BNNs) to develop bioinspired computing systems that\nare both efficient and powerful. 8 , 9 BNNs exhibit complex\ndynamics originating from the transport of\ndiscrete units, such as the flow of electrolyte ions and neurotransmitters,\nwhich introduce stochastic variations and are believed to be the key\nto the effectiveness of BNNs. 10 , 11 Traditional computing\ncores, such as central processing units (CPUs) and graphics processing\nunits (GPUs), are designed for deterministic digital computing. They\nare very inefficient when implementing bioinspired algorithms like\nspiking neural networks (SNNs). To address this, customized complementary\nmetal-oxide-semiconductor (CMOS) circuits have been developed to emulate\nbiological behaviors. 12 − 14 However, due to the intrinsic differences in operation\nmechanisms of transistors and neurons, these circuits require complex\ndesigns with substantial numbers of components, especially large capacitors,\nwhich limit the optimization of their size and power. Memristive\ndevices, with the dynamic relationships between their\nconductance and the stimulus applied to them, are promising for implementing\nbiorealistic dynamics at the single-device level. 15 − 18 Various mechanisms can induce\nmemristive behaviors, including ion transport, phase change, magnetic\npolarization, and ferroelectric polarization. 19 Among these, phase change, ferroelectric switching, and magnetization\nswitching typically involve reconfigurations of atomic lattices or\nelectron spins on lattices between specific arrangements without any\nlong-range motion, limiting their capabilities to exhibit complex\nor continuous dynamics. In contrast, ion transport stands out as it\ntypically involves stochastic long-range motion, leading to complex\ndynamic processes that are intrinsically similar to those in biological\nneurons. 20 − 22 Consequently, memristive devices based on ion migration,\nreferred to as ion-based memristive devices (IMDs) in this paper,\nare superior candidates for realizing highly desirable biorealistic\ndynamics. IMDs are built using various active ions, switching materials,\nand structural configurations, all of which influence their switching\nbehaviors. Understanding the fundamental mechanisms behind these dynamic\nbehaviors and their relation to device materials and structures is\ncrucial for developing IMDs with desired properties. This review\nbegins with an analysis of the dynamics within IMDs\nin Section 2 , exploring\nthe fundamental components of the diverse dynamic behaviors and the\npotential mechanisms involved. In Section 3 , we classify IMDs based on their switching\nmaterial types and discuss how different switching materials can lead\nto distinct dynamics. Section 4 surveys the applications of IMDs in creating essential components\nof biorealistic computing systems and compares the performances of\nIMD-based artificial components with their biological counterparts,\nhighlighting their potential and challenges. Section 5 discusses various methods to tune and optimize\nthe IMDs, providing comprehensive guidance for developing more biorealistic\ndevices. Finally, in Section 6 , we address the major challenges for the general applications\nof IMDs and propose four emerging research directions involving IMDs\nthat are critical for advancing bioinspired artificial intelligence."
} | 1,515 |
28546553 | PMC5445074 | pmc | 60 | {
"abstract": "Coral reefs are threatened by climate change as coral-algal symbioses are currently living close to their upper thermal limits. The resilience of the algal partner plays a key role in determining the thermal tolerance of the coral holobiont and therefore, understanding the acclimatory limits of present day coral-algal symbioses is fundamental to forecasting corals’ responses to climate change. This study characterised the symbiont community in a highly variable and thermally extreme (Max = 37.5 °C, Min = 16.8 °C) lagoon located in the southern Persian/Arabian Gulf using next generation sequencing of ITS2 amplicons. Despite experiencing extreme temperatures, severe bleaching and many factors that would be expected to promote the presence of, or transition to clade D dominance, the symbiont communities of the lagoon remain dominated by the C3 variant, Symbiodinium thermophilum . The stability of this symbiosis across multiple genera with different means of symbiont transmission highlights the importance of Symbiodinium thermophilum for corals living at the acclimatory limits of modern day corals. Corals in this extreme environment did not undergo adaptive bleaching, suggesting they are living at the edge of their acclimatory potential and that this valuable source of thermally tolerant genotypes may be lost in the near future under climate change.",
"introduction": "Introduction Marine ecosystems worldwide are threatened by climate change 1 . Coral reefs are particularly vulnerable due to the thermal sensitivity of the symbiotic relationship between the coral host and its algal partner of the genus Symbiodinium \n 2 . Sustained temperatures of just 1–2 °C above the average annual maxima can cause coral bleaching, a breakdown in the coral alga symbiosis, and can result in coral mortality if bleaching persists 2 , 3 . Understanding coral responses to thermal stress and the impact of bleaching on the physiology and ecology of coral-algal symbiosis is fundamental to forecasting the future of reefs. The genetic identity of the coral’s symbionts influences its thermal tolerance, with a range of thermal physiologies associated with distinct symbiont types 4 – 7 . For example, within clade D, there are certain species such as Symbiodinium trenchi (D1-4, formerly D1a) that are considered thermally tolerant. Hosting a stress-tolerant symbiont type, such as the hardy opportunists belonging to Symbiodinium clade D 8 can increase the coral holobiont’s thermal tolerance by 1–2 °C compared with more thermally sensitive variants including representatives from clade C 4 . However, the enhanced resilience of clade D symbioses can come at an energetic cost, impacting the growth rate of the host 9 , 10 . Corals can overcome the challenges associated with hosting a thermally tolerant symbiont by changing their symbiont complement in response to environmental conditions 11 – 13 . Changes in a coral’s symbiont community can occur through two distinct processes: shuffling or switching. Shuffling describes the alteration of the existing symbiont complement within a coral’s tissue to increase the abundance of the physiologically most suitable symbiont 14 , 15 . In contrast, a coral may acquire a more suitable symbiont from the environment, termed switching, which has been observed in transplants 4 and acute, experimentally-applied thermal stress 16 . Early studies on changes in the symbiont community reported the emergence of clade D symbionts after severe bleaching 17 . While alterations of the symbiont community to clade D has been observed in corals on reefs in the Caribbean 12 , 13 , other reef symbiont communities have been shown to remain stable despite bleaching 5 , 18 , 19 , suggesting local biogeographic and host specific factors may be important 18 . In order to understand how coral symbioses will respond to climate change, there has been an increasing focus on the coral symbiont communities in present day extreme reefs, particularly in highly variable tidal environments 20 – 22 . These lagoons and tide pools can experience thermal maxima that are >2.5 °C greater than neighbouring environments 23 . Consequently, corals in variable environments often host symbiont communities that are generally dominated by clade D symbionts while those on more benign neighbouring reefs are dominated by clade C symbionts 20 , 21 . The association with clade D symbionts is thought to be fundamental to the greater thermal tolerance of lagoonal/tide pool corals compared to those from stable environments, although there are exceptions 24 , 25 . The predominance of clade D in these extreme environments is consistent with observations of clade D symbionts on reefs exposed to anthropogenically elevated temperatures 26 , in corals transplanted to warmer environments 4 and in corals previously exposed to thermal stress and bleaching 12 , 13 , 17 . A notable exception to the widespread occurrence of clade D types in thermally extreme reefs exists in the Persian/Arabian Gulf (PAG). Although present in some regions of the PAG 17 , 27 , 28 , clade D symbionts are largely absent in corals on the southern coastline of the Gulf 19 , 29 , 30 . This absence is significant because these corals experience the most extreme reef temperature regime globally, with summer mean monthly maxima exceeding 34 °C annually, and with the highest reported bleaching thresholds in the world 31 . Instead, corals in the southern PAG are dominated by a variant of clade C3, Symbiodinium thermophilum , a member of an ancient symbiont lineage that is cryptically distributed outside of the PAG 32 . These symbionts have been suggested to be among the most thermally robust symbionts associated with coral and form symbioses that are stable over time with no shifts to clade D 19 . To date, this symbiont has only been identified in open-water or offshore island reef environments that are characterized by relatively limited diurnal variability in temperature, in marked contrast to the clade D dominated highly variable lagoonal and tidal systems in the northern PAG and elsewhere 20 , 21 , 27 . The composition of the symbiont community in corals in lagoonal environments of the southern Gulf is unknown, but has important implications given the extreme nature of these environments. Corals in lagoonal systems in the southern PAG are exposed to considerable chronic and acute thermal stress, with long-term summer temperatures on reefs 1.5 °C greater than the adjacent open water locations, and with diurnal temperature ranges exceeding 10 °C 33 – 35 . It is unclear whether the highly variable thermal regime impacts the stability of the symbiont community, particularly during bleaching events which have altered communities elsewhere 12 , 13 , 17 . Corals in southern PAG lagoonal reefs are exposed to conditions at the acclimatory limits of modern day coral symbioses and therefore it is essential to characterise the symbionts responsible for one of the most resilient coral communities in the world. To this end, using an amplicon sequencing approach, the symbiont community composition of a southern PAG lagoonal system was investigated before and after a severe bleaching event to assess whether these reefs are clade D or S. thermophilum dominated, and to quantify the community changes in response to one of the highest ever recorded reef temperatures.",
"discussion": "Discussion Corals in the UAQ lagoon are living at the thermal limits of present day corals and therefore provide an ideal opportunity to investigate the symbiont communities on extreme reefs. This study exploits the benefits of next generation amplicon sequencing to provide the first in-depth characterisation of southern PAG lagoonal symbiont communities. It is demonstrated that the symbiont community in a highly variable lagoon in the southern PAG is largely comprised of the C3 variant S. thermophilum and that the proportions of the dominant and background symbiont types remain stable despite severe bleaching induced by one of the most extreme temperature regimes reported for coral reefs. Extreme thermal environment Corals growing in variable thermal environments have been shown to have greater tolerance to high temperatures 21 than corals from more stable environments. While the summer maxima in these reefs can reach 35 °C 21 , the lagoonal reefs in UAQ are exposed to summer temperatures up to 37.5 °C as a result of the variability around high mean temperatures (>34 °C). The maxima experienced in UAQ are extreme even in the context of the southern PAG, where C3-dominated open water reefs only reached 35.2 °C during 2014 and records of temperatures of this magnitude are largely confined to anomaly events 53 . Considering that the UAQ reef temperatures dropped to 16.8 °C during the winter 2014–2015, the corals residing on these reefs must acclimate to one of the highest thermal ranges experienced by modern day corals. Symbiont communities dominated by Symbiodinium thermophilum The capacity of corals to acclimatise to a thermal regime varying over 20 °C annually must require a thermally tolerant symbiont. Indeed, the ability for corals to survive in highly variable reefs in other regions has, in part, been attributed to symbioses with members of clade D 21 . Previous studies have shown on both local and regional scales that reefs with high thermal variability possess greater abundances of clade D symbionts than their counterparts in more stable thermal environments 20 , 21 , as do corals on reefs with high thermal maxima 17 , 26 . Clade D corals are also common to areas with high turbidity 54 . This would suggest that clade D ITS2 types should be relatively common in the UAQ lagoon, where temperature variability is high on annual (>20 °C range) and diurnal (July/August mean range ~3 °C) scales, and corals are exposed to extreme thermal stress in summer (max. 37.5 °C). In particular, the clade D species S. trenchi might be expected in the UAQ lagoon as it is a widespread, thermally tolerant opportunist that is found on other reefs in the region 12 , 55 . However, our results show that clade D zooxanthellae are a relatively rare component of the symbiont community in corals in UAQ (<1% occurrence). Here, we show that the reefs in the lagoon are characterised by low symbiont diversity (only 3 functional OTUs present), with the coral community largely dominated by clade C3-type symbionts (97% occurrence), a pattern that was consistent across an extreme bleaching event. The dominant OTU_C3 symbionts found in UAQ belong to the recent described species, S. thermophilum . This assertion is supported by the presence of the characteristic intragenomic ITS2 variant within all colonies, although variation in the proportion of this ITS2 sequence type among coral genera suggests there may be further variation within this new species. While the superior thermal tolerance of S. thermophilum has been supported by field surveys and laboratory experiments under relatively stable thermal regimes 19 , 30 , the observations from this study suggest that this symbiont also has the capacity to withstand extreme thermally variable environments. In contrast, the rarity of clade D, and notable absence of S. trenchi in UAQ under conditions where it should thrive, implies that the lagoon environment here and the wider southern Gulf is largely unsuitable for clade D dominance, as D symbiont types (including S. trenchi ) are only found in abundance in the thermally more benign northern PAG and Gulf of Oman 17 , 27 – 29 , 55 . While the extremes in temperature may cause the outperformance of S. thermophilum over clade D symbionts, the hypersaline environment may also apply a strong selective pressure to symbiont communities. This selective pressure potentially favours S. thermophilum , which appears adapted to higher salinity environments 29 . Bleaching was widespread and recorded in all genera on UAQ lagoonal reefs in September 2014. The severe bleaching could have provided an opportunity for a shift in UAQ corals’ symbiont community to clade D, as these ITS2 types have been observed to opportunistically dominate corals after bleaching events due to their high thermal tolerance 4 , 12 , 13 . Adaptive bleaching could have occurred through switching or shuffling as clade D symbionts were present on the reef prior to bleaching and were present in background levels in individuals from all coral genera, respectively. Nevertheless, the bleaching experienced on UAQ lagoon reefs did not induce changes in the composition of the symbiont communities and no increase in the relative abundance of OTU_D1 was found. In fact, post-bleaching sampling in the genera where clade D was dominant prior to bleaching revealed that Stylophora colonies had suffered complete mortality (e.g. Supplementary Figure 1 ), however, they were a relatively minor component of the community prior to bleaching. It is apparent that despite being present on the reef prior to severe and widespread bleaching, this putatively opportunistic symbiont was unable to take advantage. Nevertheless, it is important to consider that the clade D symbionts present (e.g. D1-4-6) may not possess the thermal tolerance traits of other D types such as S. trenchi (D1-4) 7 . We postulate that the absence of adaptive bleaching results from the absence of a more tolerant symbiont than the S. thermophilum dominating these symbiont communities. In agreement with Stat and coworkers 18 , our observations from multiple coral genera with different symbiont transmission modes support the notion that changing the symbiont community in favour of D types may not be essential for resilience and recovery from extreme thermal events. Wider implications The rate of temperature increase due to climate change could exceed the rate of adaptation by corals and therefore the capacity for modern day coral symbioses to acclimate is of paramount importance. The presence of reefs in an environment as extreme as the UAQ lagoon is a positive indicator for the potential of corals to cope with the increased temperatures forecast for reefs globally. The S. thermophilum dominated community observed here shows capacity for resilience to high thermal variability and extreme events, in addition to persistence under high mean summer temperatures. As symbionts can improve host thermal tolerance 4 , the possible export of these symbionts reported in the UAQ lagoonal reefs holds promise for the acquisition of southern PAG corals’ thermal tolerance by reefs elsewhere. Nevertheless, this will require further study as the fitness of S. thermophilum may be lost outside of the PAG’s unique conditions, particularly in normal oceanic salinities 29 . The coral-algal symbioses in UAQ appear to living close to their thermal limits as the extreme summer temperatures on these reefs caused widespread bleaching and did not select for an alternative symbiont type. Considering that regional sea surface temperatures have increased by 0.5 °C per decade since the 1980s 56 , further increases to temperature at this rate may soon push these corals beyond their threshold for survival in the near future. Consequently, the long - term persistence of this important source of thermally tolerant genotypes is at risk."
} | 3,835 |
34576867 | PMC8468813 | pmc | 62 | {
"abstract": "As the problem of ocean warming worsens, the environmental adaptation potential of symbiotic Symbiodiniaceae and bacteria is directly related to the future and fate of corals. This study aimed to analyse the comprehensive community dynamics and physiology of these two groups of organisms in the coral Pocillopora sp. through indoor simulations of heat stress (which involved manually adjusting the temperature between both 26 °C and 34 °C). Heat treatment (≥30 °C) significantly reduced the abundance of Symbiodiniaceae and bacteria by more than 70%. After the temperature was returned to 26 °C for one month, the Symbiodiniaceae density was still low, while the absolute number of bacteria quickly recovered to 55% of that of the control. At this time point, the F v / F m value rose to 91% of the pretemperature value. The content of chlorophyll b associated with Cyanobacteria increased by 50% compared with that under the control conditions. Moreover, analysis of the Symbiodiniaceae subclade composition suggested that the relative abundance of C1c.C45, C1, and C1ca increased during heat treatment, indicating that they might constitute heat-resistant subgroups. We suggest that the increase in the absolute number of bacteria during the recovery period could be an important indicator of coral holobiont recovery after heat stress. This study provides insight into the cross-linked regulation of key symbiotic microbes in the coral Pocillopora sp. during high-temperature stress and recovery and provides a scientific basis for exploring the mechanism underlying coral adaptation to global warming.",
"conclusion": "5. Conclusions In this study, an indoor simulation of heat stress was conducted to elucidate the response characteristics and coordinating roles of key symbiotic members (Symbiodiniaceae and bacteria) in the environmentally sensitive coral Pocillopora sp. The heat stress would quickly disrupt the symbiotic relationship between coral host and microorganism (Symbiodiniaceae and bacteria). The numbers of both Symbiodiniaceae and bacterial populations decreased sharply. However, the surviving heat-tolerant Symbiodiniaceae and Cyanobacteria protected the basic metabolic needs of the coral holobiont. When the thermal stress was removed, bacteria recovered faster than Symbiodiniaceae. It meant that the active bacteria might play a major role in the establishment of a restored status of the coral holobiont, assisting the coral host in rapid recovery. Overall, we could describe a cross-linked model of two key symbiotic members (Symbiodiniaceae and bacteria) in coral Pocillopora sp. under heat stress and recovery. It is hypothesised that the coral response to global warming may involve a combination of a large number of bacteria with heat-resistant Symbiodiniaceae subclades.",
"introduction": "1. Introduction With global climate warming, rising sea surface temperature (SST) is a major threat to the survival and development of coral reefs [ 1 , 2 , 3 , 4 ]. Many large-scale coral bleaching events are related to the abnormal rise in SST [ 5 , 6 , 7 ], with the frequency and severity of coral bleaching tending to increase [ 8 ]. The most direct impact of high SST is the large-scale expulsion of endosymbiotic dinoflagellates (family Symbiodiniaceae) by coral, resulting in a loss of pigmentation and a disruption in the coral energy supply [ 9 , 10 ] and ultimately leading to coral bleaching or mortality. However, some species of coral hosts are adaptable and resistant to environmental stress [ 11 , 12 ], primarily because their symbiotic microorganisms are somewhat plastic and can positively respond to environmental stress [ 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 ]. Therefore, it is necessary to gain deep insight into the physiological dynamics of Symbiodiniaceae and bacteria (key members of coral–microbe assemblages), especially with respect to their coordinating role, in the coral host response to abnormal environmental stress. According to current research, a wide variety of Symbiodiniaceae provide more than 95% of the energy to coral holobionts through photosynthesis [ 23 , 24 ]. When corals are stressed by heat, the density and community structure of the symbiotic Symbiodiniaceae can change substantially [ 25 ]. For example, the absolute number of symbiotic Symbiodiniaceae changes in response to seasonal variations in SST [ 26 ]. Xu et al. evaluated the correlations between the Symbiodiniaceae density, the effective photochemical efficiency, and the seasonal variation in SST in five dominant reef coral species from the northern South China Sea (Luhuitou fringing reef). The thermal tolerance of coral is primarily and positively dependent on its Symbiodiniaceae density [ 26 ]. In addition, when corals were subjected to high-temperature stress, the Symbiodiniaceae density decreased greatly. Kemp et al. investigated the community dynamics and physiology of symbiotic Symbiodiniaceae in the coral Orbicella faveolata before, during, and after a natural thermal bleaching event off the coast of Puerto Morelos, Mexico [ 25 ]. The researchers found that the Symbiodiniaceae density and maximum photochemical quantum yield ( F v / F m ) decreased sharply during a thermally induced coral bleaching event. In addition, the community structure of Symbiodiniaceae responded positively, reflected by the increase in the relative abundance of members of heat-tolerant subclades (A3 and D1a, with higher photosynthetic efficiency). Tong et al. analysed eight potential environmental factors (temperature, salinity, NH 4 + , NO 3 − , PO 4 3− , NO 2 − , dissolved oxygen, and depth) affecting the community structure of symbiotic Symbiodiniaceae in the Galaxea fascicularis and Montipora spp. corals from three biogeographic regions at different latitudes and temperatures in the South China Sea [ 27 ]. The survey results suggested that temperature played a major role in shaping the Symbiodiniaceae community structure and coral–alga symbiosis. Thus, the Symbiodiniaceae density and community structure directly influence the adaptability and tolerance of coral hosts to changing environments. In addition, coral-associated bacteria are among the most abundant microbial organisms within coral holobionts and play an important role in energy supply, material circulation, disease occurrence, and physiological adaptability [ 15 , 28 , 29 , 30 ]. On the basis of operational taxonomic units (OTUs), next-generation sequencing (NGS) analysis has revealed more than a thousand species of coral-associated bacteria [ 31 , 32 ]. Various coral-associated bacteria respond positively to changes in the external environment. Ritchie et al. showed that when the temperature increased, the dominant community in coral ( Acropora spp.) mucus shifted from comprising antibiotic-producing bacteria to comprising conditionally pathogenic bacteria [ 33 ], which explained the decrease in resistance and increase in the case rate of coral hosts. Studies on associated microbial communities in the hard coral Acropora millepora tagged on a reef flat of Magnetic Island (Great Barrier Reef, Australia) showed that the dominant species changed to Vibrio spp. after a bleaching event induced by high temperature [ 34 ]. This may have been caused by the bacterial community responding to the lack of symbiotic Symbiodiniaceae and the rise in SST. In addition, many studies have shown that coral-associated bacteria play an important role in the response of coral hosts to abnormal environmental conditions [ 18 , 25 , 35 , 36 , 37 , 38 , 39 ]. In the above studies, the response characteristics of Symbiodiniaceae and associated bacteria in coral to abnormal environmental conditions were comprehensively characterised. However, more efforts are still needed to better understand complex coral holobionts because they represent mutually beneficial and dynamically balanced assemblages that include actinozoans, Symbiodiniaceae, and bacteria. The complete regulatory process and response characteristics of key members to abnormal environmental stress are still unclear, especially in terms of absolute quantity and community composition. In the present study, we aimed to elucidate the response characteristics and coordinating roles of key symbiotic members (Symbiodiniaceae and bacteria) in the environmentally sensitive coral Pocillopora sp. to heat stress and early recovery. A laboratory-based simulation experiment was designed to apply different temperature stresses ( Figure 1 ). Various comprehensive physiological parameters (community composition, photochemical efficiency, photosynthetic pigment content, and absolute cell numbers) related to Symbiodiniaceae and bacteria were analysed in detail under various temperature gradients. This study provides insights into the dynamic regulatory process of key coral symbionts (Symbiodiniaceae and bacteria) to high-temperature stress and new considerations for the adaptation of coral hosts to global warming.",
"discussion": "4. Discussion The community dynamics and physiology of symbiotic Symbiodiniaceae and bacteria have been extensively studied in terms of the response of coral to heat stress. The members of unique Symbiodiniaceae subclades (e.g., D1a and A3) are known to be related to the thermal tolerance of coral [ 25 ]. However, the relative abundance of the members of the three dominant subclades, i.e., C1c.C45, C1, and C1ca, in the coral Pocillopora sp. increased significantly under high-temperature stress in this study. This phenomenon may be directly related to the maintenance of the energy supply when coral is in crisis, but further experimental verification is needed. Symbiotic Symbiodiniaceae played a positive regulatory role when Pocillopora sp. was exposed to high-temperature stress. In addition, some pathogenic bacteria thrive in high-temperature environments and become highly abundant, causing pathological changes in and bleaching of their coral hosts [ 33 ]. In fact, symbiotic Symbiodiniaceae and bacteria in coral have been separately investigated in many studies [ 20 , 21 , 25 , 26 , 27 , 28 , 31 ]. We know that coral is a typical mutually beneficial symbiotic organism (the coral holobiont) that can be actively and dynamically regulated in response to environmental change. Therefore, a comprehensive understanding of the microecological regulation of the coral holobiont (including Symbiodiniaceae and bacteria) in response to temperature is helpful for a more in-depth understanding of the mechanism and evolution of coral responses to global warming. In this study, the community dynamics and physiology of Symbiodiniaceae and bacteria (including the F v / F m , Symbiodiniaceae density and subclade composition, bacterial community composition, and absolute number, and photosynthetic pigment content) during the coral response to heat stress were systematically analysed. We have generated several novel conclusions, which are discussed below. 4.1. Changes in Photosynthetic Pigments and Fv/Fm Were Inconsistent with Changes in Symbiodiniaceae Density and Subclade Composition After the planula stage, coral begins to settle on the calcareous remains of their ancestors. Most of its energy originates from photosynthesis because coral is a typical mutually beneficial holobiont that has a large number of partners (e.g., Symbiodiniaceae and photosynthetic bacteria) that provide pigments (e.g., chlorophylls and carotenoids) and carry out photosynthesis [ 23 ]. When coral faces extreme environmental stress, its energy source and maintenance can directly influence its survival. F v / F m can reflect the conversion efficiency of light energy in the photosystem II (PSII) reaction centre and is a reflection of the potential maximum photosynthetic capacity of plant and coral holobionts [ 26 , 52 , 53 , 54 ]. We simulated the stress of acute warming on the coral Pocillopora sp. to assess the influence of heat on the physiology of the coral holobiont. The content of photosynthetic pigment (chlorophyll a) and F v / F m in the coral holobiont exhibited similar change trends during the whole response to high-temperature stress and recovery. However, only the chlorophyll a and c contents increased significantly with the initial increase in temperature (at 32 °C). This phenomenon was consistent with the results reported by Nunez-Pons et al., suggesting that these contents increased presumably to meet increased metabolic demands [ 17 ]. During the recovery period, the chlorophyll (b and c) contents were also abnormally high, which may be related to the high energy demand and low member density of coral symbionts (Symbiodiniaceae and Cyanobacteria) ( Figure 4 ). However, the content of chlorophyll b related to Cyanobacteria did not change significantly throughout the whole heat-stress process but increased significantly during the recovery period ( Figure 4 B). The reason behind the change of photosynthetic pigments in coral holobiont during high-temperature stress and recovery needs to be further explored. Our results are partially inconsistent with those of a previous study [ 17 ]. In a gradual thermal stress (GTS) experiment with Exaiptasia anemones , GTS increased chlorophyll contents and decreased Symbiodiniaceae proliferation. However, the chlorophyll contents decreased in the recovery period after GTS, while the rate of symbiont division increased [ 17 ]. Moreover, in Exaiptasia experiencing thermal stress and bleaching at temperatures greater than 30 °C, the remaining photosynthates in hospite symbionts continued to be translocated but at a significant cost to the organisms [ 55 ]. Photosynthesis by Symbiodiniaceae is the main energy source of coral holobionts [ 23 , 24 ]. Although the Symbiodiniaceae density was very low in this study (e.g., at 32 °C and during recovery at 26 °C), the relative values of F v / F m were high—94 and 91%, respectively. Basing on the changes of chlorophyll b, it is speculated that Cyanobacteria might play a key role in the photosynthetic energy supply in corals in response to high-temperature stress and recovery. Especially in the recovery period, the density of the main contributors (Symbiodiniaceae) to photosynthetic energy was low, while the absolute number of bacteria rebounded significantly ( Figure 6 ). These findings further indicated that bacteria may play important roles in establishing a new state during coral recovery. As we continued to analyse the changes in Symbiodiniaceae composition, we found that the original subclade composition of Symbiodiniaceae in Pocillopora sp. included 11 types: C42 and C1c.C45, C1, and C1ca were the dominant members, all of which belonged to Symbiodiniaceae clade C (equivalent to genera) ( Figure 5 B). With increasing temperature, the relative abundance of the members of the most dominant subclade, C42, decreased significantly, while that of the other dominant subclades, C1c.C45, C1, and C1ca, increased significantly. Selective pressures in environments whose temperature widely fluctuates might have adaptive value to symbiosis specificity [ 56 ]. C42 was a temperature-sensitive subgroup, and C1c.C45, C1, and C1ca were temperature-resistant subgroups. Thus, the F v / F m should be maintained mainly by the members of the temperature-tolerant subclades C1c.C45, C1, and C1ca. However, there was no sign of recovery of Symbiodiniaceae density when Pocillopora sp. was subjected to extreme high-temperature stress and then allowed to recover at 26 °C for one month. One of the most prominent phenomena was that the chlorophyll c content related to Symbiodiniaceae increased significantly to 225% of its original value ( Figure 4 C). Therefore, photosynthetic efficiency was related to not only Cyanobacteria survival but also Symbiodiniaceae survival. This could be explained by the results of the analysis of the structural composition and the absolute number of bacteria associated with Pocillopora sp., as described below. 4.2. Potential Regulatory Role of Coral-Associated Bacteria under Heat Stress Coral-associated bacteria are very diverse and have important biological functions (material cycling, disease prevention, etc.). Previous studies have concentrated on the community dynamics of coral-associated bacteria [ 15 , 25 , 28 , 31 , 34 ], and the associated physiological functions have been focused on less [ 57 , 58 ]. In the present study, fluorescence quantitative PCR (absolute quantification) was used to analyse the absolute number of bacteria associated with Pocillopora sp. throughout the heat-stress cycle. During heating, the absolute number of bacteria decreased sharply, similar to the Symbiodiniaceae density, which was highly sensitive to high temperatures ( Figure 6 ). However, the bacterial response was more intense in the initial heating stage (at 30 °C) than in the other stage, according to the change in numerical value. The only difference was that the absolute number of bacteria substantially recovered when the temperature returned to 26 °C for one month. This microcosm change may play an essential role in the establishment of a new state of the coral holobiont during/after extreme stress. In the face of climate variability, corals are considered particularly susceptible, and the mechanisms that contribute to their recovery must be understood [ 59 ]. Ziegler et al. pointed out that symbiotic microbial adaptation constitutes another possible mechanism to assist sensitive organisms in resisting environmental changes beyond the host’s own physiological acclimatisation and assisting the migration of heat-tolerant alleles [ 21 ]. After their numbers decreased in response to extremely high temperatures, Symbiodiniaceae members were restored in terms of composition but were still present at a very low abundance. We, therefore, speculate that Symbiodiniaceae meet the metabolic needs of coral holobionts by increasing the content of photosynthetic pigments in individual cells. In addition, the rapid recovery of coral-associated bacteria, especially photosynthetic bacteria (Cyanobacteria), may play a key role in the energy supply of the coral holobiont. 4.3. Model of the Coordinated Response of the Coral Holobiont during Heat Stress Coral holobionts are complex and dynamic multifunctional organisms. To date, few studies have fully explored the changes in the community dynamics and physiology of symbiotic coral microbial organisms (especially the absolute quantity of bacteria) in response to environmental changes. On the basis of our laboratory simulation of high-temperature stress, a thermal response model of Pocillopora sp. was proposed from the surface morphology, photosynthetic pigments, symbiont composition, and Symbiodiniaceae and bacterial abundance ( Figure 7 ). Under normal conditions, coral holobionts require energy supplied by and material circulation of symbiotic microorganisms. With respect to this point, the community structure and individual populations of symbiotic Symbiodiniaceae and bacteria in the coral host were stable, both types of organisms were highly abundant, and the coral could obtain enough energy to grow healthily ( Figure 7 A). However, under heat stress, the numbers of both Symbiodiniaceae and bacterial populations decreased sharply. At this time, the balance of the coral holobiont was disrupted, and by increasing the content of photosynthetic pigments per unit cell, the surviving Symbiodiniaceae and Cyanobacteria protected the basic metabolic needs of the coral holobiont. The energy supply was essentially depleted, and the coral host was at risk of mortality ( Figure 7 B). When the thermal stress was removed, the active bacteria that reproduce faster may play a major role in the establishment of a new state of the coral holobiont, assisting the coral host in rapid recovery ( Figure 7 C). Previous studies have confirmed that the members of some Symbiodiniaceae subclades (e.g., D1a and A3) are related to the heat tolerance of coral hosts [ 25 ]. This differentiation between heat-tolerant strains of Symbiodiniaceae and bacteria, including species diversity, the rate of change, and the rate of propagation, should constitute the basis for coral adaptation to environmental changes. Therefore, we speculate that coral response strategies may involve a combination of a large number of bacteria (especially highly abundant photosynthetic bacteria) with heat-tolerant Symbiodiniaceae subclade members under future warmer climates. Therefore, our study provides novel insights into the microecological regulation of key coral symbionts during heat stress and during a recovery period."
} | 5,169 |
35050140 | PMC8780272 | pmc | 63 | {
"abstract": "Global climate change has resulted in large-scale coral reef decline worldwide, for which the ocean warming has paid more attention. Coral is a typical mutually beneficial symbiotic organism with diverse symbiotic microorganisms, which maintain the stability of physiological functions. This study compared the responses of symbiotic microorganisms and host metabolism in a common coral species, Pavona minuta , under indoor simulated thermal and cold temperatures. The results showed that abnormal temperature stresses had unfavorable impact on the phenotypes of corals, resulting in bleaching and color change. The compositions of symbiotic bacteria and dinoflagellate communities only presented tiny changes under temperature stresses. However, some rare symbiotic members have been showed to be significantly influenced by water temperatures. Finally, by using ultra-performance liquid chromatography tandem mass spectrometry (UPLC–MS) method, we found that different temperature stresses had very different impacts on the metabolism of coral holobiont. The thermal and cold stresses induced the decrease of anti-oxidation metabolites, several monogalactosyldiacylglycerols (MGDGs), and the increase of lipotoxic metabolite, 10-oxo-nonadecanoic acid, in the coral holobiont, respectively. Our study indicated the response patterns of symbiotic microorganisms and host metabolism in coral to the thermal and cold stresses, providing theoretical data for the adaptation and evolution of coral to a different climate in the future.",
"conclusion": "5. Conclusions Our work exhibited the short-term response patterns of coral–microbe assemblages to abnormal temperatures from various perspectives. The diversity and composition of symbiotic bacterial communities was not found to be significantly affected by the temperature stresses. Meanwhile, the dominant dinoflagellates were also not significantly changed. However, some rare symbiotic bacterial genera and dinoflagellate sub-clade were found to be significantly influenced by the temperature stresses. Among them, the potential probiotic members were significantly depleted in the coral holobiont under both thermal and cold stresses. Moreover, many metabolites were observed with the significantly different abundances between the corals at normal condition and temperature stresses. The thermal and cold stresses induced the decrease of anti-oxidation and the increase of lipotoxic metabolites in the coral holobiont, respectively. This work provided insight into the responses of symbiotic microorganisms and physiological metabolisms involved in coral P. minuta under temperature stresses.",
"introduction": "1. Introduction The coral reefs play a vital role in maintaining the biodiversity and ecosystem functions of marine, and global climate change threatens them all over the world [ 1 ]. Coral is a temperature sensitive marine organism, which bleaches and dyes under abnormal sea temperatures [ 2 ]. As the main consequence of climate change, ocean warming is seriously threatening the survival of coral reefs [ 3 ]. Many studies have been conducted to explore the response patterns of corals to the elevated temperature by diverse methods [ 4 , 5 , 6 ]. For example, divergent responses to heat stress by different coral taxa has been found in the Great Barrier Reef, resulting in the regional-scale shift in the composition of coral assemblages [ 7 ]. Another study also conducted on the Great Barrier Reef revealed that the populations of corals successfully adapted the warming in the future but with increasing sensitivity to random thermal fluctuations [ 8 ]. Compared with the well-known warming, climate change will also lead to extreme weather events, such as abnormal cold in winter in tropical or subtropical regions. The abnormal temperature drop also can affect the growth and stability of coral reefs. However, there is a lack of research on the response patterns of coral reef to the relatively cold seawater temperature. A variety of microorganism mutuality symbiosis with the corals and their interaction is indispensable in the normal physiological function of coral reefs [ 9 , 10 ]. Particularly, coral bleaching of coral reefs under the abnormal temperatures is the result of dysfunction of the symbiotic relationship and expulsion of the symbionts from coral host [ 11 ]. Among these symbionts, bacteria and dinoflagellates are more worthy of attention. Corals provided ideal habitats for bacteria, and numerous novel bacteria have been detected in them to complete substantial functions linked with nutrient cycling, the degradation of pollutants, and host health [ 12 , 13 , 14 ]. The dinoflagellates, mainly Symbiodiniaceae, is an obligate symbiotic family residing within the tissues of corals, which has been demonstrated in a particular relationship with coral resilience to ambient temperatures [ 15 ]. There is a two-way exchange of metabolites between these symbionts and the coral hosts to maintain a healthy state of mutualism [ 16 ] Changes in water temperature conditions may shift the symbiotic microbial communities of corals and destroy of the homeostasis of holobiont metabolism. Therefore, understanding the response patterns of symbiotic microorganisms and host metabolism in corals under temperature stresses is necessary to improve the formulation of coral reef conservation. Recent advances in the omics-based technologies have provided more comprehensive methods to obtain the symbiotic microbial communities and metabolic changes in corals. High-throughput sequencing based on target biomarkers, such as 16S rRNA and ITS genes, has been frequently used to assess the diversity and composition of bacteria and dinoflagellates among diverse ecosystems, including the coral holobiont [ 17 , 18 ]. Moreover, metabolomics based on mass spectrometry have applied to explore the variations of metabolic functions in corals with different physiological states [ 19 ]. In this study, a multi-omics method combined with the high-throughput sequencing based on 16S rRNA and ITS genes and the un-targeted mass spectrometry-based metabolomics was established. Pavona minuta , a typical massive scleractinian coral from Weizhou Island, Beibu Gulf of China, was selected as the research object. The response patterns of the symbiotic microorganisms and host metabolism as well as the relationships among them under extremely high and low seawater temperature stresses were evaluated. The findings of this study will assist future research into the mechanisms of coral devastation under different environments and climate change.",
"discussion": "3. Discussion Coral is a typical mutually beneficial symbiotic organism (together termed the coral holobiont), and the compositions coordinate with one another when coping with environmental stress. Therefore, a synthesized understanding of the multiple responses of the coral holobiont (including bacteria and dinoflagellate) to temperature is favorable for a deeper understanding of the ecological and evolutionary mechanisms of coral responses to climate change. Many previous studies have reported the responses of coral-microbe assemblages to extreme thermal stress, and showed inconsistent results of resistance or susceptibility of coral communities to climate change [ 20 , 21 , 22 ]. The data present herein illustrated the response patterns of symbiotic microorganisms and host metabolism in coral P. minuta exposed to different temperature stresses. The fitness of coral holobiont has been proven to have an essential relationship with coral-associated bacterial communities [ 23 ]. Bourne et al. [ 24 ] demonstrated conserved bacterial banding profiles during a bleaching event by denaturing gradient gel electrophoresis analysis. The development of next-generation sequencing has been facilitating the research on diversity and community structure of microorganisms, and more compete symbiotic bacterial communities of corals have been detected by recent studies [ 14 , 18 ]. Proteobacteria, Bacteroidetes, Cyanobacteria, Chloroflexi and Firmicutes were found to be the dominant bacteria phyla in this study ( Figure 2 d), and this result was consistent with other studies [ 14 , 25 ]. For dinoflagellate, it is an obligate symbiotic member residing within the tissues of coral reefs, and most coral–dinoflagellate associations have host specificity [ 26 ]. As the advances of next-generation sequencing, the IST2 rDNA genotyping based sequencing were more frequently used to assess dinoflagellate diversity to obtain more comprehensive results [ 17 , 27 ]. The dinoflagellate community primarily consisted of Clade C (Cladocopium), in which C1 was the dominant symbiont ( Figure 3 ). Other research about corals in marine areas, such as Okinawa of Japan and Jeju Island of Korea, were also found sub-clade of C1 as the majority member of the dinoflagellate–coral associates [ 28 , 29 ]. In contrast to the dominant symbiotic microorganisms, some rare members in symbiotic communities observed significant changes under different temperature stresses ( Figure 2 e,f, and Figure 3 ). In marine ecosystems, rare members of bacterial com-munities have been indicated to play a more active role than the dominant members [ 30 ]. Besides, rare members of dinoflagellate have long been recognized to contribute to the resilience of coral–algal associations [ 22 ]. Members of dinoflagellate subclades (e.g., D1a and A3) with relatively lower abundances in corals were known to be related to the thermal tolerance [ 31 ]. Several rare bacterial genera, including Muricauda , Loktanella , Acrophorminum , Pseudophaeobacter , and Hautella, were found to be significantly respond to the temperature stresses ( Figure 2 d,e). Among them, Muricuada had a relative abundance below to 3% and other genera were all below to 0.3%. The genus of Muricuada have been reported to possess the capacity of zeaxanthin biosynthesis [ 32 ], which is a natural pigment with critical role in the prevention of age-related macular degeneration in human beings [ 33 ]. The significant enrichment of Muricuada in corals from HT group indicated that this bacteria genus could respond to the tolerance of high temperature ( Figure 2 d). In addition, the bacteria genera that depleted in the HT or LT groups, such as Loktanella and Pseudophaeobacter , were showed to have the algicidal activity on the toxic dinoflagellate [ 34 ] or probiotic effect against to pathogens [ 35 ]. These results suggested the temperature stress on corals could cause the enrichment of harmful symbionts due to the weakness of competition. Moreover, we observed that Clade A16 of dinoflagellate experienced significant fluctuation in the results of this study, although its proportion in the dinoflagellate community of P. minuta was low ( Figure 3 b). According to the information mentioned above, we inferred that the sub-clade A16 of dinoflagellate could be inferred to have toxic on coral reefs, but the related evidence has not been reported in any previous studies. Temperature stress can result in the imbalance of metabolism in the coral holobiont. Hillyer et al. [ 36 ] studied the metabolite profiles of symbiont and host during heat-stress and bleaching in a model cnidarian-dinoflagellate symbiosis, and detected elevated pools of polyunsaturated fatty acids (PUFAs) in the symbiont, but reductions of PUFAs in the host. Moreover, another study reported that elevated ambient temperatures could induce the alteration of carbohydrate composition, cell structural lipids, and signaling com-pounds in the reef-building coral [ 37 ]. In this study, a significant increase in the contents of some PUFAs, such as PA (14:0/19:1(9Z)), 16,17-epoxy-DHA, PC (0-12:0/0-2:0), and cannabidiolic acid, were also found in corals under thermal stress ( Figure 5 a). Moreover, the most decreased metabolites in corals under thermal stress were serval members of MGDG ( Figure 5 a). MGDG has been shown to have anti-oxidation ability, which can clear excess active oxygen free radicals in the hosts [ 38 ]. For cold stress, 10-oxo-nonadecanoic acid was the metabolites with the highest abundant increase (~20-fold increase, below 5-fold for other DAMs, Figure 5 b), which is an inducer of lipotoxic effects and result in apoptosis [ 39 ]. These findings indicated the harmful effects of temperature stresses on the coral reefs through the changed of metabolic profiles. In our study, almost all dominant bacteria and dinoflagellates seemed to have tiny variations under the temperature stress ( Figure 2 d and Figure 3 ). There is a theory called “limited symbiont shuffling”, which indicates that the composition of symbionts may change over time with their hosts simultaneously, but the change proportion of most of symbiotic members are relatively low and the occurrence of symbiont shuffling may be rare [ 26 ]. In this study, our results emphasized that the relative abundances of different clades kept relatively stable, which is consistent with limited symbiont shuffling theory [ 7 , 40 ]. The symbiotic microbial community studied in here presented a rich diversity ( Figure 2 and Figure 3 ). This high diversity might hamper the predictions about their responses to simulated temperature stresses [ 41 ]. A useful method is to classify microbial communities into different functional groups instead of detecting the complicated responses of hundreds of microbial taxa. Furthermore, due to the heterogeneity of the sensitivity of different lineages, it would be intended to focus on these key species or their relatives in this kind of research. Future research may take a phylogenetic, trait-based framework involved in predicting coral-associated microbial responses to climate change. More indicators, such as trait diversity, should be introduced to the study of coral holobiont."
} | 3,463 |
40212700 | PMC11935265 | pmc | 66 | {
"abstract": "Separate memory and processing units are utilized in conventional von Neumann computational architectures. However, regarding the energy and the time, it is costly to shuffle data between the memory and the processing entity, and for data‐intensive applications associated with artificial intelligence, the demand is ever increasing. A paradigm shift in traditional architectures is required, and in‐memory computing is one of the non‐von‐Neumann computing strategies. By harnessing physical signatures of the memory, computing workloads are administered in the same memory element. For in‐memory computing, a wide range of memristive material (MM) systems have been examined. Moreover, developing computing schemes that perform in the same sensory network and that minimize the data shuffle between the processing unit and the sensing element is a requirement, to process large volumes of data efficiently and decrease the energy consumption. In this review, an overview of the switching character and system signature harnessed in three archetypal MM systems is rendered, along with an integrated application survey for developing in‐sensor and in‐memory computing, viz., brain‐inspired or analogue computing, physical unclonable functions, and random number generators. The recent progress in theoretical studies that reveal the structural origin of the fast‐switching ability of the MM system is further summarized.",
"conclusion": "7 Conclusion A market for effective and high‐performance training and inference systems, both on the edge and in the cloud, has been generated by the rapid growth of AI, e.g., deep neural networks. A vital role in defining the future of computing is represented by mobile devices, which are impeded by energy limitations. The leveling of the cost per transistor as the transistor size becomes smaller is another reason. Various device manufacturers could be motivated to maintain previous technology nodes or feature sizes but enhance the device performance with high‐efficacy computing elements including a computational memory. The integration of the MM system with a variety of front‐end CMOS technologies is facilitated, since most MM systems are compatible with the BEOL integration. In‐memory computing based on the MM system is expected to have a substantial impact on enhancing the area and energy efficiency, along with the latency. This could pave a way for the next generation of non‐von Neumann computing in a favorable market setting.",
"introduction": "1 Introduction Modern computation systems are constructed on the foundation of a von Neumann computing architecture in which data are transferred to a processing unit. [ \n \n 1 \n , \n 2 \n , \n 3 \n \n ] Substantial costs in the energy and the latency, i.e., the delay between an instruction to transfer data and the same data being transferred, are incurred when a large amount of data are shuttled between the memory and the processing unit for performing different computing workloads. For many applications, e.g., vital artificial intelligence (AI)‐type tasks, the latency related to retrieving data from memory units is a major performance roadblock (researchers have called it a bottleneck). Between the time utilized for accessing data in the memory and the processing element, there is an increased difference, and this phenomenon is described as a memory wall. Since computational systems are energy constrained owing to the increase in the number of edge‐computing devices and cooling limitations, the energy cost of moving data is also a substantial challenge. The latency cost of multiplying two numbers in processing units is smaller compared to that of retrieving the numbers from the memory for traditional complementary metal–oxide–semiconductor (CMOS) technologies. [ \n \n 4 \n , \n 5 \n , \n 6 \n \n ] The difficulty of avoiding data movement exists for conventional strategies, including the utility of application‐specific processors that are customized for targeted applications or many processors connected in parallel, viz., graphic processing units. Thus, new computing architectures in which the memory and the processing entity are more co‐located are required. Inserting a monolithic‐compute unit nearer to monolithic‐memory units physically is one of the concepts proposed. [ \n \n 7 \n , \n 8 \n , \n 9 \n \n ] The recent enhancements in hardware stacking technologies and commercialization of advanced memory types including the high‐bandwidth memory and hybrid memory cubes have benefitted the idea, which is termed as near‐memory computing, substantially. [ \n \n 10 \n , \n 11 \n , \n 12 \n \n ] The 3D monolithic integration was utilized further for attaining a smaller hardware size and a denser connectivity between the memory and processing units. [ \n \n 13 \n , \n 14 \n , \n 15 \n \n ] However, a separation between the memory and the compute unit still exists physically for conventional schemes that target to minimize the distance and the time for memory retrieval. A different strategy wherein computing workloads are implemented in the same memory entity is described as in‐memory computing ( Figure \n \n 1 a,b ). [ \n \n 16 \n , \n 17 \n , \n 18 \n \n ] This is attained by harnessing the physical character of material systems, device designs, array‐type configurations, peripheral circuitries, and memory controllers. [ \n \n 5 \n , \n 19 \n , \n 20 \n \n ] In‐memory computing describes a computing workload that is achieved in a memory unit. Moreover, in‐memory computing is a promising candidate for enhancing the time complexity, viz., the amount of computational time utilized to run an algorithm, of computing workloads. [ \n \n 21 \n , \n 22 \n , \n 23 \n \n ] A high degree of parallelism enabled by a large ensemble of memory units that administer a computing task results in the improved time complexity. Furthermore, the time complexity decreases with an increase in the degree of connectivity between memory units. Thus, a substantial improvement in the computing efficiency is achieved when the boundary between the memory and the processing module becomes negligible, which mimics an energy efficient human brain where the processing and memory elements are interlinked. [ \n \n 24 \n , \n 25 \n , \n 26 \n \n ] \n Figure 1 In‐memory computing and in‐sensor computation using MM systems. a) In a traditional computational platform, the data D are transferred to a processing element, when a function f is implemented on the D , resulting in substantial costs in power and time. b) With reference to the in‐memory computation, by harnessing physical characters of the memory hardware, the f ( D ) is administered in the same computational storage entity, therefore avoiding the requirement to transfer the D to the processing element. The MM technologies including phase‐change memory (PCM), resistive switching memory (RSM), and magnetic tunnelling memory (MTM) can operate as units of the computational‐storage element. c) Traditional sensory computational design. The analog outputs from different sensors are altered to digital signals, which are retained in the storage unit. The processing elements access the data from the storage and subsequently transfer the output signal back to the storage element for long‐term retention. d) The in‐sensor computational design. The processing operations are included in distinct sensors for front‐end processing. To avoid data transfer between sensors and processors, the sensor can cooperate to implement information aggregation and compression, and data processing. © 2024 WILEY‐VCH GmbH Memristive materials (we term them MMs) are promising candidates for achieving in‐memory computing. [ \n \n 27 \n , \n 28 \n , \n 29 \n , \n 30 \n , \n 31 \n \n ] Upon the application of an external electrical stimulus, the MM system discloses programmable conductance states. [ \n \n 32 \n , \n 33 \n , \n 34 \n \n ] The prototypical MM operations, enabled by the phase‐change, i.e., thermally induced crystalline–amorphous transitions, tunnel magnetoresistance, viz., spin‐dependent tunnel conductance, and electrochemical reaction, e.g., redox and ion migration, are based on the switching of a dielectric layer in a two‐terminal metal–dielectric–metal configuration. [ \n \n 35 \n , \n 36 \n , \n 37 \n \n ] The MM systems, which have a small footprint, low programming energy, high reliability, and short switching time, exhibit a marked contrast in the electrical conductance as a result of the dependence of conductance states on the history of electrical stimuli. [ \n \n 38 \n , \n 39 \n , \n 40 \n \n ] Furthermore, through the utilization of physical processes to perform complex signal alterations, the data can be processed in the MM system inherently for enhancing both the energy efficiency and the area efficacy of many applications, such as hardware security, and neuromorphic and analog computing. [ \n \n 41 \n , \n 42 \n , \n 43 \n \n ] \n The analog sensory data are altered to digital signals through an analog‐to‐digital conversion, and subsequently stored in the memory, in traditional computing architectures (Figure 1c,d ). Finally, the data are moved from the memory unit to processing units. However, a low‐energy efficiency, as well as long latency, results from the traditional data conversion and transmission strategy. Nevertheless, various connected sensors or single self‐adaptive sensor types process the sensory data directly, which integrate computing operations and sensing functions, in an in‐sensor computation architecture. [ \n \n 44 \n , \n 45 \n , \n 46 \n , \n 47 \n \n ] In this review, we provide an overview of the switching signature and system character utilized in three typical MM systems, along with a combined application survey for developing in‐memory computing, as well as in‐sensor computation ( Figure \n \n 2 \n ). We also disclose an outlook on the challenge and opportunities. Furthermore, we summarize theoretical studies that reveal the physical origin of the rapid‐switching ability of the MM system. Figure 2 Memristive material development. Various types of memrisitive material systems, i.e., phase‐change memory layers, resistive switching memory layers, and magnetic tunneling memory layers, are disclosed. This survey presents four key modes of research domains that utilizes the memristive material system. © 2024 WILEY‐VCH GmbH"
} | 2,570 |
40103995 | PMC11915212 | pmc | 67 | {
"abstract": "Self-powered sensors are increasingly valued for their eco-friendly and energy-efficient design, making them ideal for sustainable applications. As global energy demand rises and carbon emissions increase, there is a shift toward renewable energy sources like solar and wind. Advanced sustainable energy devices, such as piezoelectric and triboelectric nanogenerators, show promises for capturing untapped energy, supporting the development of portable, green devices. While commercialization of triboelectric materials is limited, they hold strong potential for large-scale energy harvesting. This study investigates how tailored surface topography can enhance the electrical output of a hybrid nanogenerator. We developed a hybrid piezoelectric and triboelectric nanogenerator (HBNG) using a BaTiO 3 -PDMS composite (containing 10–20 vol% barium titanate in polydimethylsiloxane). Micron-sized pyramid structures of 20% BT/PDMS were created on the film through optical lithography, while scanning electron microscopy and X-ray diffraction were used to assess the composite's crystal structure and phase characteristics. Altering the film's surface morphology led to substantial improvements in electrical performance, with voltage increasing from 28 V in the pristine film to 92 V in the micro-pyramid patterned film, and current rising from 2.7 μA to 11.0 μA. The enhanced power density and cyclic test suggests that surface topography optimization is highly effective, supporting long-term cyclic operation, and energy storage in capacitors. This work highlights the potential of surface-engineered nanogenerators in advancing sustainable, self-powered technologies.",
"conclusion": "Conclusion In summary, a Hybrid Triboelectric and Piezoelectric Nanogenerator was fabricated by using PDMS and barium titanate as the primary materials. The surface morphology of the composite film was changed by employing micro-pyramids to its surface with the help of lithography. For the nanogenerator, PDMS-BT film was used as a tribonegative material, meanwhile, Al foil was used as a tribopositive layer as well as the electrode and contact separation mode was utilized to couple piezoelectric and triboelectric effects together. The pressing of film, induced dipoles in it due to deformation, meanwhile the triboelectric charges were formed due to the contact and separation. The results of output voltage and current indicated that the surface morphology plays a huge role in enhancing the triboelectric effects as increasing the contact surface area can significantly improve its output performance. The density of patterns plays an important role in increasing the friction between the surfaces as well and hence our pattern density ( D ) of about 10 000 cm −2 also contributed largely towards the increments in the values of output voltage and current. These results offer promising strategies for fabricating high-performance and sensitive HBNGs for wearable devices and sensors.",
"introduction": "Introduction In recent decades, extensive research has focused on low-power energy harvesting from environmental sources such as vibrations, light, and wind, which would otherwise be wasted. Among these sources, kinetic energy in the form of vibrations, friction, and random displacement has garnered particular attention from researchers. This energy can be harnessed to power low-power and ultra-low-power electronic devices with wireless connectivity. 1–8 To extract kinetic energy, there are various energy harvesting strategies using electromagnetic, 9 electrostatic, 10 piezoelectric, 11 triboelectric, 12 pyroelectric, and thermoelectric mechanisms. Among these, piezoelectric and triboelectric energy harvesting utilize quite efficient mechanisms because they are inexpensive, environmentally friendly, and have high output performance. 10,12–20 Nanogenerators have gained significant attention in renewable energy research for their capability to harness mechanical energy from the surrounding environment. Among them, triboelectric nanogenerators (TENGs) stand out due to their straightforward design, a wide variety of material options, easy of fabrication, and cost efficiency. 21 Recently, TENGs have been employed to capture mechanical energy from various sources, including wind, blood flow, eye movements, typing actions, and ocean waves. Moreover, nanogenerators have found applications in powering a variety of sensors, such as pressure, mercury, photodetector, humidity, ion, and health monitoring sensors. 21 However, a key challenge in TENG technology is the instability in surface friction, which leads to variations in triboelectric performance over time. Frictional heat generated during the operation of TENGs is known to impair performance and longevity. To address this, shape memory materials like polyurethane utilize this heat, improving performance by altering key parameters such as the dielectric constant and charge density. Additionally, TENGs perform better in critical touch/non-touch states when approaches like ferromagnetic cilia-based TENGs (FC-TEGs) are employed, as they help sustain output without increased wear. The application of nonpolar semisolid lubricants further reduces friction and enhances the mechanical lifespan of TENGs. These innovations help mitigate the limitations of friction and wear, leading to more efficient and durable energy-harvesting devices. 22 However, it is not appropriate to fabricate by pristine materials because they exhibit low sensitivity and output power. The power density of the energy harvester using conventional materials for TENG is generally low, about 0.03–0.9 mW cm −2 . 23 Piezoelectric Nanogenerators (PENG) produce relatively low output power when used independently. 24 The output is in the ranges of nano-amperes and low voltages. Recent approaches in architected triboelectric and piezoelectric materials have shown significant advancements in energy harvesting and sensing applications. Nanogenerators based on piezoelectric or triboelectric materials have emerged as an attractive cost-effective technology for harvesting energy from renewable sources and for human sensing and biomedical applications. 25 The development of novel biocompatible soft materials and micro/nano-structured or chemically functionalized interfaces has opened new opportunities in this field. One interesting approach is the hybridization of piezoelectric and triboelectric effects in coupled nanogenerators (HBNG). These devices can make full use of mechanical energies and achieve both higher output and sensing performance. 26 The integration concept and performance enhancement strategies of HBNG have been focused on structural simplification and efficiency improvement, leading to the development of all-in-one mechanical energy-scavenging and sensing devices. Hence it is quite effective to incorporate piezoelectric materials with triboelectric materials to fabricate hybrid PENG and Triboelectric Nanogenerator (TENG) devices. 26 However, Hybrid piezoelectric/triboelectric nanogenerators (HBNG) have delivered higher outputs in the ranges of nano-amperes and low voltages. The global market revenue was estimated to be about 1B in 2020, which is anticipated to rise to € 3.2B by the end of 2027, with the contribution of nanogenerators in automobile and biomedical products reaching ∼ € 59M-with an annual growth rate of more than 10%. This rapidly growing HBNG revenue in these sectors are inspiring to overcome the manufacturing barriers and material limitations and their performance in different medical sectors, which are now incorporated into millions of patients and this number is predicted to increase tenfold by 2030. 27 The main component is the TENG, and PENG is a supporting material to enhance electrical power generation. TENG is based on the phenomena of triboelectrification and electrostatic induction in which a material becomes electrically charged after it is brought into contact with a different material through friction, therefore the output performance of TENG is related to materials that are being utilized, and the surface morphology corresponding to the contact area. 28–35 Although there has been a great use of these techniques, they are very likely to become less popular as they have the disadvantage of producing very low output voltage and when scaled to a lesser size, their output performance is affected which leads to the problem of not being able to miniaturize the device. When an improved and creative structure is implemented with the usage of superior materials to increase the output, the cost of the device also goes up, which is not so effective in the long run and limits its application as a sustainable energy harvesting device. For TENG devices, neither magnets nor coils are required, they are inexpensive, have light mass, and can be fabricated using organic material as well. Furthermore, a self-powered electronic skin nanogenerator based on triboelectric and mechanoluminescent properties has been developed, capable of distinguishing multiple stimuli through a strain-sensitive mechanoluminescent layer. 36 Due to their simplicity, wide range of materials, and cost-effective production, TENGs have become widely popular in different applications. Researchers have explored numerous materials, including polymers, metals, and composites, to enhance the triboelectric effect and thus improve the device's efficiency. Common materials include polydimethylsiloxane (PDMS), 37 polytetrafluoroethylene (PTFE), 38 and other polymers that offer a high surface charge density. Despite change in the crystal structure of the materials, the surface morphology plays a vast role in increasing the output performance of TENG devices, increasing the surface contact area can generate more triboelectric charge during the contact and friction process. Different structures and designs have been investigated to produce high-performance TENG devices. 28–35 Initially, the scope for TENG was limited to micro-scale energy harvesting to be used for electronic network applications, but gradually their scope was extended into many useful applications including biomedical sensors, self-powered devices, actuators. 28–35 This can be done by patterning the surface with different structures at the micro-scale i.e. , cubes, semi-circles, and pyramids. Among the choice of materials in the triboelectric series, PDMS is an excellent candidate as it has excellent thermal, elastic, and mechanical properties and is environmentally stable. PDMS has a significant ability to gain electrons and stands out due to its flexibility, transparency, and durability. 37,39–45 To improve the electrical, thermal, and dimensional stability of PDMS, ceramic fillers can be incorporated into it to synthesize a composite. 26 For this purpose, barium titanate (BT) is a good choice of material to be used as the ferroelectric and piezoelectric material in the hybrid nanogenerator. The pyroelectric effect in BaTiO 3 facilitates charge generation due to temperature variation, offering further energy harvesting potential. This phenomenon is well-suited for use in self-powered sensing and energy harvesting applications and works in conjunction with triboelectric and piezoelectric effects. 20 BT is a ferroelectric material; it has a perovskite structure, and can attain four crystal structures according to the change in temperature. 46–48 The phase diagrams of BT show that the phase transitions occur at quite low temperatures which makes them an attractive candidate for probing mechanisms of enhanced piezoelectricity. 49–52 Above the Curie temperature BT shows a cubic crystal system, but at temperatures lower than the T c , its structure changes from cubic to tetragonal, with the orientation of dipole moment, which results in ferroelectricity in the material. The piezoelectric coefficient ( d 33 ) value of BT synthesized by a solid-state reaction is almost 190 pC N −1 and the dielectric permittivity for BT is 1670 at ambient temperature, hence it is suitable to be used for energy harvesting purposes owing to its high dielectric permittivity, which leads to high power capabilities from E = (1/2) CV 2 . 53–57 Several efforts have been made to better utilize the spontaneous polarization of BaTiO 3 , including controlling the morphology of the materials, building heterojunctions, and creating small-sized devices for energy harvesting applications. Combining the above-mentioned advantages into BT-based NG is promising research and surface morphological changes make it a potential candidate for various applications. 53–57 The Hybrid Nanogenerator (HBNG) can be used to draw high electrical energy outputs. The devices can potentially be used as wearable tactile sensors and for energy harvesting. The constant motion of water, industrial vibrations, human body movement, and road vibrations can be effectively transformed into energy and used as a constant supply of electricity without any interruption. 58–61 The present article investigates the impact of modifying the surface morphology of the BT-PDMS film to enhance its contact area. Also, we propose highly sensitive HBNG with enhanced voltage and current by using the micro pyramid pattern incorporated with BT and PDMS. The BT particles having different concentrations of PDMS were investigated and optimized. With these changes, the output voltage of the HBNG device showed a noticeable increase, accompanied by a rise in current. The HBNG fabrication technology holds a viable potential in electrical flexible devices, biomedical, sensors applications.",
"discussion": "Results and discussions The crystal structure of synthesized BaTiO 3 powder samples were characterized by an X-ray diffractometer with Cu/Kα radiation. In Fig. 3(b) , the diffraction peaks for the prepared sample calcined at 850 °C matched with the JCPDS card no. 01-079-2265, corresponding to tetragonal phase BaTiO 3 with space group P 4 mm . The observed peaks at 2 θ of 22.06°, 31.57°, 38.80°, 44.85°, 50.09°, 56.26°, 65.91°, 74.91°, 79.17° correspond to the (100), (101), (111), (002), (102), (211), (202), (301), and (113) planes, respectively. There are some extra peaks belonging to BaCO 3 and Ba 2 TiO 4 corresponding to JCPDS card no. 00-041-0373 and ICDD card no. 00-035-0813 respectively. The presence of BaCO 3 and Ba 2 TiO 4 indicates that the reaction was insufficient at the calcination temperature, therefore by products can be removed when the material is sintered at higher temperatures. The lattice parameters of BaTiO 3 by Rietveld analysis of sintered sample in Fig. 3(a) are shown in Table 1 . The values of a , c , and the c / a ratio exhibit index of tetragonality behavior of the BaTiO 3 . Full-width half maximum (FWHM) and concentration of the tetragonal phase was analyzed and measured by Rietveld analysis. The crystallite size of BaTiO 3 was 4–7 μm. The absence of secondary peaks or impurities in the XRD pattern of the sintered BaTiO 3 sample at 1150 °C indicates a pure phase formation without significant side reactions. The sharp diffraction peaks of BaTiO 3 (JCPDS 01-079-2265) and the split (200) peak around 2 θ range in 44°–45°, as illustrated in Fig. 3(c) using Rietveld analysis, confirm the presence of the tetragonal phase and demonstrate the sample's high crystallinity. The refined lattice parameters suggest a c / a ratio, which is a hallmark of the tetragonal phase essential for piezoelectric activity as mentioned in Table 1 . Fig. 3 (a and b) XRD patterns of barium titanate (BaTiO 3 ) powder and sintered samples at 850 °C and 1150 °C, showing phase evolution with temperature, with peaks for BaTiO 3 (BT), BaCO 3 (B), and Ba 2 TiO 4 (B 2 ). (c) Rietveld analysis of the 1150 °C sintered sample, demonstrating precise peak fitting and quantitative phase composition, which confirms high crystallinity and structural refinement of BaTiO 3 . (d) Crystal structure schematic of BaTiO 3 , illustrating the perovskite lattice with Ba at the corners, Ti at the center, and O ions on the faces, a structure crucial for its ferroelectric properties. Table 1 Summarize the lattice parameters of the BaTiO 3 powder by Rietveld analysis Centre peak FWHM Crystallite size (μm) Concentrate (%) \n a \n \n c \n \n α \n \n c / a Tetragonal \n 111 \n 4.007 4.038 90 1.00775 111T 38.8784 0.22509 7.913 100 \n 200 \n 002T 45.2234 0.36821 4.089 70.358 200T 44.8564 0.47837 5.346 29.642 \n Fig. 3(a) shows the schematic illustration of the specification and layout of pyramid surface morphology engineered on the BT-PDMS (20%) surface. By increasing the contact area through these micro-pyramids in the composite film, it is expected to yield a higher electrical output due to more effective energy transfer and interaction with its uniform pyramid environment as compared to the pristine film. Scanning electron microscopy (SEM) and optical images in Fig. 4(b and c) provide a visual confirmation of the mold's micro-pyramid-structure, captured at different magnifications. These images show a detailed, magnified structure that reveal the perfect texture of the mold, leading to the uniformity of the structure, ensure the structure meets the intended design (base of the pyramid is 61 μm × 61 μm, the top of the pyramid is 13 μm × 13 μm and the height is 34 μm). The pyramids are at the intervals of 40 μm, and due to the regularity and high-density topography of the film produced by the mold. This consistency is a key factor to maintaining uniform contact across the film, which ultimately affects HBNG performance. Fig. 5(a–f) shows the SEM images of BT particles at different magnifications. The images revealed that the particles have a uniform structure and are well dispersed with a spherical morphology. The sizes of particles range from 150 nm to 270 nm. Such consistent shapes and dispersion are advantageous, to ensure the uniform distribution of BT particles in PDMS matrix, leading to the film's homogeneous structural and morphological properties. Elemental analysis and mapping of BT was performed by the EDS detector attached to the SEM. EDS analysis provide the information of elemental composition and confirmed the desired elements barium (Ba), titanium (Ti) and oxygen (O), respectively. Mapping along with Energy Dispersive X-ray spectroscopy (EDS) visualized the distribution of Ba, Ti and O in BT, verifying the uniform distribution of these elements with the BT materials to avoid agglomeration in the film as shown in Fig. 5(h) . Fig. 4 (a) Schematic showing detailed specifications of the micro-pyramidal patterning with precise measurements of dimensions. (b) SEM images at various magnifications display the uniform arrangement and accuracy of the micro-pyramidal structures on the silicon mold. (c) Optical images at 20× magnification reveal the patterned surface of the silicon mold, demonstrating the consistency and clarity of the micro-structured design. Fig. 5 (a–f) SEM images of barium titanate (BaTiO 3 ) powder at various magnifications, displaying particle morphology and size distribution. (g) EDS analysis for elemental composition, confirming the presence of Ba, Ti, O, and C in the sample. (h) Elemental mapping images for Ba, Ti, and O, showing uniform distribution of these elements within the BaTiO 3 powder, indicating consistent material composition. In Fig. 6 , micro pyramid pattern shows the clear and uniform distribution of micro-pyramids pattern of the BT-PDMS (20%) film structure. This regularity in patterning indicates a well-structured topography, essential for HBNG that relies on consistent surface properties. The BT-PDMS (20%) films reveal the clear integration of the BT particles in the PDMS matrix. These particles are submerged in the PDMS matrix, ensuring the film's mechanical stability and functional surface area. The EDS analysis was also conducted to identify the elemental composition of BT within the PDMS film. Fig. 6 (a–f) SEM images showing the surface topography of the BT-PDMS composite film with micro-pyramidal structures at various magnifications, illustrating the uniformity and precision of the micro-pyramids. (g–k) SEM images of the pristine BT-PDMS film at different scales, highlighting the distribution of BaTiO 3 particles within the PDMS matrix. (l) EDS elemental analysis of the BT-PDMS film, confirming the presence and distribution of elements (Ba, Ti, O, Si, and C) in the composite. Thermogravimetric analysis was conducted to study the thermal stability and decomposition behavior of the synthesized BT material, tracking the weight loss of the material as a function of temperature as shown in Fig. 7(a) . There are several weight loss-steps that occurred during the measurements when the temperature is increased from 25 °C to 1050 °C at a rate of 10 °C min −1 . Throughout this temperature span, the BT material experienced an overall weight loss of 0 to 4.5%. In the initial phase (P1), the significant first weight loss is observed sharply at 300 °C. This dramatic weight loss is likely due to the removal and burnout of residual organics arising during the synthesis of material, along with the release of moisture during high temperatures. In the second phase (P2), additional, more gradual weight loss is observed at around 800 °C, likely caused by the attributed to the amount of chemisorbed water or hydroxyl ion incorporated in the BaTiO 3 samples. The first phase (P1) shows pronounced weight loss than in to second phase (P2), due to the presence of organics and physiosorbed water in the materials, as supported by prior studies. 46–48 Fig. 7 (a) Thermogravimetric analysis (TGA) of barium titanate powder, showing weight loss as a function of temperature. The two main phases (P1 and P2) indicate the burnout of organics, release of physiosorbed water, and incorporation of hydroxyl ions into the BaTiO 3 lattice. (b) Raman spectroscopy of barium titanate powder, displaying characteristic peaks at 185, 304, 519, and 722 cm −1 , which correspond to vibrational modes in the BaTiO 3 structure. The Raman spectroscopic data for the calcined BaTiO 3 (BT) powder is shown in Fig. 7(b) , providing comprehensive details on the material's geometry, local symmetry, and crystal structure. Through changes in vibrational modes, Raman spectroscopy, which is renowned for its sensitivity to local crystal structures and symmetry, efficiently records structural changes. A change in the local crystal structure of BT was indicated by an increase in vibrational amplitude. 22 Given that the greater vibrational amplitude indicates a more stable and ordered lattice structure, this shift implies a decrease in lattice disorder. The spectra revealed the tetragonal crystal structure of BaTiO 3 , characterized as four peaks were observed, each mode corresponding to a particular vibration within the crystal structure. These specific peaks confirm the existence of the tetragonal phase, known as a stable structure for BaTiO 3 , particularly after calcination. The calcination process appears to have enhanced the material's structural order, as indicated by the increased vibrational amplitudes and the clarity of the peaks. These different modes at 185, 304, 519 and 722 cm −1 confirm the presence of the tetragonal phase and unique vibrational in BT powder, which is consistent with previous studies. 23,24 These peaks show the reliability of the calcination process not only stabilized the crystal structure but also reduced the crystal disorder, leading to more distinct vibrational response and enhancement of the material's structural integrity. Previous work achieved a significant rise in the dielectric constant even below the percolation threshold, in contrast to earlier research that claims an elevation of the dielectric constant near the percolation threshold. The dielectric characteristics of composite systems having a conductive phase scattered throughout an insulating matrix, such as PZT–Ag and PZT–Pt, 62 have been explained by certain research. 59 Such metal–dielectric composites' dielectric response is in good agreement with percolation theory, indicating that the observed rise in the dielectric constant might be attributed to the higher dielectric field created around the metal particles. The electrical and dielectric characteristics of the sintered sample are reported in Table 2 . 62 Table 2 Summarize the dielectric and electrical properties of sintered sample Sr # Temperature Sintered density Dielectric constant (10 kHz) Tangent loss Breakdown strength Figure of merits (10 kHz) BaTiO 3 1150 °C 98% 2560 0.023 1.8 4608 It is known that using piezoelectric materials in TENG devices could improve their output performance, 3,63,64 Therefore, in our study, a 20% volume fraction of BT was incorporated into PDMS and the resulting composite film was paired with aluminum foil. This choice was made because PDMS and aluminum are positioned far apart in the triboelectric series, suggesting enhanced charge generation. The composite layer acquires a negative charge due to its tendency to gain electrons, while aluminum foil tends to lose electrons, resulting in a positive charge distribution ( Fig. 8 ). Fig. 8 Dielectric constant ( ε ) and tangent loss ( D ) of sintered barium titanate as a function of frequency. The dielectric constant decreases with increasing frequency, while the tangent loss initially decreases before sharply increasing at higher frequencies. In the initial stage of discussing the piezo response separately in the presence of BT, when the layers have not in contact with each other, the electrical output would be zero as there is no application of pressure now, leaving no net charge on the film surface. When the films come into contact with each other by pressing on the application of force, piezoelectric charges are generated due to the imbalance of electric dipoles caused by the stress induced orientation of particles inside the composite film. This leads to the creation of a piezoelectric potential gradient within the film by the appearance of net positive and net negative charges on the opposite surfaces of film, which results in flow of charges and hence the production of current. When the pressing has been done, again there is no flow of charges as the deformation of the film has been stopped. As the compressive force is being released, the charges flow back to maintain the equilibrium leading to an output current again and after the compressive force is fully removed, the original state is reached, completing the cycle. Alongside the piezo response, there is also the generation of triboelectric charges as shown in Fig. 9 . The device won't produce any triboelectric charges in the initial state as there is no contact or friction between the triboelectric layers. However, when pressure is applied, the surface of the layers gets in contact with each other, and triboelectric charges are generated. The composite layer would gain electrons from Al foil as it has a stronger affinity for a negative charge, due to this Al foil would attain a positive charge while the composite layer becomes negatively charged. Now the electric polarity has been generated after the pressure is removed and the layers are separated. Due to electrostatic induction, the composite would induce its oppositely generated charges from Al foil to the bottom Al foil, leading to a positive output current when electrons were driven to the top Al foil. The current generated at this point also included the piezoelectric induced charges as mentioned previously, leading to dual phenomenon of triboelectric and piezoelectric effects occurring at the same time which results in enhanced electrical output. When the layers are fully separated by the distance of 1 cm, then there would be no charge transfer due to electrical equilibrium. The pyramid structures strongly influenced the electrical output of the device and there was a huge increase in the values of voltage and current as compared to the device with the regular surface as shown in Fig. 10 , due to the increased contact area and there is more charge transfer during friction. Fig. 9 Schematic illustration of the charge generation mechanism in the HBNG device, highlighting the interaction of dipoles, triboelectric charges, and piezoelectric charges within the composite film and aluminum foil layers. The central image shows the practical application of the device, with arrows indicating the charge flow during mechanical deformation. Fig. 10 Output current and voltage of the HBNG device with different surface morphologies: (a and b) standard BT/PDMS composite at 10, 20% concentration, showing moderate voltage and current under compression and release; (c) micro-pyramidal BT/PDMS structure at 20% concentration, exhibiting significantly higher voltage (up to 90 V) and current output due to enhanced surface morphology, demonstrating the performance advantage of the micro-pyramid design. To calculate the output voltage and current of the devices, an oscilloscope was used. To measure the current, a 100 kΩ resistor was used externally to obtain current graphs as shown in Fig. 10 . For the HBNG, the contact-separation method was employed to generate the voltage and current, the mechanical force of 800 N was used as the operation condition and both devices had a contact area of 4 × 4 cm 2 with the test distance between the two layers of about 1 cm. For the contact separation mode, two dissimilar dielectric films are facing each other, and electrodes are attached at the top and bottom of the stacked films, when the films come in contact with each other, it creates oppositely charged surfaces, and a potential drop is created as the films are separated. When the films are connected by a load, free electrons flow in order to balance the electrostatic field, and the potential is dropped again as films come in contact with each other again as shown in Fig. 10 . The oscilloscope readings revealed that the output voltage of the nanogenerator with a regular surface reached at most about 11, 27 V and the output current reached around 2.1, 2.7 μA for 10 and 20% BT/PDMS, meanwhile for the nanogenerator with pyramid topography, the maximum output voltage reached up to 92 V and the output current reached around 11 μA as shown in Fig. 10(b) . From the values, it is evident that the value of voltage increased up to three folds, meanwhile the current increased up to five folds as compared to pristine nanogenerator. In the previous studies, Kai-Hong et al. employed three different microneedle patterns on AL/PDMS TENG device, these three patterns included: Overlapped Microneedle (OL-MN), Overlapped Two-Height Microneedle (OL-TH-MN), and Overlapped Deep Two-Height Microneedle (OL-DTH-MN). 65 The density of patterns ( D ) was such that OL-MN had a pattern density of 654 MN cm −2 while OL-TH-MN and OL-DTH-MN had a pattern density of 965 MN cm −2 , similarly, the surface contact area of OL-MN, OL-TH-MN, and OL-DTH-MN was 22.91 × 10 3 , 24.38 × 10 3 , and 29.69 × 10 3 mm 2 and the output current of these devices were 109 μA, 117 μA, and 129 μA as mentioned in Table 3 . The parameters of micro-needles and our work are compared, and the results indicated that the output performance of TENG devices is closely related to the pattern density. From the results, it is evident that surface morphology has played a huge role in enhancing the output performance of the device. Some of the recent work done on the surface morphology of TENG devices has been mentioned in Table 4 . Researchers have used different methods to employ complex morphology to their material including lithography, chemical treatment, and CO 2 laser ablation as shown and indicated that surface morphology plays a huge role in enhancing the output performance of the TENG devices. The output performance of triboelectric nanogenerators is dependent on the contact surface area of the friction layers and thus for the case of nanogenerator with pyramid topography, the contact area has increased as the surface area would be calculated as “the number of pyramids” multiplied by the “pyramid surface area”. 65 The more the number of pyramids, the more would be the effective contact area for enhancing triboelectric properties during contact and separation deformation. The density of patterns (pyramids per cm or PD per cm) is associated with the output performance of the TENG device and is calculated as such: Density ( D ) = N / A where as N represents the number of pyramids and A is the total area of the PDMS/BT plane. In our case, the density of patterns is 10 000 cm −2 and the area of a single truncated pyramid is 1159 μm 2 . Table 3 Comparison of patterned parameters and their output performance Samples Height (μm) Width (μm) Distance between patterns (μm) Pattern density (N cm −2 ) Output current (μA) Output voltage (V) OL-MN 13 118 268 1680 654 109 123 OL-TH-MN 1235 264 13 390 965 117 127 OL-DTH-MN 1528 274 15 787 965 129 167 PD (this work) 34 61 39 10 000 11 92 Table 4 Comparison of recent work done on surface morphology and their output performance Morphology Materials Fabrication method Operation mode Output current (μA) Output voltage (V) Ref. Pyramids ITO/PET-PDMS Lithography Linear motor 0.7 18 \n 66 \n Nano-pattern textile Ag/PDMS-ZnO Chemical treatment Mechanical force simulator 65 120 \n 67 \n Cubes ITO/PET-PDMS Lithography Linear motor 0.7 18 \n 66 \n Nano-pillars Au/PDMS ICP etcher 130 N, 3 Hz 3.2 83 \n 68 \n Micro-pillars Al/PDMS Lithography 10N, 5 Hz 8.3 72 \n 69 \n Micro-needle Al/PDMS CO 2 laser ablation Hand tapping 43.1 102.8 \n 70 \n Overlapped micro-needle Al/PDMS CO 2 laser ablation Hand tapping 109.7 123 \n 65 \n Overlapped 2-height micro-needle Al/PDMS CO 2 laser ablation Hand tapping 117.6 127 \n 65 \n Overlapped deep 2-height micro-needle Al/PDMS CO 2 laser ablation Hand tapping 129.3 167 \n 65 \n Octet truss BaTiO 3 3D printing Steel ball drop 52 101 \n 22 \n Microstructured surface Cu/PTFE 3D printing Mechanical force simulator 8 3860 \n 71 \n Curved architect PTFE/Cu/PP/Cu/PET 3D printing Mechanical force simulator 0.01 40 \n 72 \n Micro pyramids Al/PDMS-BT Lithography Hand tapping 140 1300 This work \n Fig. 11 shows measured output voltages and currents of the energy harvester based on the BT film with 20% concentration of PDMS. In HBNG, the performance parameter varies with the load resistance in case of micro-pyramidal BT/PDMS structure. Specifically, V oc and I sc shows the opposite trends according to the load resistance. To measure the generated output power of the energy harvesting system, a load resistor or capacitor was used to measure the maximum output power and energy. The maximum power density was obtained by optimizing the load resistance. As the load resistance increases, the maximum current produced by HBNG decreases. This decline is primarily due to the ohmic losses due to higher resistance, it limits the current flow, leading to the reduction in current output from 250 to 0.2 μA. In contrast, V oc increases as load resistance increases due to the high resistance circuit, the V built across the load. This behavior led to the increase in output load voltage from 76 to 0.5 V. Fig. 11 (a) Open-circuit voltage (black) and short-circuit current (blue) of the micro-pyramidal BT/PDMS HBNG as functions of load resistance, showing the device's electrical response. (b) Power density of the micro-pyramidal BT/PDMS HBNG as a function of load resistance, highlighting optimal power output. (c) Stored voltage in a capacitor circuit using the HBNG, demonstrating energy storage capability. (d) Reliability comparison of the micro-pyramidal HBNG, illustrating performance stability over time. \n Fig. 11(b) illustrates the output power generated by the energy harvester, calculated based on the voltage and current across an applied external load. The output voltage and current were calculated by varying the external load resistance, ranging from 100 Ω to 100 MΩ, connected to the BT/PDMS film energy harvester. The output power ( P ), was determined by: P = I L V L where I L and V L represent the output current and voltage across the load resistance, respectively. As shown in Fig. 10(b) , the output power of the energy harvester first increased and then decreased. The maximum output power was 2500 μW at an optimized load resistance of 37 MΩ, corresponding to a voltage of 23 V and a current of 92 μA. After this peak value, the power generated output decreased. This trend indicates the load resistance at the maximum power transfer occurs for the energy harvester. Additionally, the power density can be expressed as: Power density = Generated output power/volume The generated output power of the piezoelectric energy harvester based on the micro-pyramidal BT/PDMS film was 250 mW cm −3 based on the device size of 3 × 3 cm 2 . Numerous studies have been conducted on HBNG, though most tend to show limitations and low power density. To address this issue, this micro-pyramid BT/PDMS film structural design elevates the frictional contact and surface charge density, resulting in a tremendous increase in power density. Table 4 compares the voltage and current of various TENG, PENG and HBNG, which uses various materials and fabrication techniques under similar pressure conditions. The table shows that our micro-pyramidal BT-PDMS HBNG achieves a significantly higher power density than those in previous studies, demonstrating the effectiveness of this approach in energy output. The current load ( I L ) can be described by the equation: I L = V /( R HBNG + R L ), where R HBNG and R L represent the resistances of the micro-pyramidal BT/PDMS film and the applied external load, respectively. Consequently, the output power across the load P L can be written as: P L = I 2 R L = ( V /( R HBNG + R L )) 2 R L = V 2 R HBNG 2 / R L + 2 R HBNG + R L The maximum value of P L occurs at the minimum value of the denominator, and therefore, the derivative of the denominator of R L can be expressed as: d/d R L ( R HBNG 2 / R L + 2 R HBNG + R L ) = − R 2 HBNG / R 2 L +1 = 0. Consequently, the maximum P L value occurs when R HBNG = R L . In our study, Fig. 11(c and d) show the stored voltage and reliability results of the energy harvester based on the micro-pyramidal BT/PDMS film. Under conventional ceramic capacitors with capacitances of 1 μF, the stored voltage of the energy harvester increased up to 8 V within 50 seconds, when mechanical forces were applied showing the HBNG impressive charging capability. In energy harvesting applications, ensuring electrical output despite the periodic external forces by repeated and intense friction are essential for the performance of the HBNG. A cyclic test, to assess long-term stability, was performed to examine the reliability of the output performance of the film. During this test, HBNG underwent 1000 cycles of pressing and releasing by mechanical forces without any observable drop in the open-circuit voltage ( V oc ). The maximum value of the voltage is the same at the beginning and the end of the cycle test, showing the stable power generation, the energy harvester showed excellent mechanical resilience and a stable output performance even under constant external pressure. The results demonstrate that the micro-pyramid BT/PDMS film-based energy harvester delivers exceptional output performance with no signs of degradation or fatigue over time. \n Fig. 12 presents the output performance and versatile detection capabilities of our HBNG nanogenerator, designed with a micro-pyramidal structure, as depicted in the graphical representation in Fig. 13 under a range of everyday and environmental conditions. The HBNG effectively detected various human and environmental motions, such as pressing, tapping, and stepping, with output signals strengthening as the applied pressure increased (pressing < tapping < stepping). During stepping, the V oc reached a maximum of 92 V, though this output fluctuated due to the irregularity of impact forces. When the device was attached to a human foot, it enabled precise monitoring of real-time activity by capturing subtle V oc pulse patterns, demonstrating its potential for accurate human motion tracking. In addition, the device showed robust performance in daily scenarios, generating output voltages of 92 V, 24 V, and 2.8 V under stepping, tapping, and pressing conditions, respectively. These tests were conducted at standard room conditions (23 °C and 47% humidity), underscoring the nanogenerator's adaptability and high sensitivity for a variety of applications, including wearable technology, environmental sensing, and human activity monitoring across different conditions. This performance highlights the HBNG's suitability for flexible, real-world applications. Fig. 12 Output voltage response of the HBNG nanogenerator under different mechanical actions: (a) stepping shows a strong response up to 90 V, ideal for high-impact motion detection; (b) tapping generates peaks up to 30 V, indicating sensitivity to moderate pressure; (c) pressing produces consistent signals around 2.8 V, suitable for low-force sensing applications. Insets depict the testing setup for each action. Fig. 13 Schematic illustration of the HBNG nanogenerator with a micro-pyramidal structure, showing the layered composition of Kapton tape, aluminum foil, and PDMS-BaTiO 3 composite. The micro-pyramidal substrate design enhances contact sensitivity, facilitating higher output voltages under mechanical stress."
} | 10,427 |
27066184 | PMC4802747 | pmc | 68 | {
"abstract": "ABSTRACT Arbuscular mycorrhizal (AM) fungi form mutualistic interactions with the majority of land plants, including some of the most important crop species. The fungus takes up nutrients from the soil, and transfers these nutrients to the mycorrhizal interface in the root, where these nutrients are exchanged against carbon from the host. AM fungi form extensive hyphal networks in the soil and connect with their network multiple host plants. These common mycorrhizal networks (CMNs) play a critical role in the long-distance transport of nutrients through soil ecosystems and allow the exchange of signals between the interconnected plants. CMNs affect the survival, fitness, and competitiveness of the fungal and plant species that interact via these networks, but how the resource transport within these CMNs is controlled is largely unknown. We discuss the significance of CMNs for plant communities and for the bargaining power of the fungal partner in the AM symbiosis.",
"conclusion": "Conclusions AM fungi and their CMNs play a significant role in plant ecosystems and control the fitness and competitiveness of the plant individuals within their CMNs. Our current understanding about resource exchange in the AM symbiosis is primarily based on experiments with root organ cultures or with single plants that are colonized by one AM fungus. 6,8 The transferability of these experiments to CMNs, however, is very limited, because in natural ecosystems both partners in the AM symbiosis can choose among multiple trading partners and do not depend on a single partner for their nutrient or carbon supply. Plants play a critical role for the carbon supply of their CMNs and the composition of the plant community within one CMN has been shown to affect the abundance or extension of CMNs in soils. 29,30 Very little is known about how AM fungi allocate nutrient resources or infochemicals within their CMN, or how host plants compete with other plants for nutrients that are available for their CMNs. More research is needed to better understand how the costs and benefits of the AM symbiosis are controlled in CMNs, and how fungal networks affect the inter-fungal or inter-plant competitiveness of both partners in natural ecosystems."
} | 556 |
31048787 | PMC6497639 | pmc | 69 | {
"abstract": "Perturbations in natural systems generally are the combination of multiple interactions among individual stressors. However, methods to interpret the effects of interacting stressors remain challenging and are biased to identifying synergies which are prioritized in conservation. Therefore we conducted a multiple stressor experiment (no stress, single, double, triple) on the coral Pocillopora meandrina to evaluate how its microbiome changes compositionally with increasing levels of perturbation. We found that effects of nutrient enrichment, simulated predation, and increased temperature are antagonistic, rather than synergistic or additive, for a variety of microbial community diversity measures. Importantly, high temperature and scarring alone had the greatest effect on changing microbial community composition and diversity. Using differential abundance analysis, we found that the main effects of stressors increased the abundance of opportunistic taxa, and two-way interactions among stressors acted antagonistically on this increase, while three-way interactions acted synergistically. These data suggest that: (1) multiple statistical analyses should be conducted for a complete assessment of microbial community dynamics, (2) for some statistical metrics multiple stressors do not necessarily increase the disruption of microbiomes over single stressors in this coral species, and (3) the observed stressor-induced community dysbiosis is characterized by a proliferation of opportunists rather than a depletion of a proposed coral symbiont of the genus Endozoicomonas .",
"introduction": "Introduction In natural systems, disturbances or stressors rarely occur in isolation. Anthropogenic impacts disrupt individual animal physiology, alter whole populations or community dynamics, and drive shifts in system-level processes thereby putting biodiversity in peril 1 – 4 . Therefore, it is imperative to characterize how multiple stressors interact to disrupt natural systems. We are using the operational definitions of types of interactions between multiple stressors as defined by Folt et al . 3 and Vinebrooke et al . 4 . An additive effect, or null interaction, occurs when the combined effect equals the sum of the separate effects. A synergistic interaction occurs when the combined effect of multiple stressors is greater than the additive effect. And lastly, an interaction is deemed antagonistic when combined stressors produce a biological response that is less than the additive effect. Despite the existence of multiple interaction outcomes, synergies are often emphasized in conservation literature, perhaps due to the risk of negative feedbacks accelerating ecosystem decline and degradation 5 . A balanced research agenda that looks for synergies and antagonisms is necessary to fully understand how mitigating local stressors will or will not compensate for global stressors 6 . For instance, improving water quality and decreasing water turbidity in seagrass systems may exacerbate the damaging effects of heat stress from global warming 6 . Similarly, marine invertebrates and their microbiomes are often faced with global stressors associated with climate change and local stressors such as nutrient pollution or overfishing 7 – 9 . Yet few studies empirically test the individual and combinatorial effects of more than two stressors on host microbiomes 7 – 10 . Current statistical methods and models for microbiome studies 11 , such as those that evaluate alpha and beta diversity and differential abundance, can be combined with multi-stressor experimental designs and used to statistically quantify the interacting effects of multiple stressors. For instance, patients with Crohn’s, a disease associated with gut microbiome dysbiosis, were treated with either antibiotics or a diet of exclusive enteral nutrition 12 . The two therapies likely disrupt the gut microbiome through different mechanisms and are independently associated with dysbiosis. In one case, the stressors produced opposite responses in the abundance of a single bacterial genus, Alistipes . Yet, an antagonistic interaction was not tested for but easily could be with a crossed design with patients receiving both therapies. In an environmental example, warm- or cold-stressed oysters crossed with bacterial infection by vibrios showed evidence of synergy as warm-stressed oysters experienced the highest mortality following infection 13 . When evaluating the oyster hemolymph microbiome, an interaction term of stress × infection in the univariate analysis of alpha diversity and multivariate analysis of beta diversity was not included, but if included in statistical methods, would clarify the type of interaction between the two stressors. Using robust statistical methods and interaction models benchmarked in the microbiome field 14 , we investigated how a global stressor, thermal stress, interacts with local stressors, nutrient pollution and predation, to alter the coral microbiome. Corals, currently experiencing major threats of climate change and nutrient pollution, can function as environmental sentinels and are thereby prime candidates for multiple stressor experiments. Previous studies of the coral microbiome have shown that stress tends to increase species richness 8 , 9 , 15 – 17 and cause shifts in community composition from potentially beneficial symbiotic bacteria that dominate healthy corals to potentially opportunistic or pathogenic bacteria that dominate stressed corals 8 , 9 , 15 , 17 – 20 . Beta diversity, or species turnover between samples, has also been reported to increase with stress 7 , 9 , 21 , 22 , and stressed corals have microbial communities distinct from control corals 23 – 26 . Therefore, we designed a fully-crossed experiment to investigate biological responses including alpha and beta diversity indices and differential abundance modeling of individual taxa with stress. For the purposes of this study, we define a stressor to be any external disturbance from the host’s environment that causes a quantifiable change in microbial community structure. We utilized univariate and multivariate statistical techniques to parse the main effects and interactions among stressors. The coral Pocillopora meandrina was exposed to increased seawater temperatures, pulse nitrate and ammonium enrichment, and simulated predation in a factorial mesocosm tank experiment with all possible combinations of these stressors. We hypothesized that local stressors like nutrient pollution and predation would interact synergistically with thermal stress to reduce the host’s ability to regulate its microbial community which would be manifested by: (1) an increase in the compositional heterogeneity and variability (beta diversity) among stressed corals compared to the controls, and (2) an increase in community evenness in stressed corals as a result of (3) shifts from few dominant symbiotic bacterial taxa to a myriad of potentially opportunistic bacterial taxa that bloom and become overrepresented in stressed corals.",
"discussion": "Discussion Contrary to our hypothesis, our overall results suggest the global and local stressors tested in this tank experiment generally do not act synergistically to induce dysbiosis in the coral microbiome of Pocillopora meandrina . In fact, we find that the biological response in the microbial community to stress does not scale positively with increasing number of stressors. We predicted that when sequentially adding stressors to the system, we would see a concurrent increase in deterministic changes to the microbiome (Fig. 5a,c,e ). For beta diversity, deterministic changes would produce clusters with increasingly distant locations from the control community (Fig. 5a ). Stochastic changes would likewise produce communities that were more dispersed or variable (Fig. 5c ). Instead, we found that the greatest deterministic changes in the microbial community resulted from single stressors, while interactions produced an intermediate level of change resulting in antagonisms that decreased the individual effects (Fig. 5b ). For stochastic changes, any environmental stressor was sufficient to induce dispersion around the centroid of healthy corals (Fig. 5d ), although this dispersion was likely a result of the single dominant taxon (Fig. 1 ). The changes in alpha diversity, however, did not scale positively with the number of stressors, and single stressors appear to increase community diversity more than two or three stressors combined (Fig. 5e,f ). We also found that stress induced a myriad of opportunists to invade the community, shifting species dominance away from coral symbionts. The dynamics observed in species’ abundance profiles of the microbial community following a perturbation may be explained by each particular microbes’ nutrient preference and competitive ability 31 . Figure 5 Conceptual description of predicted ( a , c , e , g ) vs observed ( b , d , f , h ) patterns with multiple stressors. Location ( a , b ) and dispersion ( c , d ) effects represent measures of beta diversity. Community evenness ( e , f ) represents patterns in Simpson’s index. Patterns in taxa differential abundance in log2FoldChange using the generalized linear model framework are displayed in g and h with the gray line denoting no effect or no interaction. Colors represent the type of stressor combination applied to the corals: none = teal, single = yellow, double = purple, triple = red. In contrast to more heterogeneous communities that may be more robust to changes in community evenness, the control corals were dominated by a single taxon initially and thus exhibited low evenness. We would predict then that any perturbation to the system would only increase evenness reflected in higher Simpson’s diversity (e.g., Fig. 2 ). As such, when stressors were applied to the coral host, bacterial community evenness increased when the dominance shifted from the OTU-Endo symbiont to other taxa such as Desulfovibrionaceae 8 , 19 , 32 , Enterobacteraceae 20 , 23 , Amoebophilaceae 33 , 34 , Moraxellaceae 35 , 36 , and Rhodobacteraceae 8 , 16 , 37 , 38 (Fig. 1 ). Contrary to previous work 8 , we did not see an increase in species richness with stress. This may be a result of the mesocosm tanks restricting natural presence/absence dynamics on the reef. Instead, the results suggest a reshuffling of microbial members rather than an increase of new species. Many microbiome studies seek to understand whether dysbiosis, or an imbalance in microbiota, is marked by an invasion or proliferation of pathogens or a depletion of beneficial bacteria 39 , 40 . Yet taxon relative abundance measures alone do not provide enough information to answer this question. Individual responses of taxa to stress and their contribution to microbiome dysbiosis can be assessed with differential abundance analysis 37 , 41 , 42 . Unlike diversity measures which are driven by the changes of dominant taxa in the community, differential abundance testing can identify changes in minor players in the community. DESeq2 can be used to model the abundance of each taxon independently, while accounting for the discrete positive nature of count data and the compositionality of the community using a generalized linear model (GLM) 43 . Using the linear model framework, we expected main effects to increase opportunistic bacteria, and interactions to produce synergistic effects as the community becomes increasingly compromised (Fig. 5g ). Instead, we found that two-way interactions produced antagonistic responses among opportunistic taxa. This apparent antagonism may be a result of a dominance effect, in which one stressor accounts for most or all of the biological response, changing susceptible taxa such that the second stressor has no additional effect 5 . Stressors may provide some benefit or resource that normally limits the abundance of opportunistic taxa. For instance, high temperature may increase bacterial reproduction and metabolic rate, while scarring may increase free nutrients in the form of amino acids or open niches. These results suggest that opportunists such as Rhodobacter or Desulfovibrio spp. are not co-limited by the resources provided by high temperature and scarring. For instance, opportunistic taxa may be proliferating at such a high rate due to increased temperature and increased reproductive rates, that additional free nutrients from scarring do not compound the effect. In contrast, three-way interactions resulted in synergies as invading taxa continued to increase in abundance (Fig. 5h ). This suggests that opportunistic taxa that had maximized their biological response under two stressors, were in fact co-limited by some resource provided by a third stressor. For instance, the addition of nitrogen may have allowed some opportunistic taxa to surpass the maximum threshold of reproductive or metabolic potential set by high temperature and scarring. Alternatively, the difference in interaction type between two-way and three-way interactions may be a result from the coral host’s compromised immune system 44 – 46 . The coral host effectively regulates it’s associated microbial community under two stressors with a heightened immune response. However, with the addition of a third stressor, innate immunity could be overwhelmed, and the host could no longer regulate its community, thereby allowing a synergistic proliferation of opportunists. Despite the current bias in interaction literature toward identifying synergies 5 , our study highlights multivariate and univariate statistical tools that can be combined with standard methods in microbial ecology to more accurately characterize interaction types to host-microbiome systems. Community diversity measures are standardly conducted in microbiome research 14 , however, they have rarely been used to explicitly test for antagonisms or synergisms between environmental stressors using a microbiome dataset 7 , 9 , 12 , 13 , 37 . Although there is no evidence that these measures respond linearly to stress, these analyses revealed unexpected patterns of community response to increasing amounts of stress. This study presents an initial evaluation of the utility of these community diversity measures in characterizing interactions between different combinations of stressors that are known to damage the coral host and produce compositional changes in its microbiome."
} | 3,616 |
30899212 | PMC6416793 | pmc | 71 | {
"abstract": "Over the past few years, Spiking Neural Networks (SNNs) have become popular as a possible pathway to enable low-power event-driven neuromorphic hardware. However, their application in machine learning have largely been limited to very shallow neural network architectures for simple problems. In this paper, we propose a novel algorithmic technique for generating an SNN with a deep architecture, and demonstrate its effectiveness on complex visual recognition problems such as CIFAR-10 and ImageNet. Our technique applies to both VGG and Residual network architectures, with significantly better accuracy than the state-of-the-art. Finally, we present analysis of the sparse event-driven computations to demonstrate reduced hardware overhead when operating in the spiking domain.",
"conclusion": "8. Conclusions and Future Work This work serves to provide evidence for the fact that SNNs exhibit similar computing power as their ANN counterparts. This can potentially pave the way for the usage of SNNs in large scale visual recognition tasks, which can be enabled by low-power neuromorphic hardware. However, there are still open areas of exploration for improving SNN performance. A significant contribution to the present success of deep NNs is attributed to Batch-Normalization (Ioffe and Szegedy, 2015 ). While using bias less neural units constrain us to train networks without Batch-Normalization, algorithmic techniques to implement Spiking Neurons with a bias term should be explored. Further, it is desirable to train ANNs and convert to SNNs without any accuracy loss. Although the proposed conversion technique attempts to minimize the conversion loss to a large extent, yet other variants of neural functionalities apart from ReLU-IF Spiking Neurons could be potentially explored to further reduce this gap. Additionally, further optimizations to minimize the accuracy loss in ANN-SNN conversion for ResNet architectures should be explored to scale SNN performance to even deeper architectures.",
"introduction": "1. Introduction Spiking Neural Networks (SNNs) are a significant shift from the standard way of operation of Artificial Neural Networks (Farabet et al., 2012 ). Most of the success of deep learning models of neural networks in complex pattern recognition tasks are based on neural units that receive, process and transmit analog information. Such Analog Neural Networks (ANNs), however, disregard the fact that the biological neurons in the brain (the computing framework after which it is inspired) processes binary spike-based information. Driven by this observation, the past few years have witnessed significant progress in the modeling and formulation of training schemes for SNNs as a new computing paradigm that can potentially replace ANNs as the next generation of Neural Networks. In addition to the fact that SNNs are inherently more biologically plausible, they offer the prospect of event-driven hardware operation. Spiking Neurons process input information only on the receipt of incoming binary spike signals. Given a sparsely-distributed input spike train, the hardware overhead (power consumption) for such a spike or event-based hardware would be significantly reduced since large sections of the network that are not driven by incoming spikes can be power-gated (Chen et al., 1998 ). However, the vast majority of research on SNNs have been limited to very simple and shallow network architectures on relatively simple digit recognition datasets like MNIST (LeCun et al., 1998 ) while only few works report their performance on more complex standard vision datasets like CIFAR-10 (Krizhevsky and Hinton, 2009 ) and ImageNet (Russakovsky et al., 2015 ). The main reason behind their limited performance stems from the fact that SNNs are a significant shift from the operation of ANNs due to their temporal information processing capability. This has necessitated a rethinking of training mechanisms for SNNs."
} | 979 |
31652510 | PMC6829311 | pmc | 72 | {
"abstract": "Memristor devices are generally suitable for incorporation in neuromorphic systems as synapses because they can be integrated into crossbar array circuits with high area efficiency. In the case of a two-dimensional (2D) crossbar array, however, the size of the array is proportional to the neural network’s depth and the number of its input and output nodes. This means that a 2D crossbar array is not suitable for a deep neural network. On the other hand, synapses that use a memristor with a 3D structure are suitable for implementing a neuromorphic chip for a multi-layered neural network. In this study, we propose a new optimization method for machine learning weight changes that considers the structural characteristics of a 3D vertical resistive random-access memory (VRRAM) structure for the first time. The newly proposed synapse operating principle of the 3D VRRAM structure can simplify the complexity of a neuron circuit. This study investigates the operating principle of 3D VRRAM synapses with comb-shaped word lines and demonstrates that the proposed 3D VRRAM structure will be a promising solution for a high-density neural network hardware system.",
"conclusion": "5. Conclusions In this study, a 3D VRRAM structure was newly proposed as the synapse of a neural network system. It was concluded that 3D VRRAM implemented as synapses can increase the chip area efficiency and simplify the neuron circuits. This study investigates the operating principle of 3D VRRAM using comb-shaped WL synapses and proves that this structure has promise for a neural network system. The accuracy of a neural network with 3D VRRAM synapses was measured by classifying 7 × 7 alphabet letter images using a circuit simulator. The guide training algorithm was optimized for hardware implementation because it does not include a backpropagation algorithm. Therefore, the guide training algorithm and winner-take-all methods were used to validate the performance accuracy of the 3D VRRAM synapses in a HSPICE simulation. The simulation results showed 80% accuracy until the inverted pixel count reached 12%. This means that 3D VRRAMs are usable as synaptic mimic circuits in neural network systems. A 3D vertical synapse with an integrated 3D VRRAM structure will be a promising solution for a high-density neuromorphic chip.",
"introduction": "1. Introduction In recent years, neuromorphic computing has emerged as a complementary system to the von Neumann architecture. Much of the research on neural network hardware implementation discusses how to connect large numbers of neurons and synapses. As a consequence, various memory devices such as static random-access memory, resistive random-access memory (RRAM), floating-gate (FG) memory, and phase change memory have been implemented as the synapse model in neural network hardware systems [ 1 , 2 , 3 , 4 ]. The most popular device-level component chosen to implement the synapses is the “memory resistor”, or memristor, because the resistance value of a memristor is a function of its historical activity. Moreover, energy efficiency is a key challenge of neuromorphic computing and RRAM is attractive for large-scale system demonstration due to its relatively lower energy consumption as compared with other synaptic devices [ 5 ]. The most common use of the memristor two-dimensional (2D)-crossbar is as a multiple memristor synapse since a single memristor cannot represent the positive and negative weights of synapses. However, 2D crossbar array synapses are not suitable for the implementation of deep neural networks (DNN) because the chip area depends on both the depth of the neural network and the number of input and output nodes. The three-dimensional (3D) vertical resistance random-access memory (VRRAM) promises to minimize the area of a resistive memory. It can be categorized into two types based on its word line structures [ 6 ]: 3D VRRAM with a word line (WL) planar structure uses metal planes as WL electrodes, while a 3D VRRAM with a WL even/odd structure has comb-shaped WLs separated by etching. This structure is more promising than a WL plane structure for the VRRAM architecture because it has the same performance as a double cell bit [ 7 , 8 ]. Therefore, if a 3D VRRAM is used for synapses instead of a 2D crossbar array, as shown in Figure 1 , the chip area of a DNN system can be effectively reduced. Recently, several works have evaluated the synaptic RRAM using 3D VRRAM. A high-density 3D synaptic architecture based on Ta/TaOx/TiO 2 /Ti RRAM is proposed as a neuromorphic computation hardware and the analog synaptic plasticity is simulated using the physical and compact models [ 9 ]. The potentiality of the VRRAM concept for various neuromorphic applications is investigated with one synapse being emulated by one VRRAM pillar [ 10 ]. Yet many of these studies have focused on experimental demonstration at a single RRAM cell level, and the idea that neuromorphic applications are possible is only presented as a concept. There are some previous studies related to 3D VRRAM with a WL planar structure. For example, the four-layer 3D RRAM integrated with FinFET (Fin Field-Effect Transistor) was developed for brain-inspired computing and in-memory computing [ 11 ], and 3D vertical array of RRAM was proposed for storing and computing large-scale weight matrices in the neural network [ 12 ]. However, a 3D VRRAM with comb-shaped WLs is more promising for a more efficient synaptic RRAM architecture because it has a double cell bit. Although research on 3D VRRAM with comb-shaped WLs has been published, it has focused on RRAM device variation, and explored the concept of many devices connected to one pillar operating as one synapse to overcome the variation [ 13 ]. Implementing a single synapse with multiple devices reduces the benefits of using 3D VRRAM. Moreover, reported previously related studies did not evaluate the circuit level properties of 3D VRRAM with comb-shaped WLs. Theoretical investigations are insufficient for exploring the relationship between synapse weight change and memory device resistance in 3D VRRAM. In this study, we propose a new optimization method for machine learning weight changes that considers the structural characteristics of 3D VRRAM. This study investigates the operating principle of 3D VRRAM synapses with comb-shaped WLs and demonstrates that this structure is a promising synaptic model for neural network systems. The remainder of this paper is organized as follows: Section 2 describes a new 3D VRRAM crossbar array synapse incorporating a synaptic memristor model and learning operations for a guide training algorithm [ 14 , 15 ]. In Section 3 , the accuracy of a neural network with 3D VRRAM synapses is measured by classifying 7 × 7 alphabet letter images using HSPICE circuit simulation. The conclusions are presented in Section 4 .",
"discussion": "4. Discussion In order to determine the appropriate number of training epochs, the learning accuracy was evaluated by varying the number of training epochs from 1 to 300. Figure 11 a shows the accuracy of pattern classification according to the number of training epochs. Only the original image was used in the test, and the accuracy of the pattern classification increases as the number of training epochs increases. The accuracy of the training after 100 epochs, however, is almost unchanged. Thus, we set 100 epochs as the default for neural network training simulation. In order to verify how accurately the pattern classification can be performed even if noise is added to the input image, simulations were performed with an increasing number of inverted pixels as shown in Figure 11 b. Obviously, as the noise increases in the input image, the accuracy of the pattern classification decreases. The simulation results, however, show 80% accuracy until the inverted pixel percentage increases to 12%. This means that 3D VRRAMs are usable as synapses in a neural network system. Therefore, using 3D VRRAM as the synapse structure of a neural network can greatly improve chip area utilization. In this study, we evaluated the accuracy of a neural network consisting only of input and output nodes with no hidden layers. A 3D VRRAM synapse with comb-shaped WLs structured with hidden layers is a subject for future work, and we will demonstrate the effects of 3D VRRAM synapses by performing simulations in a more diverse learning environment."
} | 2,098 |
23145489 | PMC3740780 | pmc | 74 | {
"abstract": "Background The symbiosis between reef-building corals and photosynthetic dinoflagellates ( Symbiodinium ) is an integral part of the coral reef ecosystem, as corals are dependent on Symbiodinium for the majority of their energy needs. However, this partnership is increasingly at risk due to changing climatic conditions. It is thought that functional diversity within Symbiodinium may allow some corals to rapidly adapt to different environments by changing the type of Symbiodinium with which they partner; however, very little is known about the molecular basis of the functional differences among symbiont groups. One group of Symbiodinium that is hypothesized to be important for the future of reefs is clade D, which, in general, seems to provide the coral holobiont (i.e., coral host and associated symbiont community) with elevated thermal tolerance. Using high-throughput sequencing data from field-collected corals we assembled, de novo , draft transcriptomes for Symbiodinium clades C and D. We then explore the functional basis of thermal tolerance in clade D by comparing rates of coding sequence evolution among the four clades of Symbiodinium most commonly found in reef-building corals (A-D). Results We are able to highlight a number of genes and functional categories as candidates for involvement in the increased thermal tolerance of clade D. These include a fatty acid desaturase, molecular chaperones and proteins involved in photosynthesis and the thylakoid membrane. We also demonstrate that clades C and D co-occur within most of the sampled colonies of Acropora hyacinthus , suggesting widespread potential for this coral species to acclimatize to changing thermal conditions via ‘shuffling’ the proportions of these two clades from within their current symbiont communities. Conclusions Transcriptome-wide analysis confirms that the four main Symbiodinium clades found within corals exhibit extensive evolutionary divergence (18.5-27.3% avg. pairwise nucleotide difference). Despite these evolutionary distinctions, many corals appear to host multiple clades simultaneously, which may allow for rapid acclimatization to changing environmental conditions. This study provides a first step toward understanding the molecular basis of functional differences between Symbiodinium clades by highlighting a number of genes with signatures consistent with positive selection along the thermally tolerant clade D lineage.",
"conclusion": "Conclusion Transcriptome-wide analysis confirms the presence of deep evolutionary divisions among the four most common Symbiodinium clades associated with reef-building corals. Nevertheless, phylogenetic tree-based comparisons of relative rates of nucleotide substitutions highlight a number of genes and functional categories that are candidates for the functional differences that have been attributed to clade D Symbiodinium . Top candidates include a fatty acid desaturase, three molecular chaperones and several proteins involved in photosynthesis and the thylakoid membrane. These data provide the first genomic-scale exploration into the adaptive thermotolerance of clade D Symbiodinium . The use of high-throughput sequencing data from field-collected corals also allowed for a detailed examination into the composition of naturally occurring Symbiodinium communities. Our results provide additional support for the prevalence of multiple Symbiodinium clades within individual coral colonies, and therefore the capacity of individual colonies to adjust to changing conditions by ‘shuffling’ the proportions of their resident endosymbionts.",
"discussion": "Discussion Genetic divergence between clades Phylogenetic studies looking at a handful of highly-polymorphic, non-protein coding genetic markers have demonstrated strong evolutionary divergence between Symbiodinium clades, and based on these sequences, divergence times between clades A-D have been estimated at ~30-65 million years (Figure \n 1 )\n[ 10 , 18 , 53 ]. However, little is known about genome-wide levels of divergence, especially at protein-coding genes, which typically experience higher levels of mutational constraint (though see\n[ 59 ]). Our results illustrate substantial genetic divergence throughout most coding genes between all four of the clades most commonly found within Scleractinian corals (A-D), with raw median nucleotide divergences ranging from 18.5-27.3% and Kimura 2-parameter corrected distances\n[ 60 ] of 21.6-34.3% (Figure \n 2 ). Although sequencing and alignment errors may be adding to these high levels of divergence, there is also reason to believe that these divergences may actually be underestimates because we only calculated sequence divergence at loci that were similar enough, at the amino acid level, to be identified as orthologs and similar enough, at the nucleotide level, to fall within our maximum tree length cutoffs. Additionally, these estimates are based only on single nucleotide polymorphisms (SNPs); they do not include insertion and deletion differences. It has been a long debate in the Symbiodinium research community as to what level of divergence these clade designations represent (e.g., species, genus, family); however, accurate characterization of genome wide divergence has been limited by the lack of genomic resources available for these taxa. The primates represent one well-studied system for which several similar estimates of genomic divergence are available. George et al.\n[ 61 ] report exome-wide divergences between Homo sapiens and seven non-human primates spanning three families and seven genera. The average inter-genera divergence is 2.35% and the average inter-family divergence is 2.91%. The lowest level of differentiation estimated in our study, between clades B and C (putatively belonging to the same genus), is more than four times higher than the highest level of differentiation reported by George et al. (4.19% between human and tamarin)\n[ 61 ]. Although the use of taxonomic divisions is not standardized among groups, the high levels of genomic divergence seen among these four Symbiodinium clades certainly suggests that they should be recognized at some higher level of taxonomic division, such as family or order. This is consistent with early phylogenetic work based on rDNA sequences, which demonstrated that within a single locus, genetic diversity within Symbiodinium was equivalent to order-level differences seen in other dinoflagellate groups\n[ 62 ]. The observed relative levels of genomic differentiation are consistent with previously estimated phylogenetic relationships\n[ 18 , 53 ], with clades B and C exhibiting the lowest level of divergence, clade D showing intermediate levels of divergence to B and C, and clade A showing high levels of divergence to all other clades. Assuming a 10–30 day generation time, in hospite [ 63 , 64 ] , the resulting estimate for Symbiodinium substitution rates (~3.1-3.4×10 -9 /site/year) equates to ~3.8-12.5×10 -8 substitutions per site per generation. This is fairly consistent with mutation rate estimates for Caenorhabditis elegans and Drosophila : 2.1×10 -8 and 0.8×10 -8 per site per generation, respectively (reviewed in\n[ 65 ]). Genes with elevated d N /d S along the clade D lineage One of the best-described biochemical differences between relatively heat tolerant and sensitive types of Symbiodinium is related to the saturation state of the thylakoid lipid membranes of the chloroplast. Tchernov et al.\n[ 10 ] demonstrated that several different relatively heat tolerant strains of zooxanthellae (isolates from clades A, B and E) had thylakoid membranes with substantially lower proportions of unsaturated polyunsaturated fatty acids making them more thermally stable and less susceptible to attack by reactive oxygen molecules. No clade D symbionts were included in the analysis of Tchernov et al.\n[ 10 ], but given the generality of their results across relatively thermally tolerant subtypes in clades A, B and E, saturation state of the thylakoid membranes is likely to be important in clade D as well. Our analysis revealed one desaturase involved in unsaturated fatty acid synthesis with significantly elevated d N /d S along the clade D lineage (4-taxa ortholog 515, ω D =0.094, ω BG =0.041; Additional file\n 6 : Table S5). This ortholog is very similar to the palmitoyl-monogalactosyldiacylglycerol delta-7 desaturase (FAD5) in Arabidopsis thaliana (BLASTX evalue=3×10 -104 ), which is involved in early stages of desaturation of fatty acids used in the synthesis of thylakoid membranes\n[ 66 ]. A nonsense mutation in this gene in A. thaliana resulted in a change in the composition of the lipid thylakoid membrane, a reduction in chlorophyll content and delayed recovery of photosystem II after photoinhibition by high light stress\n[ 66 , 67 ]. Therefore, this gene is a strong candidate for involvement in the described functional differences between our clade C and D type Symbiodinium [ 31 ]. No genes involved in fatty acid synthesis exhibited elevated d N /d S along the clade C lineage. In addition to this fatty acid desaturase, seven additional genes related to the chloroplast exhibited elevated d N /d S along the clade D lineage (4 significant after SGoF; Figure \n 4 , Additional file\n 10 : Table S9), including five that compose parts of the thylakoid membrane. Based on protein matches in Swiss-Prot, these orthologs include two subunits of the photosystem I reaction center (subunit II = 4-taxa ortholog 707, ω D =0.118, ω BG =0.021; subunit IV = 4-taxa ortholog 202, ω D =0.153, ω BG =0.032; Additional file\n 6 : Table S5), part of the light-harvesting complex of photosystem II (Fucoxanthin-chlorophyll a-c binding protein F = 4-taxa ortholog 352, ω D =infinite, ω BG =0.044; Additional file\n 6 : Table S5) and a membrane protein involved in the insertion of proteins into the inner membrane of the thylakoid (Inner membrane ALBINO3-like protein 2 = 3-taxa ortholog 58, ω D =0.078, ω BG =0.04; Additional file\n 7 : Table S6). In contrast, we found only one gene involved in photosynthesis with elevated d N /d S along the clade C lineage, peridinin-chlorophyll a-binding protein (3-taxa ortholog 433, ω D =infinite, ω BG =0.046; Additional file\n 9 : Table S8). Therefore, proteins related to the thylakoid and photosynthesis appear to be additional candidates for functional differences between clade C and D Symbiodinium . Several additional functional categories of interest are marginally enriched within the orthologs exhibiting elevated d N /d S along the clade D lineage, including the biological process category ‘protein folding’ (5 orthologs; 4 significant after SGoF) and the related molecular function category ‘unfolded protein binding’ (i.e., chaperones, 3 orthologs), which include matches to heat-shock protein 90 (3-taxa ortholog 1308, ω D =infinite, ω BG =0.0001; Additional file\n 7 : Table S6), prefoldin subunit 3 (4-taxa ortholog 424, ω D =0.099, ω BG =0.028; Additional file\n 6 : Table S5) and chaperone protein DnaJ (4-taxa ortholog 365, ω D =infinite, ω BG =0.018; Additional file\n 6 : Table S5). Chaperones are known to have a diverse set of non-stress related cellular roles, primarily involved in preventing inappropriate interactions among proteins in non-native conformations\n[ 68 ]. However, it is also well documented that chaperone proteins are often upregulated in response to a multitude of stressors, including high temperatures, which can cause protein denaturation\n[ 68 ]. To our knowledge there has not yet been any large-scale gene expression study dedicated to Symbiodinium (reviewed in\n[ 69 ]). Therefore, it is unclear whether these chaperones are also upregulated in response to heat stress in zooxanthellae. However, Leggat et al.\n[ 41 ] identified heat-shock protein 90 as the fourth most abundant transcript in an EST library constructed for Symbiodinium sp. C3 extracted from Acropora aspera . In contrast to these results, no chaperones or protein binding genes were identified with elevated d N /d S along the clade C lineage. Higher rates of amino acid evolution of these proteins in clade D may be related to the higher thermal tolerance of clade D Symbiodinium . It is worth noting that a number of the orthologs exhibiting significantly high d N /d S along the clade D lineage also show significantly low d N /d S along the clade C lineage (Additional file\n 11 : Table S10). This result emphasizes the different patterns of evolution along these two lineages, but it also a good reminder that the analyses of d N /d S along our two focal branches are not independent. Therefore, from the current analysis it is not possible to determine whether the significant results are due to stronger than average positive selection along the D lineage, stronger than average purifying selection along the C lineage or some combination of the two. It is, of course, also possible that rates of change along additional braches of the tree could be influencing the patterns described. Furthermore, temperature tolerance is just one phenotypic difference that exists between clades C and D; therefore, genes involved in temperature tolerance are expected to represent only a subset of the loci highlighted in these analyses. However, the main purpose of this study is to identify candidate genes for the functional differences between clades C and D. Further work is necessary to explore each of these candidates in detail and to elucidate what, if any, role they may have played in the thermal adaptation of clade D Symbiodinium . Additional genes that have been under positive selection along the C and D lineages may have been missed due to high levels of synonymous sequence divergence among clades, which can mask signatures of selection. High levels of sequence divergence appear to be inevitable when comparing Symbiodinium clades in Scleractinian corals (Figure \n 2 ); however, studies have begun to document high levels of functional divergence even within Symbiodinium clades and sub-clade types (e.g.,\n[ 10 , 28 , 34 , 70 ]). For example, type C1 has been shown to be even more thermally tolerant than some clade D symbionts in association with particular coral species\n[ 34 ], and similarly heat-tolerant Symbiodinium types have also been described in clades A and B\n[ 10 ]. Subclade-level genetic comparisons within clade D will be critical for pinpointing the exact evolutionary lineage(s), within clade D, on which selection for increased heat-tolerance has occurred. This type of analysis is also likely to highlight additional genes involved in thermal tolerance of the clade D type evaluated in this study. Similar subclade-level studies in the other major coral-associated Symbiodinium clades will also be important to see if similar mechanisms are involved in heat-tolerance among the different clades. Increased heat-tolerance may have evolved multiple independent times within Symbiodinium , and therefore, heat-tolerant types within the different clades may have evolved different mechanisms for coping with increased temperatures. Mixed assemblages of symbiodinium The majority of genotyping methods for Symbiodinium have a detection threshold for background types of >5-10% of the total symbiont population\n[ 20 , 21 , 71 ]. Quantitative real-time PCR techniques have demonstrated that because of this limitation, traditional methods have greatly underestimated the number of corals hosting multiple clades of Symbiodinium [ 20 ]. Our characterizations of symbiont communities using high-throughput sequencing data provide further evidence for the prevalence of background strains of Symbiodinium at low frequencies. All 16 colonies of Acropora hyacinthus sampled in this study exhibit some proportion of both clades C and D when results are pooled across the two samples from each colony (background frequencies: 0.1-49.4%), despite the fact that PCR-based cp23S genotyping detected only a single clade in each of these colonies. The ubiquitous presence of both clades C and D within A. hyacinthus colonies in Ofu illustrates the strong potential within this population for symbiont ‘shuffling’ (i.e., changing the proportions of different symbionts already coexisting within a coral colony) to play a role in acclimatization of individual colonies in response to environmental fluctuations. It is possible that this level of co-occurrence of clades C and D is unusual for corals. American Samoa has been shown to have a relatively high incidence of both clades C and D\n[ 27 ]. By contrast, type D is less abundant in many other Pacific localities\n[ 27 ]. Co-association with both clades C and D may be less prevalent in situations where one clade is particularly rare. However, Mieog et al.\n[ 20 ] demonstrated similarly high levels of co-occurrence of clades C and D in four species of coral across 11 geographic locations on the Great Barrier Reef. Therefore, high-levels of co-occurrence are clearly not unique to American Samoa. A surprise in our work is the suggestion that different parts of the same colony house different proportions of symbiont clades C and D (Table \n 1 ). Symbiont community compositions are known to systematically vary within colonies of some coral species based on irradiance gradients (e.g., tops vs. sides of colonies in Montastrea spp. in the Caribbean;\n[ 24 ]). However, A. hyacinthus in Ofu, American Samoa occur as flat plates ~20-60 cm across with little obvious depth or light heterogeneity from branch to branch. To date, we can detect no clear environmental gradient within plates that would explain large differences in Symbiodinium community compositions across a colony."
} | 4,434 |
34832720 | PMC8623428 | pmc | 75 | {
"abstract": "With the rapid growth of numerous portable electronics, it is critical to develop high-performance, lightweight, and environmentally sustainable energy generation and power supply systems. The flexible nanogenerators, including piezoelectric nanogenerators (PENG) and triboelectric nanogenerators (TENG), are currently viable candidates for combination with personal devices and wireless sensors to achieve sustained energy for long-term working circumstances due to their great mechanical qualities, superior environmental adaptability, and outstanding energy-harvesting performance. Conductive materials for electrode as the critical component in nanogenerators, have been intensively investigated to optimize their performance and avoid high-cost and time-consuming manufacture processing. Recently, because of their low cost, large-scale production, simple synthesis procedures, and controlled electrical conductivity, conducting polymers (CPs) have been utilized in a wide range of scientific domains. CPs have also become increasingly significant in nanogenerators. In this review, we summarize the recent advances on CP-based PENG and TENG for biomechanical energy harvesting. A thorough overview of recent advancements and development of CP-based nanogenerators with various configurations are presented and prospects of scientific and technological challenges from performance to potential applications are discussed.",
"conclusion": "4. Conclusions and Outlook In conclusion, high-performance, stretchable and uniform nanogenerators are excellent options for human energy conversion. In the past years, due to the huge advancement and development of nanostructured materials, especially conducting polymers, PENG and TENG have been intensively investigated. In this review, the CPs-assisted PENG and TENG with different configurations are discussed and illustrated. The ability of CPs to be solution processed and polymerized in situ contributes to the ease of fabricating their composites in a variety of wearable device forms, such as fibers, nanorods, films, sponges, foams, aerogels, and textiles. The use of composites of conductive polymers to overcome brittleness and processability while keeping electrical conductivity and desirable biological features such as cell adhesion has been investigated. The required electrical conductivity of conductive polymers is routinely sacrificed in order to improve the mechanical characteristics of electrodes. A fundamental understanding of the relation between both the conductive polymer filler and the non-conductive polymer matrix would also result in a synergistic effect in the mechanical and electrical properties of the composites. For minor strain detection and energy harvesting, conformal and nanostructured CP helps boost a stable signal between the electrode and soft skin. However, there are still significant challenges to solve in the practical application of conducting polymer-based nanogenerators, as detailed below. First, it is important to improve the performance of CPs themselves. Controlling the electrical conductivity of pure CPs is difficult. Future study should focus on novel synthetic methodologies and assembly technologies for mass-producing CPs nanoparticles. CPs with great structural order are known to be electrically conductive. It is essential to propose effective strategies for managing the crystalline/amorphous ratio in CP structures. This will be an important study area for CP-based nanomaterial fabrication techniques. Choosing the right CPs and preparing them correctly appears to be the key to successful nanocomposites. This field could benefit greatly from future research on CPs with acceptable sizes, nanostructure, and properties. Simple, efficient, scalable, and economical nanomanufacturing of multi-component nanocomposites is required. Second is to conduct large-scale, green-fabrication methodology. Despite major advancements in flexible nanogenerator technology, there is still much work to be done in terms of device integration and devising large-area, low-cost, clean-room, pollutant-free procedures. There are two parts to this issue that stem from device production and long-term usage: the chemical compatibility and environmental stability of the device, and the degradation of the nanogenerator to reduce e-waste. The key research path in the future will be to produce textile-based nanogenerators with stable and machine-washable performance. It is preferable that all of the materials used are widely acknowledged by the textile industry and that the device is constructed utilizing miniature industrial machinery. Third is to achieve all-in-one integrated electronics with standardized evaluation. Development of multifunctional sensing and energy-harvesting systems that may be used for continuous monitoring and power support of patient′s health status is urgently required. Because of its versatility, an all-in-one electronics system that includes strain/pressure sensors, temperature sensors, a gas sensing device for measuring humidity, electrophysiological (EP) sensors, and energy harvesting devices with wireless signal communication is highly desirable for both daily and clinical applications, as previously stated. In addition, it is crucial for standardized evaluation and application of nanogenerator technologies on various configurations.",
"introduction": "1. Introduction A variety of highly flexible and high-performance electronic devices have been developed for next-generation applications in different fields, including health monitoring, smart communication, flexible displays, energy storage, green electronics, and artificial intelligence systems [ 1 , 2 , 3 ]. The increasing demand for the integration of electronic technology into flexible and wearable devices is driven by the need for continuous, long-term, and individual monitoring at any location and at any time, which is becoming increasingly important. Many different types of wearable and flexible electronics have been proposed to detect physical activity and bio-signals from the human body, including biomechanical signals (bending, pressure, motion, tactile, vibration, heartbeat, breath), temperature, humidity, and electrograms (electromyography, electrocardiography, electro-oculography, and electro-encephalography) [ 4 ]. Our bodies consist of a variety of energies, such as chemical, thermal, and mechanical energy, the latter of which is the most plentiful. In recent decades, wearable energy-harvesting systems and generators have been increasingly significant in recent decades when it comes to self-powered electronic sensors and devices. The research is primarily concerned with the extraction of the greatest amount of useable energy and the conversion of this energy into electric energy. Traditional energy harvesters, on the other hand, continue to encounter difficulties in properly capturing motion, in order to boost the power production while maintaining the system′s stability. However, their improvements must still face numerous hurdles, including how to minimize weight and size, as well as the restricted power supply. As a result, downsizing of flexible, intelligent, and self-powered wearable devices has emerged as a critical research area. To utilize this mechanical energy and address the previously described obstacles, a nanogenerator (NG) that can convert mechanical energy into electric energy was designed in 2006; it has made amazing progress in the last 15 years. There are two types of devices that transform small quantities of mechanical energy into electrical energy via the piezoelectric or triboelectric effects, known as piezoelectric nanogenerators (PENGs) and triboelectric nanogenerators (TENGs) [ 5 ]. Body motions that are observed and used can be divided into two categories: limb motion (i.e., walking, running, jogging, bending, stretching) and subtle motion (i.e., skin tension, heart beats, breathing, blood pressure, phonation) [ 6 ]. Human physical vital signs generate strain and/or pressure changes are collected and converted to electrical signals and energy by nanogenerators, which are significant markers for human energy harvesting. Generally speaking, it may be classified into two categories: (i) energy with mild strain variation, such as heartbeat and blood pressure; (ii) energy with large strain variation, such as variable moving states, finger bending, and facial expression. The major criteria and challenges for high-performance flexible nanogenerators are to: (i) detect output signals sensitively and in a timely manner, (ii) have better sensing capabilities with resilience and recyclability, and (iii) limit potential pollution as green electronics. Traditional wearable electronics are generally made of silicon and carbon-based materials, which have poor biological compatibility and a hard substrate for skin adhesion. Among them, because of their high electrical conductivity, simplicity of synthesis, and innate flexibility, conducting polymers (CPs) have been considered the most suitable materials for flexible electronics [ 7 ]. Polymers are usually considered as electrical insulators before the discover of conducting polymers; yet, these organic polymers exhibit distinctive electrical and optical properties, which are comparable to widely used inorganic semiconductors. The electrical and optical properties of CPs are attributed to the conjugated carbon chains, which are composed of alternating single bonds and double bonds and are determined by highly delocalized, oriented, and electron-dense p bonds. Typical CPs include polyacetylene (PA), polyaniline (PANI), polypyrrole (PPy), poly(3,4-ethylenedioxythiophene):poly(styrenesulfonate) (PEDOT:PSS), poly- thiophene (PTH), poly(para-phenylene) (PPP), poly(phenylenevinylene) (PPV), and polyfuran (PF). The different methods of CPs synthesis and fabrication have been well discussed based on our research experience [ 8 , 9 , 10 ]. Conducting polymers with high flexibility and stretchability, as well as mechanical and electrical responsiveness, have contributed to the recent acceleration of the study on CPs-based in flexible and wearable electronics. Achieving electrical control over a film′s volume and conductivity has been achieved according to the tunable electrical conductivities by doping and de-doping of the polymer. Chemical polymerization, electrochemical polymerization, and photo-induced polymerization are all methods for producing conducting polymers with a variety of nanostructures by controlling the nucleation and growth processes of polymerization. Furthermore, hybridization of CPs with other nanomaterials such as metal, metal oxides, chalcogenides, carbonaceous materials, and metallic oxide resulted in promising functional nanocomposites with improved harvesting performance. Until now, there are several published review papers about conductive polymer nanocomposites and flexible electronics for sensing and biomedical applications. Mokhtar et al. gave a mini review on conducting polymers in wearable devices. Liu et al. reviewed the fabrication and structures of conductive polymer composites for flexible strain sensors [ 11 ]. In addition, there are also some review articles about flexible conductive polymer composites for biomechanical sensors, which not only focus on conducting polymers, but also include other conducting (metal, carbonaceous, metallic oxide) fillers [ 12 , 13 , 14 , 15 ]. Some review articles summarized conducting polymers for sensing [ 7 , 16 ], energy storage [ 17 ], environmental protection [ 18 ], and electromagnetic interference shielding [ 19 ], but did not focus on the biomechanical energy harvesters [ 20 , 21 ]. Furthermore, research also provided summaries for some conducting polymers, such as PEDOT [ 22 , 23 ], PANI [ 24 ] and PPy [ 25 ]. However, at present, the conducting polymers nanogenerators with various structures for wearable energy harvesting have not been systematically summarized. Different to the previous study, this review aims to summarize the recent advances on CP-based PENG and TENG for biomechanical energy harvesting. As shown in Figure 1 , CP-based nanogenerators with various structures and related performances are presented and prospects of challenges to potential applications are discussed."
} | 3,079 |
37801497 | PMC10558124 | pmc | 76 | {
"abstract": "Spiking neural networks (SNNs) aim to realize brain-inspired intelligence on neuromorphic chips with high energy efficiency by introducing neural dynamics and spike properties. As the emerging spiking deep learning paradigm attracts increasing interest, traditional programming frameworks cannot meet the demands of the automatic differentiation, parallel computation acceleration, and high integration of processing neuromorphic datasets and deployment. In this work, we present the SpikingJelly framework to address the aforementioned dilemma. We contribute a full-stack toolkit for preprocessing neuromorphic datasets, building deep SNNs, optimizing their parameters, and deploying SNNs on neuromorphic chips. Compared to existing methods, the training of deep SNNs can be accelerated 11×, and the superior extensibility and flexibility of SpikingJelly enable users to accelerate custom models at low costs through multilevel inheritance and semiautomatic code generation. SpikingJelly paves the way for synthesizing truly energy-efficient SNN-based machine intelligence systems, which will enrich the ecology of neuromorphic computing.",
"introduction": "INTRODUCTION Recently, artificial neural networks (ANNs), such as convolutional neural networks (CNNs) ( 1 ), recurrent neural networks (RNNs) ( 2 ), and transformers ( 3 ), have defeated most other methods and even surpassed the average ability levels of humans in some areas, including image classification ( 1 , 4 , 5 ), object detection ( 6 – 8 ), machine translation ( 3 , 9 – 11 ), speech recognition ( 12 , 13 ), and gaming ( 14 , 15 ). These achievements are computer science–oriented because ANNs are mainly driven by gradient-based numerical optimization methods ( 16 , 17 ), big data ( 18 , 19 ), and massively parallel computing with graphics processing units (GPUs) ( 20 , 21 ). Although neuroscience plays a diminished role in ANNs ( 22 ), insights from neuroscience are critical for building general human-level artificial intelligence (AI) systems ( 23 , 24 ). The human brain is one of the most intelligent systems, having overwhelming advantages over any other artificial system in cognition and learning tasks such as transfer learning and continual learning ( 24 ). The neuroscientific community has been exploring biologically plausible computational paradigms to understand, mimic, and exploit the impressive feats of the human brain. Correspondingly, neuroscience-oriented spiking neural networks (SNNs) have been derived and are sometimes regarded as the next generation of neural networks ( 25 ). SNNs process and transmit information using short electrical pulses or spikes, making them similar to biological neural systems. The neurons in SNNs are lower-level abstractions of biological neurons that collect signals from dendrites and process stimuli with nonlinear neuronal dynamics, which enable SNNs to be competitive candidates for processing spatiotemporal data ( 26 , 27 ). Spike-based biological neural systems are extremely energy efficient, e.g., the human brain only consumes a power budget of approximately 20 W ( 28 ). Benefiting from event-driven computing on spikes, SNNs are up to 1000× more power efficient than ANNs ( 29 ) when running on tailored neuromorphic chips, including True North ( 30 ), Loihi ( 31 ), and Tianjic ( 29 ). The biological plausibility, spatiotemporal information processing capabilities, and event-driven computational paradigm of SNNs have attracted increasing research interest in recent years. Emerging spiking deep learning methods Because of their nondifferentiable spike trigger mechanisms and the complex spatiotemporal propagation processes, designing high-performance learning algorithms for SNNs is challenging. Traditional learning algorithms mainly incorporate biologically plausible unsupervised learning rules and primitive gradient-based supervised learning methods. Unsupervised learning algorithms inspired by the biological nervous system are applied in SNNs, including Hebbian learning ( 32 ), spike timing–dependent plasticity (STDP) ( 33 ), and their variants ( 34 – 38 ). Primitive supervised learning methods including SpikeProp ( 39 ), Tempotron ( 40 ), ReSume ( 41 ), and SPAN ( 42 ) achieve higher performance than biologically plausible unsupervised methods. However, these approaches are limited. Most SpikeProp-based methods only allow spiking neurons to fire no more than one spike, while Tempotron, ReSume, and SPAN cannot train SNNs with more than one layer. Thus, these primitive supervised learning methods can only solve tasks that are no harder than classifying the Modified National Institute of Standards and Technology (MNIST) dataset ( 43 ). One of the key technologies that have led to the rapid progress of ANNs is deep learning ( 44 ), which optimizes the parameters of multilayer ANNs via backpropagation and learns high-dimensional representations of data with multiple levels of abstraction. To overcome the challenge of training SNNs, researchers have explored the application of deep learning methods to SNNs and achieved substantial performance improvements. Two of the most commonly used deep learning methods for SNNs are the surrogate gradient method ( 45 – 47 ) and the ANN-to-SNN conversion (ANN2SNN) ( 48 – 53 ). SNNs trained by the surrogate gradient method achieve high performance on complex datasets such as the Canadian Institute for Advanced Research (CIFAR) dataset ( 54 ), the Dynamic Vision Sensor (DVS) Gesture dataset ( 55 ) and the challenging ImageNet dataset ( 19 ) using only a few simulation time steps ( 56 – 61 ), while SNNs converted from ANNs attain almost the same accuracy as that of the original ANNs on the ImageNet dataset with dozens of simulation time steps ( 51 , 62 , 63 ). Because of the rapid progress achieved by deep learning methods, the applications of SNNs have been expanded beyond classification to other tasks including object detection ( 64 – 66 ), object segmentation ( 67 , 68 ), depth estimation ( 69 ), and optical flow estimation ( 70 ). The boom exhibited by the research community indicates that spiking deep learning has become a promising research hotspot. Demands for frameworks Experience derived from the development of ANNs shows that software frameworks play a vital role in deep learning. Modern frameworks, including TensorFlow ( 71 ), Keras ( 72 ), and PyTorch ( 73 ), provide user-friendly frontend application programming interfaces (APIs) implemented by Python and high-performance backend accelerated by C++ libraries, e.g., Intel MKL and NVIDIA CUDA. Statistics show that the numbers of both new adopters and projects increase exponentially after the release of modern frameworks ( 74 ), indicating that these frameworks substantially reduce the workload required to build and train ANNs, help users quickly realize ideas, and make great contributions to the growth of deep learning research. The rapid development of machine learning frameworks has additionally accelerated the progress of the research community, which also highlights the importance of frameworks for spiking deep learning. However, there is no mature framework available for spiking deep learning. Most existing frameworks for SNNs, including NEURON ( 75 ), NEST ( 76 ), Brian1/2 ( 77 , 78 ), and GENESIS ( 79 ), can build detailed spiking neurons with complex neuronal dynamics, use numerical methods to approximate ordinary differential equations (ODEs), and simulate the biological neural system with high precision but do not integrate automatic differentiation, which is the core component required for gradient-based deep learning. These frameworks construct SNNs that are highly biologically plausible and can be used to investigate the functionality of real neural systems but are not designed to solve machine learning tasks. Nengo ( 80 ), SpykeTorch ( 81 ), and BindsNET ( 82 ) use simplified neurons with smaller numbers of ODEs than detailed neurons. With their low computational complexity brought by simplified neurons, these frameworks can implement some primary machine learning and reinforcement learning algorithms but still lack the modern deep learning capabilities of SNNs. Because of the lack of available frameworks, researchers who want to combine advanced deep learning methods with SNNs have to build basic spiking neurons and synapses from scratch, resulting in repetitive and uncoordinated efforts. Deep SNNs involve a large number of matrix operations in both spatial and temporal dimensions of the data, which requires researchers to refine their codes to create high-performance programs that are accelerated by GPUs. Such workloads increase the burden on researchers. In the neuromorphic community, SNNs are frequently used to process data from neuromorphic sensors and deployed on neuromorphic computing chips, but data processing and deployment also require considerable time and effort. After skillful researchers implement their projects, inconsistent programming languages, coding styles, and model definitions generated by different authors will be difficult to reuse and divide the community. The efficiency of scientific research can be greatly improved if there exists a modern spiking deep learning framework that has at least the following three characteristics: exploits and accelerates spike-based operations; supports both simulations on CPUs/GPUs and deployment on neuromorphic chips; and provides a full-stack toolkit for building, training, and analyzing deep SNNs. SpikingJelly: A modern framework for spiking deep learning To solve the above issues and promote research on spiking deep learning, we present SpikingJelly, an open-source deep learning framework, to bridge deep learning and SNNs. Figure 1A shows a hierarchical overview of the architecture of SpikingJelly. On the basis of one of the most commonly used machine learning frameworks, PyTorch, SpikingJelly supports the simulation of SNNs on both CPUs and GPUs with autograd-enabled computation. Additional CUDA kernels are used for GPU-level acceleration beyond that provided by PyTorch. To achieve a balance between ease of use, flexible extensibility, and high performance, the deep learning aspects in SpikingJelly are elaborated into four sections: Components, Functions, Acceleration, and Networks (see the “Subpackages of deep learning” section in the Supplementary Materials). Components provide essential modules such as spiking neurons and synapses to build deep SNNs. Functions contain practical functions for training, simulating, analyzing, converting, quantizing, and deploying SNNs. Some modules in Functions are the functional formulations of the corresponding modules in Components. In this way, both object-oriented programming and procedural programming are supported to satisfy the diverse demands of users. Acceleration accelerates the SNN simulation process with extra semiautomatically generated CUDA kernels, thus exploiting the efficiency of low-level programming languages and reducing the development cost via the code generation technique. On the basis of the above subpackages, Networks provide classic and large-scale network structures such as Spiking ResNet for fast model reuse and primary SNN applications for beginners, which is favorably received by the community. Considering that neuromorphic datasets obtained from neuromorphic sensors ( 83 – 85 ) such as ATIS, DAVIS, and DVS are widely used in SNNs, SpikingJelly integrates neuromorphic dataset processing methods, including downloading, unifying data layouts, and reading interfaces with the general NumPy ( 86 ) format. With the quantize and exchange packages provided in Functions, compatibility with neuromorphic chips is also effectively implemented by providing low-bit quantizers for network weights and exchange functions for deploying SNNs. Fig. 1. Overview of SpikingJelly. ( A ) Architecture of the whole framework. ( B ) A code example of building and running an SNN, whose architecture is shown in ( C and D ). The execution times required for a single training iteration with T = 2, 4, 8, 16, and 32 and inference with T = 128 using Spiking ResNet-18 built by SpikingJelly, Norse, and snnTorch. ( E ) Comparisons of the ecological niche of SpikingJelly with those of other frameworks. ( F ) Typical adoptions from research based on SpikingJelly. As a full-stack framework, SpikingJelly enables researchers to build SNNs with flexible and convenient APIs, simulate SNNs with extremely high efficiency, and deploy SNNs to edge AI devices. With SpikingJelly, a method for synthesizing a truly spike-based energy-efficient machine intelligence paradigm enriches the ecology of the research in this field.",
"discussion": "DISCUSSION Although SNNs outperform ANNs in terms of biological plausibility and power efficiency, their applications were restricted to neuroscience rather than computational science because of the lack of available learning methods. With the introduction of deep learning methods, the performance of SNNs has been greatly improved, making spiking deep learning a new research hotspot. However, this emerging research area faces the dilemma that the classical software frameworks focus on neuroscience rather than deep learning, while new frameworks have not been developed. SpikingJelly is designed to satisfy the booming research interests of spiking deep learning (see the “Statistical trends” section in the Supplementary Materials). Spiking deep learning is an emerging interdisciplinary field where researchers should be well versed in both neuroscience and deep learning. However, researchers who specialize in one research area may not be familiar with the other, which is consistent with our developers’ experience when answering questions and participating in discussions posed by users in GitHub. To mitigate the learning and the cost of use, SpikingJelly provides brief and convenient APIs. Classic models and frequently used training scripts are also included. With a few lines of code, users can easily build various types of SNNs and run their models even if they are not seasoned developers aware of the underlying implementation. This design philosophy of SpikingJelly frees users from painstaking coding operations when implementing more creative work. Complex and diverse spiking neurons and synapses are the core components of SNNs. Modifying neurons and synapses by mimicking the biological neural system ( 56 , 98 , 141 ) or referring experience from deep ANNs ( 60 , 89 ) is a practicable method for improving spiking deep learning. Researchers want to modify existing modules to define new classes of spiking modules by changing certain functions and properties. Researchers also expect that they should only need to write a few codes to make large changes to the behavior and performance of a model. Such a research paradigm is supported by SpikingJelly’s flexible APIs. Most of the modules in SpikingJelly are created by inheriting from their parents, overriding functions, and adding/deleting new attributes, which also provides a perfect reference for researchers to define new modules. Deep learning typically uses datasets with massive numbers of samples and large-scale models ( 22 ). A larger number of training epochs is also used to achieve better performance ( 159 ). All the above characteristics are computationally intensive and hold for spiking deep learning. Moreover, because of the additional temporal dimension, deep SNNs have a higher level of computational complexity than deep ANNs. Thus, the simulation efficiency of deep SNNs is critical, especially for the recent research progress in terms of evaluating deep SNNs with more than 50 layers on the ImageNet dataset containing 1.28 million images has been a widely used performance baseline ( 58 , 60 , 62 , 88 , 89 ). Computational efficiency is emphasized in the design of SpikingJelly. The simulation process using SpikingJelly benefits from its infrastructure, PyTorch, which enables CPU acceleration with OpenMP/MKL and GPU acceleration with cuBLAS/cuDNN. Advanced acceleration with fusion operations is introduced by merging dimensions, semiautomatically generated CUDA kernels, and just-in-time (JIT) compiling, which brings extremely high training/inference efficiency. Equipped with these acceleration methods, SpikingJelly achieves state-of-the-art simulation speed, which frees researchers from wasting too much time waiting for lengthy deep SNN training processes. Deep learning methods boost the performance of SNNs and make SNNs practical for use in real-life tasks ( 26 ). With the full-stack solution provided by SpikingJelly for building, training, and deploying SNNs, the boundaries of deep SNNs have been extended from toy dataset classification to applications with practical utility, including human-level performance classification, network deployment, and event data processing. Beyond the classic machine learning tasks, several frontier applications of deep SNNs have also been reported, including a spike-based neuromorphic perception system consisting of calibratable artificial sensory neurons ( 160 ), a neuromorphic computing model running on memristors ( 161 ), and the design of an event-driven SNN hardware accelerator ( 115 ). All the above evidence indicates that the advent of SpikingJelly will accelerate the boom of the spiking deep learning community. The future development scheme of SpikingJelly will keep track of the advances in neuromorphic computing, and the rapid symbiotic growth of SpikingJelly and the community will be witnessed."
} | 4,403 |
32620171 | PMC7334858 | pmc | 77 | {
"abstract": "Background Anaerobic digestion (AD) is a globally important technology for effective waste and wastewater management. In AD, microorganisms interact in a complex food web for the production of biogas. Here, acetoclastic methanogens and syntrophic acetate-oxidizing bacteria (SAOB) compete for acetate, a major intermediate in the mineralization of organic matter. Although evidence is emerging that syntrophic acetate oxidation is an important pathway for methane production, knowledge about the SAOB is still very limited. Results A metabolic reconstruction of metagenome-assembled genomes (MAGs) from a thermophilic solid state biowaste digester covered the basic functions of the biogas microbial community. Firmicutes was the most abundant phylum in the metagenome (53%) harboring species that take place in various functions ranging from the hydrolysis of polymers to syntrophic acetate oxidation. The Wood-Ljungdahl pathway for syntrophic acetate oxidation and corresponding genes for energy conservation were identified in a Dethiobacteraceae MAG that is phylogenetically related to known SAOB. 16S rRNA gene amplicon sequencing and enrichment cultivation consistently identified the uncultured Dethiobacteraceae together with Syntrophaceticus , Tepidanaerobacter , and unclassified Clostridia as members of a potential acetate-oxidizing core community in nine full-scare digesters, whereas acetoclastic methanogens were barely detected. Conclusions Results presented here provide new insights into a remarkable anaerobic digestion ecosystem where acetate catabolism is mainly realized by Bacteria . Metagenomics and enrichment cultivation revealed a core community of diverse and novel uncultured acetate-oxidizing bacteria and point to a particular niche for them in dry fermentation of biowaste. Their genomic repertoire suggests metabolic plasticity besides the potential for syntrophic acetate oxidation. Video Abstract",
"conclusion": "Conclusions By combining metagenomics with 16S rRNA gene amplicon sequencing and enrichment cultivation, this study allowed us to link the identified organisms with their potential function in the AD food web. The metabolic reconstruction of metagenome assembled genomes from a full-scale biogas reactor and enrichment cultures identified novel putatively acetate-oxidizing bacteria with high metabolic plasticity. The unique ability to reverse the WL pathway or likely using other pathways for syntrophic acetate oxidation is key for SAOB to compete with the carbohydrate-fermenting majority. Beyond the genomic evidence, syntrophic acetate oxidation was demonstrated by the enrichment of Syntrophaceticus , Tepidanaerobacter , and Clostridia distantly related to Caldicoprobacter. Acetate-dependent catabolic processes in the biogas reactors were largely attributed to bacteria as acetoclastic methanogens were completely outnumbered by the SAOB. Thermophilic dry-fermentation of biowaste possibly provides an exceptional ecosystem to study SAOB and for the enrichment and isolation of novel SAOB. Assuming that the identified SAOB are actually active as acetate-oxidizers, this underpins the importance of these bacteria for process functioning and overall organic carbon mineralization. Further studies applying proteomics or transcriptomics and isotope labeling techniques will be crucial to determine their in situ activity and to quantify syntrophic acetate oxidation versus acetoclastic methanogenesis.",
"discussion": "Results and discussion Genomic potential in thermophilic dry fermentation of biowaste A total of ~ 645 megabases of quality trimmed metagenomic sequencing data was recovered from a thermophilic digester that treats biowaste. Both 16S rRNA gene fragments extracted from the metagenome and total quality trimmed reads were used for taxonomic assignment (Fig. 1 a, Additional file 2 : Supplementary Methods). Classification of reads that could be mapped to a 16S rRNA gene reference database revealed Firmicutes as dominant bacterial phylum (53.2%) followed by Bacteroidetes (11.1%) and Thermotogae (7.7%). All other bacterial phyla together accounted for 11.8% (Fig. 1 a). Archaea contributed 3.1% to the prokaryotic community and was strikingly dominated by the genera Methanoculleus (2.6%) and Methanothermobacter (0.4%) (Fig. 1 a, Additional file 1 : Fig. S2). Assembly and binning allowed the reconstruction of 14 MAGs that showed more than 60% of marker gene completeness although the sequencing depth was rather low, which is indicated by the average assembly coverage of 4.7-fold. MAGs could be recovered from almost all major phyla that were identified from total metagenome reads (Fig. 1 a), and 57% of the metagenome reads could be mapped back to the MAGs. The phyla Atribacteria , Bacteroidetes , Halanaerobiota , Firmicutes , Synergistetes , and Thermotogae were represented in the metagenome by at least one medium-quality MAG (> 50% complete). The majority of MAGs (7 out of 14) affiliated with the dominant phylum Firmicutes . Most of the MAGs contained less than 4% contamination, though the Methanoculleus MAG-R14 showed a duplication of almost all single copy marker genes (Additional file 1 : Table S1). Moreover, the Defluviitoga MAG-R13 ( Thermotogae ) contained 14% contamination from closely related species. Concatenated ribosomal proteins and the taxonomic assignment of the annotated coding sequences (CDS) were used for a more detailed phylogenetic identification of the MAGs (Fig. 1 b). A comparison of the common modules and those that were mostly absent suggests that fermentation rather than respiration account for the majority of carbon mineralization (Fig. 1 b), which agrees with the very low concentration of common electron acceptors typically found in AD systems. Moreover, the most widespread hydrogenases found in the MAGs and most abundant in unbinned metagenome sequences were electron confurcating group A3 [FeFe] hydrogenases (Additional file 1 : Fig. S1) that couple the fermentative evolution of hydrogen from NADH with the oxidation of reduced ferredoxin [ 41 ].\n Fig. 1 Taxonomic assignment of quality trimmed metagenome reads summarized on phylum level ( a ) based on total reads (outer ring) and 16S rRNA gene fragments (inner ring). Phylogeny of metagenome assembled genomes based on concatenated marker protein sequences and selected metabolic functions encoded in their genomes ( b ). Further information about contamination and completeness of the MAGs is provided in Additional file 1 : Table S1. RAxML bootstrap support > 70% is indicated by open circles, and bootstrap support > 90% is indicated by closed circles. Multifurcations were introduced for branches with bootstrap support < 50% AD of organic matter starts with the depolymerization of macromolecular material. Biowaste can be rich in polymers derived from plant material like cellulose. The large fraction of Defluviitoga (MAG-R13) in this system indicates the relevance of these bacteria for the hydrolysis of polysaccharides. Defluviitoga utilize a remarkable variety of complex carbohydrates such as xylan, pullulan, lichenan, galactan, chitin and cellulose, among others [ 42 , 43 ] producing acetate, ethanol, hydrogen, and carbon dioxide. Defluviitoga tunisiensis was also identified as widespread and abundant suggesting a central role in thermophilic biogas reactors [ 44 ]. Among the highest number of intracellular and extracellular carbohydrate-active enzymes (CAZymes) were identified in the Defluviitoga MAG-R13, which further support their importance for the initial breakdown of macromolecular compounds (Fig. 2 ). In addition, genes involved in oligosaccharide and sugar degradation as well as the relevant transporters were prevalent among the bacterial MAGs. This is in coherence with a high functional redundancy that has been reported for the first steps of AD and a gradually more specialized community in the later steps of AD [ 44 – 46 ]. Several complete modules for amino acid metabolism and degradation were detected in the Sphingobacteriales MAG-R11 suggesting uncultured Bacteroidetes as degrader of proteinaceous material besides their capability to utilize carbohydrates (Fig. 2 ). Genes for dissimilatory sulfur metabolism (sat and apr) were only identified in Halanaerobiaceae MAG-R8, but their sulfur-metabolizing capacity remains unclear as closely relates species such as Halocella cellulolytica and Halothermothrix orenii were described as obligate-fermenting bacteria [ 47 , 48 ]. The final step of AD in this system is facilitated by hydrogenotrophic methanogens of the genera Methanoculleus (MAG-R14) and Methanothermobacter (Fig. 1 a, Additional file 1 : Fig. S2) that potentially act as syntrophic partner for the bacteria described below.\n Fig. 2 Number and groups of carbohydrate active enzymes identified in the MAGs. Major functions associated with selected groups are indicated. Glycosyl hydrolase families (GH) primarily include carbohydrate degrading enzymes Metagenomic perspective on syntrophy and microbial interactions Several MAGs were assigned to syntrophic or putatively syntrophic bacteria (Fig. 3 ). Here, we particularly focused on SAOB as acetoclastic methanogens were not detected in the MAGs and unassembled metagenome reads (Fig. 1 ), which suggests that acetate is mainly turned over by bacteria.\n Fig. 3 Putative functions of the MAGs in the anaerobic digestion food web ( a ). The asterisk (MAG-R9) indicates a potential function that merely proposed in literature (see text for details). Wood-Ljungdahl pathway genes in MAGs of syntrophic and putatively syntrophic bacteria as well as the corresponding energy-conserving mechanisms ( b ). Hyd, [FeFe] hydrogenase group A3; Hdr, heterodisulfide reductase; NiFe H2ase, energy-conserving group 1a [NiFe] hydrogenase; Rnf, Na + /H + translocating ferredoxin:NAD + oxidoreductase; Q, quinone; ETF, electron transfer flavoprotein; ETF-Q OR, electron transfer flavoprotein/quinone oxidoreductase; FDH, formate dehydrogenase; Ftr, formate transporter; EM, electron mediator; CM, cytoplasmic membrane; THF, tetrahydrofolate. Genes involved in the Wood-Ljungdahl pathway are acetate kinase (1), phosphotransacetylase (2), acetyl-CoA synthase complex including corrinoid protein (CoFeSP) (3), carbon monoxide dehydrogenase (4), methyltransferase (5), methylene-THF reductase (6), methylene-THF dehydrogenase (7), methenyl-THF cyclohydrolase (8), formyl-THF synthethase (9), and formate dehydrogenase (10) Evidence from biochemistry and transcriptomics revealed that some SAOB use the reversed Wood-Ljungdahl (WL) pathway for acetate oxidation [ 49 – 52 ]. The WL pathway can operate in both the reductive and oxidative direction, and diverse acetogenic bacteria use this pathway to produce acetyl-CoA from CO 2 with various electron donors or use the pathway as electron sink during carbohydrate or organic acid fermentation [ 53 ]. WL pathway genes of SAOB were usually scattered in the genome not organized in an operon structure [ 54 ], and based on gene homology, it is further not possible to predict if the pathway could be reversed for acetate oxidation [ 52 ]. Therefore, it is not possible to identify SAOB from coding information in their genome alone. Phylogenetic analysis of ribosomal proteins revealed Syntrophaceticus MAG-R2 most closely related to the well characterized SAOB Syntrophaceticus schinkii and Thermacetogenium phaeum (Fig. 1 b). Although the MAG showed only 64% marker gene completeness, it contained the WL pathway except a methylene-tetrahydrofolate (methylene-THF) reductase gene that is missing (Fig. 3 b). Moreover, MAG-R2 shared several genes encoding energy conservation mechanisms with S. schinkii and T. phaeum including the respiratory group 1a [NiFe] hydrogenase that is absent in the other known SAOB [ 54 ]. The genomic repertoire for syntrophic acetate oxidation was further identified in Dethiobacteria MAG-R7. The 89% complete MAG only lacked the gene for acetate activation, the acetate kinase (Fig. 3 b). In the oxidative WL pathway, one ATP has to be invested for acetate activation, and substrate level phosphorylation of methyl-THF to formate would yield one ATP. Thus, there is no net ATP gain and it is still unknown how SAOB conserve energy. To circumvent energy-consuming acetate activation by an acetate kinase to acetyl phosphate, an alternative route via acetaldehyde was recently proposed for T. phaeum [ 55 , 56 ]. Dethiobacteria MAG-R7 contained several genes encoding an aldehyde ferredoxin oxidoreductase that could enable this alternative way of acetate activation, or the acetate kinase gene is missed in the assembly. A nearly complete 16S rRNA gene was encoded in MAG-R7 that further confirmed the phylogenetic placement based on ribosomal proteins. Interestingly, MAG-R7 grouped with Ca . Syntrophonatronum acetioxidans, a SAOB important in haloalkaline environments [ 30 , 57 ]. Ca . Contubernalis alkalaceticum, another haloalkaliphilic bacterium that was described as SAOB [ 58 ], and other syntrophic Dethiobacteraceae [ 57 ] were closely related to MAG-R7. Together, the phylogeny and the genomic repertoire for syntrophic acetate oxidation suggest the Dethiobacteria MAG-R7 as a potential SAOB. Besides carbohydrate fermentation, a complete beta-oxidation pathway was found in MAG-R7 suggesting a high metabolic flexibility potentially including fatty acid degradation (Fig. 1 b, Fig. 2 ). Mosbæk et al. [ 59 ] suggested members of the DTU014 clade ( Firmicutes ) as SAOB. Protein-based stable isotope probing combined with metagenomics identified heavily labeled proteins affiliating with the DTU014 in incubations with 13 C-labeled acetate although a complete WL pathway was not identified in the respective genome bins [ 59 ]. Both DTU014 MAGs obtained in this study (R5 and R6) shared > 99% average nucleotide identity with the MAGs reported by Mosbæk et al . [ 59 ] and also lack key genes such as the carbon monoxide dehydrogenase/acetyl-CoA synthase complex including the corrinoid iron sulfur protein (Fig. 3 b). The oxidative WL pathway is so far the only known route used by SAOB to drive acetate oxidation. However, not all isolated and characterized SAOB encode the complete gene set for this pathway as well. The genomes of four isolated SAOB are completely sequenced, but Clostridium ultunense and Pseudothermotoga lettingae lack several genes of the WL pathway indicating that other pathways for syntrophic acetate oxidation could exist. Consequently, acetate-oxidizing DTU014 might use such an alternative route. Low energy yield and slow growth rates of syntrophic SCFA-oxidizing bacteria could be a possible explanation that those organisms were often found in low abundances in AD. The high coverage of DTU014 MAG-R5 and MAG-R6 together with their large fraction of 16S rRNA gene fragments (Additional file 1 : Table S1, Fig. S2) indicate a high abundance of these species, which contradicts the typical proportions of SAOB in AD communities and rather points to a function different than syntrophic acetate consumption in this biogas reactor. Other potential syntrophs identified in the metagenome were Caldatribacterium (MAG-R9) and Syntrophomonas (MAG-R1) species. The putative functions of the poorly characterized candidate phylum Atribacteria are as yet little understood. Atribacteria associated with AD were proposed as syntrophic propionate-oxidizing bacteria based on metagenomics and transcriptomics [ 60 , 61 ]; however, genes encoding the methylmalonyl-CoA pathway for propionate oxidation were largely absent in Caldatribacterium MAG-R9. Saccharolytic fermentation and syntrophic acetate oxidation via a hypothetical pathway using the glycine cleavage system was also proposed [ 61 , 62 ]. There is as yet no experimental evidence for syntrophic acetate oxidation among the Atribacteria , and physiological inferences are almost exclusively based on metagenome data. MAG-R1 affiliated with cultured Syntrophomonas species (Fig. 1 b). Like the isolated relatives, Syntrophomonas MAG-R1 appears to be specialized for the degradation of fatty acids using the beta-oxidation pathway. MAG-R1 and MAG-R9 did not encode a WL pathway and thus likely did not belong to the acetate-degrading community (Fig. 3 b). Potential for syntrophic acetate oxidation in enrichment cultures Mineral medium containing no electron acceptors other than bicarbonate/CO 2 was inoculated from a thermophilic biogas reactor (Additional file 1 : Table S2). Acetate was repeatedly fed to a concentration of 10 mM. The cultures were incubated at 55 °C and 1–10% was regularly transferred into fresh medium. After 117 days of incubation and three successive transfers, the active cultures were selected for amplicon sequencing to identify the acetate-consuming microbial community (Fig. 4 ). The enrichments ENR-ALA, ENR-AEA, and ENR-Ac were inoculated from the same reactor but treated slightly different (see “Methods” section for details). Bacteria known as SAOB such as Syntrophaceticus and Tepidanaerobacter were enriched but also acetoclastic methanogens of the Methanosarcina (Fig. 4 ). In ENR-Ac, Firmicutes classified as Caldicoprobacter and Methanosarcina were enriched whereas known SAOB could not be detected. Active acetate-degrading enrichment cultures with no known SAOB were selected, and cultivation was pursued in the same way (ENR-Ac). The microbial communities were screened again after 211, 244, and 491 days of incubation. Caldicoprobacter -related bacteria were enriched from 2% relative abundance in the inoculum up to 63% and were consistently identified as the dominant microbial group in these cultures although their relative abundance was fluctuating (Fig. 4 ). Even though it is unclear if the detected Methanosarcina performed acetoclastic or hydrogenotrophic methanogenesis, the presence of strictly hydrogenotrophic Methanothermobacter strongly suggest syntrophic acetate oxidation. In addition, uncultured Anaerolineaceae ( Chloroflexi ), Symbiobacterium , and Acetomicrobium were also enriched in the culture ENR-Ac (Fig. 4 ; Additional file 1 : Fig. S3). Uncultured Anaerolineaceae were previously linked to homoacetogenesis in AD using a combined metagenomics and transcriptomics approach [ 63 ]. The high acetate concentrations fed to the cultures and phases of starvation were stress factors for the microbes. Such phases of imbalance might allow the coexistence of homoacetogenic bacteria, SAOB, acetoclastic, and hydrogenotrophic methanogens. Acetomicrobium was identified in all cultures and was often among the most abundant. The partial 16S rRNA sequence of the dominant OTU (~ 400 bp length) from the enrichment cultures was closely related to A. hydrogeniformans (> 99% identity), which was described to ferment simple sugars and amino acids [ 62 ]. We tested all three related species ( A. hydrogeniformans DSM22491, A. thermoterrenum DSM13490, and A. mobile DSM13181) with both Methanoculleus thermophilus DSM2373 and Methanothermobacter thermoautotrophicus DSM1053 in a defined co-culture for syntrophic growth on acetate. Moreover, the Acetomicrobium sp. purified from our enrichment cultures was also tested for growth with both methanogens. Under the applied conditions, no acetate consumption or methane formation was observed for any of the defined co-cultures. Interestingly, Acetomicrobium was also enriched in thermophilic propionate-degrading enrichment cultures inoculated from the same biogas reactor [ 64 ]; however, their role in these syntrophic communities still remains enigmatic.\n Fig. 4 Relative abundance of major bacterial and archaeal groups in acetate-fed enrichment cultures. The microbial community was either analyzed by 16S rRNA gene amplicon sequencing (iTags) or metagenome sequencing (Meta) Enrichment ENR-Ac was further sampled at day 244 for metagenome sequencing. Two largely complete MAGs with less than 2% contamination (Additional file 1 : Table S1) that affiliated with Firmicutes were recovered from the metagenome. Both MAGs did not contain 16S rRNA gene fragments that could be used for phylogenetic identification. Based on ribosomal proteins, MAG-E1 was distantly related to Caldicoprobacter sp., whereas MAG-E2 was related to Calderihabitans maritimus (Fig. 1 b). Amplicon sequencing revealed that the clade assigned to Caldicoprobacter was dominated by two major OTUs at day 491. Together with the coverage information (Additional file 1 : Table S1), this suggest that the MAGs E1 and E2 represented the dominant species of the Caldicoprobacter clade. Firmicutes MAG-E1 that showed 97.6% marker gene completeness lacked central genes of the WL pathway whereas a complete WL pathway could be identified in MAG-E2 (Fig. 3 b). A comparison of their genomic repertoire revealed that the metabolism of MAG-E2 is more versatile (Figs. 1b , 2 , 3 b). Central genes of butyrate and propionate metabolism, carbohydrate fermentation, beta oxidation, and amino acid metabolism were detected. Moreover, a respiratory complex I, genes for a dissimilatory sulfite reductase (DsrAB) together with DsrCJKMOP as well as a truncated denitrification pathway including nitrate reductase, nitrite reductase, and nitric-oxide reductase were encoded in the genome (Additional file 1 : Fig. S4, Fig. S5). This strongly suggests that MAG-E2 can utilize various electron acceptors. The MAG-E2 contained genes encoding a NAD-dependent FDH and a FDH complex (Fig. 3 b) that is likely membrane associated and periplasmatically oriented as indicated by a TAT (twin-arginine translocation) signal peptide as well as transmembrane helices. The structure of this complex is comparable to the membrane-bound FDH of Thermacetogenium phaeum , which is only expressed during syntrophic growth on acetate [ 55 ]. Reverse electron transport could be driven in a similar manner to that of Syntrophomonas wolfei when growing syntrophically on butyrate [ 65 – 67 ]. Electrons may be shuttled by an electron transfer flavoprotein (ETF) to an ETF-quinone oxidoreductase that reduces the quinone pool. The periplasmatically oriented FDH complex could then reoxidize the reduced quinone pool. MAG-E2 further encodes an electron confurcating FeFe-hydrogenase group A3 (Ech) and a heterodisulfide reductase (Hdr) complex that is possibly involved in reverse electron transport (Additional file 1 : Fig. S4). Stable isotope probing further suggested Caldicoprobacter -related Firmicutes as acetate-oxidizing bacteria in thermophilic methanogenic communities fed with 13 C-labeled acetate [ 68 ]. The major fraction of the heavily labeled bacterial DNA was assigned to the Caldicoprobacter -related organisms [ 68 ]. SAOB are fastidious and slow growing bacteria that are difficult to culture. So far, no SAOB were isolated together with hydrogenotrophic methanogens from our enrichment cultures. Thus, additional research is indispensable to confirm syntrophic acetate metabolism among these uncultured bacteria and to unravel their mechanism for syntrophy. Amplicon sequencing revealed stable bacterial and archaeal communities in thermophilic dry-fermentation of biowaste To examine if the identified putatively SAOB were a regular component of the microbial community in thermophilic solid-state digestion of biowaste, we selected nine full-scale anaerobic reactors located in Germany and the Netherlands for 16S rRNA gene amplicon sequencing (iTags) of the V3-V4 region (> 400 bp). The reactor that was used for metagenomic analysis was also included for the amplicon sequencing approach (PFL9). The combination of amplicon sequencing and metagenomics enabled us to link the metabolic insights to a stable core population thus allowing more general implications and a broader perspective on microbial interactions in thermophilic dry-fermentation of biowaste. All sampled biogas reactors treat biowaste in dry fermentation (approximately 30% dry matter content) and were operated between 54 and 55 °C. More than 550,000 quality trimmed amplicon reads were kept for phylogenetic classification (Additional file 1 : Table S3). Archaea usually contributed 1–2% to total prokaryotic iTags (Fig. 5 a), which reflects the importance of low abundant but highly active microorganisms for the functionality of a community. Only in one reactor (PFL6), the archaeal 16S rRNA gene abundance reached 4.9%, and Archaea were also enriched in the separated liquid fraction (PLF1a) of reactor PFL1. The archaeal fraction of total prokaryotes is probably slightly underestimated as the average 16S rRNA gene copy number per genome for Bacteria ( n = 5) is higher than those of the dominant methanogens ( Methanothermobacter , n = 2; Methanoculleus , n = 1). Methanoculleus , Methanothermobacter , and Methanomassiliicoccus were ubiquitous throughout the sampled biogas reactors with the first two genera predominating the archaeal community. Together, Methanoculleus and Methanothermobacter contributed 83.2–98.9% to archaeal iTags (Fig. 5 a). In nearly half of the systems (4 out of 9), acetoclastic methanogens of the genera Methanosaeta and Methanosarcina were not detected using amplicon sequencing (Fig. 5 a). In five reactors, Methanosarcina accounted for 0.1–0.7% of Archaea , which is equivalent to a maximum of 0.008% of the total prokaryotic iTags. This very low relative abundance is contradicting to studies that identified Methanosarcinales as abundant core members of the archaeal community in AD [ 4 , 6 , 7 , 44 ]. More quantitative methods such as fluorescence in situ hybridization will be necessary to estimate exact Bacteria to Archaea ratios; however, the dominance of strictly hydrogenotrophic over acetoclastic methanogens within the archaeal community points to a favorable opportunity for SAOB in these AD systems.\n Fig. 5 Microbial community composition in nine full-scale thermophilic biowaste digesters. The archaeal fraction of total prokaryotes and the relative abundance of archaeal amplicons was revealed by 16S rRNA gene amplicon sequencing ( a ). Percentages of taxa relative to total bacterial amplicons are only shown for major groups that contributed ≥ 1% to all bacterial 16S rRNA sequences in at least one sample ( b ). Bray-Curtis dissimilarity based on OTUs clustered at 97% identity is depicted at the bottom Like for the Archaea, the bacterial community showed a recurring diversity pattern among the biogas reactors (Fig. 5 b). Alpha diversity measures further indicate a low diversity within the digesters (Additional file 1 : Table S4). This is in accordance with the findings that thermophilic systems harbor a lower number of species compared to mesophilic reactors [ 69 , 70 ]. In contrast, the degradation of complex substrates was previously linked to a high diversity whereas a smaller number of species was identified in reactors that degrade simple substrates [ 44 ]. Overall, the microbial community identified by amplicon sequencing was in good coherence with the outcomes of the metagenomic approach for PFL9 (Figs. 1a and 5 ), and the identified MAGs adequately covered the major groups of the microbial community. Firmicutes , Bacteroidetes, and Thermotogae were consistently identified as the dominant clades on phylum level in accordance with previous findings that reported a high abundance of those three phyla in thermophilic and biowaste-treating digesters [ 7 , 44 ]. Together, they accounted for 89.6–95.2% of the bacterial community (Additional file 1 : Fig. S6). Among the Firmicutes , a large fraction of reads affiliated with uncultured bacteria of the MBA03 and the DTU014 clade. Together, they contributed 28–46.5% to all bacterial iTags and make up more than half of the Firmicutes (Fig. 5 b). The MBA03 and DTU014 groups cluster sequences of uncultured bacteria at order to family level. However, the MBA03 clade was represented only by one single OTU that accounted for more than 97% of the reads, and the DTU014 clade comprised two OTUs that contributed 92.3–96.6% of their iTags revealing very low species diversity within these groups. Despite their abundance and the identification as members of the core microbiome in mesophilic and thermophilic AD, their metabolic potential and physiology is still poorly characterized [ 6 , 71 ]. Together with the metagenomic analysis, this suggests that few MBA03 and DTU014 species are central to carbohydrate fermentation throughout these thermophilic AD reactors. As most of the MAGs from the metagenome did not contain 16S rRNA gene fragments, they could not directly linked to the OTUs or clustered groups. Instead, the phylogenetic position based on concatenated ribosomal proteins, the taxonomic classification of all coding sequences, and coverage information were used for the assignment (Fig. 5 ). Unlike the Archaea , potential acetate-consuming bacteria of the genera Syntrophaceticus and Tepidanaerobacter were consistently identified contributing 0.4–3.6% and 0.2–1% to the bacterial community, respectively. Syntrophaceticus schinkii was described as a mesophilic species operating at a temperature range from 25 to 40 °C [ 29 ]; however, thermophilic representatives of the Syntrophaceticus were recently identified as drivers of syntrophic acetate oxidation in acetate-fed chemostats [ 72 ] suggesting a wide temperature range for acetate degradation within this genus. Uncultured Dethiobacteraceae were consistently identified with similar abundances like the other SPOB but exceptionally high in abundance (~ 5%) in PFL9, the reactor used for our metagenome survey. Moreover, iTags affiliating with Caldicoprobacter always contributed more than 1% to total bacterial amplicons (Fig. 5 b). The OTUs identified in the enrichment cultures were also among the dominant Caldicoprobacter OTUs in the biogas reactors. The OTUs only share approximately 91% sequence identity to isolated Caldicoprobacter spp. thus indicating a phylogenetic and physiological differentiation between these bacteria. Taken together, uncultured Dethiobacteraceae and unclassified Clostridia distantly related to Caldicoprobacter might harbor novel putatively syntrophic acetate-consuming bacteria that are widespread in thermophilic dry fermentation of biowaste. The reactors harbor a diverse acetate-consuming community impressively dominated by SAOB over acetoclastic methanogens. Ecological considerations for syntrophic acetate metabolism in thermophilic AD of biowaste Despite the fact that hydrogenotrophic and acetoclastic methanogens generally coexist, it has been shown by radioisotope probing combined with metagenomic sequencing that the acetoclastic pathway for methanogenesis can be dominant while hydrogenotrophic methanogens showed higher abundances [ 73 ]. However, the absence or clear minority of Methanosarcina and other acetoclastic methanogens but ubiquity of diverse potential SAOBs on the other hand strongly suggests that acetate is primarily oxidized by bacteria in the studied biogas reactors. Digesters that treat manure or protein-rich material such as biowaste can have high levels of total ammonia nitrogen (TAN) above 4 g NH 4 + -N/kg (e.g., [ 7 , 45 , 73 ]) and ammonia inhibition is a major issue in AD (reviewed by [ 74 ]). It has been demonstrated that a thermophilic community in AD of biowaste can tolerate significantly more free ammonia compared to a mesophilic community [ 15 ]. This was also shown specifically for the hydrogenotrophic methanogens [ 26 ]. The toxic-free ammonia increases with elevating temperature and pH, and its concentration was suggested as major factor that determines the assemblage of the acetate-consuming bacterial and archaeal community [ 17 , 19 , 22 , 24 , 25 , 44 ]. TAN concentrations in reactor PFL1 and PFL8 have been regularly measured and usually ranged between 1.7 and 3 g NH 4 + -N/kg (Additional file 1 : Fig. S7) and were in the lower range for AD of biowaste (e.g., [7, 45]). The composition of biowaste varies over time thus fluctuations of TAN concentrations are expected. However, the community in reactor PFL1 was very similar when sampled at different time points (PFL1b and PFL1c, Additional file 1 : Table S2). Although it requires additional research to unravel the environmental factors that shape this consistent community, it appears that dry matter content, temperature, and the substrate were more important than the TAN level alone. Interestingly, such adapted community towards syntrophic acetate oxidation might be more tolerant to ammonia fluctuations caused by high protein loadings and consequently more stable with respect to process failure due to ammonia inhibition. In AD, a major fraction of the carbon flow to methane occurs via acetate. A better understanding of SPOB will be important and could provide a basis for the management of microbial AD communities."
} | 8,288 |
38278811 | PMC10817900 | pmc | 79 | {
"abstract": "Superhydrophobic surfaces demonstrate excellent anti-icing performance under static conditions. However, they show a marked decrease in icing time under real flight conditions. Here we develop an anti-icing strategy using ubiquitous wind field to improve the anti-icing efficiency of superhydrophobic surfaces during flight. We find that the icing mass on hierarchical superhydrophobic surface with a microstructure angle of 30° is at least 40% lower than that on the conventional superhydrophobic plate, which is attributed to the combined effects of microdroplet flow upwelling induced by interfacial airflow and microdroplet ejection driven by superhydrophobic characteristic. Meanwhile, the disordered arrangement of water molecules induced by the specific 30° angle also raises the energy barriers required for nucleation, resulting in an inhibition of the nucleation process. This strategy of microdroplet movement manipulation induced by interfacial airflow is expected to break through the anti-icing limitation of conventional superhydrophobic materials in service conditions and can further reduce the risk of icing on the aircraft surface.",
"introduction": "Introduction The unexpected icing phenomena on the aircraft surface is always inevitable when it traverses the supercooled cloud layer, which causes serious energy and security problems 1 – 3 . Therefore, numerous approaches have been proposed to impede the growth of ice layer in order to eliminate safety risks during flight, such as reducing the contact of supercooled droplets on the surface 4 , 5 , delaying the ice nucleation process 6 , inhibiting the expansion of ice crystals and decreasing the adhesion strength between the ice and the surface 7 , 8 . Depending on the considerations of additional energy consumption and equipment, these methods can be classified into active and passive anti-icing technologies 9 . Conventional active anti-icing technology always has various shortcomings including short-term validity, excessive energy consumption and intricate equipment 10 , 11 . Even for the widely utilized electrothermal de-icing method, apart from increasing the energy burden of aircraft during flight, it also introduces undesirable electromagnetic interference on the aircraft surface 12 . Consequently, in light of the objective to reduce droplet-surface contact, the superhydrophobic anti-icing technology with droplet self-ejection effect is identified as a promising passive anti-icing approach 13 . Observations have previously demonstrated that various droplet bouncing behaviors on the superhydrophobic surface can be achieved by adjusting the surface structure and wettability 14 – 16 . Moreover, the contact time between the droplet and the surface can be minified by four times on certain superhydrophobic surfaces 17 . Recently investigations have revealed that numerous air pockets retained on the superhydrophobic surface can diminish the supercooling between droplets and the substrate at low temperatures, resulting in the postponement of ice nucleation 18 , 19 . Furthermore, ice nucleation can be delayed by controlling the surface energy and microstructure under low humidity condition 20 . Particularly, the superhydrophobic surface can even prevent droplets from freezing for 7360 s in a static environment of −10 °C 21 . Although the anti-icing performance of superhydrophobic material has been widely recognized, these materials often begin to freeze within seconds under real flight conditions due to the enhanced heat exchange between the microdroplets and the surface which is induced by high-speed impact of numerous subcooled microdroplets and low-temperature airflow 22 . Meanwhile, a certain subcooled microdroplets may be directly pinned and frozen inside the microstructure on the superhydrophobic surface, leading to an anti-icing efficiency limitation of about 30% without the assistance of external field 23 – 27 . To further enhance the anti-icing property, photothermal materials such as cermet, metal oxide and carbon-based materials are introduced into the superhydrophobic surface. Afterwards, solar radiation can be converted into heat to further impede the nucleation process of supercooled droplets on the superhydrophobic surface 28 , 29 . It is noteworthy that despite the anti-icing efficiency of superhydrophobic surface has been effectively improved through introducing photothermal materials, it is still difficult to ensure the anti-icing property of surface without sufficient sunlight 30 – 38 . Even though the superhydrophobic surface was constructed by phase change materials (PCMs), it was still unable to guarantee the all-weather anti-icing ability due to the limitation of energy storage capacity 39 . Hence, drawing inspiration from the concept of anti-icing strategies aided by external field, the ubiquitous wind field accompanying aircraft during flight may also contribute to improving the anti-icing efficiency of superhydrophobic surface around the clock. Previous research demonstrated the significant influence of airflow velocity and surface wettability on the dynamics of single droplet motion 40 . The supercooled droplets on the superhydrophobic surface can be separated from the surface rapidly under the drive of airflow. Furthermore, the arrangement and distribution of numerous microdroplets (around 20 μm in flight environment) were determined by the airflow model near the wall and the liquid water content per unit 41 . Presently, additional surface resistance is always raised by conventional superhydrophobic surface without aerodynamics-designed microstructures under flow conditions, leading to an increase in overall energy consumption during flight. However, superhydrophobic microstructures with specific aerodynamic shapes are expected to directionally control the motion behavior of microdroplets on the surface without affecting the surface resistance. Furthermore, the contact process between the supercooled droplet and the surface can be effectively abbreviated by disturbing the near-wall flow field in low-temperature environment, giving rise to improving the anti-icing capability of the superhydrophobic surface. In this work, a hierarchical structure surface with both aerodynamic and hydrophobic characteristics is designed and fabricated in order to control the airflow near the surface, and the anti-icing performance is characterized and analyzed in low-temperature and high-velocity inflow environments. Combined with the simulation and anti-icing experiment, it reveals that the aerodynamic microstructures with certain angles can effectively delay ice formation by disrupting the ordering of water molecules and increasing nucleation barriers. Particularly, the microdroplet flow upwelling induced by interfacial airflow and microdroplet ejection driven by the superhydrophobic characteristics together provide a higher capability to prevent ice from accumulating on the superhydrophobic surface. The implementation of this strategy can effectively improve the anti-icing performance without relying on specific components and morphology of superhydrophobic materials, expanding the application scope of superhydrophobic materials in the anti-icing field.",
"discussion": "Discussion On account of the actual flight environment, the microstructures were designed and verified in order to minimize additional energy consumption induced by airflow. Afterwards, the superhydrophobic hierarchical structures were successfully constructed on the surface by electrodeposition without affecting the aerodynamic performance. The icing delay behavior revealed that the superhydrophobic hierarchical structure sample with an angle of 30° exhibited a remarkable icing delay time of 1381 s for larger droplets at 253.15 K, and consistently possessed a better icing delay property at lower temperatures. Moreover, the specific angle of 30°contributed to a slight reduction in microdroplet icing on the superhydrophobic surface. This was due to the fact that the disordered arrangement of water molecules induced by the specific angle imposed higher energy barriers for nucleation, thereby inhibiting the icing behavior. Moreover, the fluid dynamics simulation predicted that the microdroplets were always concentrated above a few microstructures at the front of the microstructure sample and gradually diminished along the flow direction. The subsequent icing wind tunnel test verified that the A-30 sample exhibited superior anti-icing performance, which was only 60% of the icing mass observed on the superhydrophobic plate. Notably, it was difficult for microdroplets to bounce off the plate surface since their movement velocity was greatly decelerated by the low-speed airflow near the wall. Conversely, the microdroplet flow upwelling induced by microstructures and microdroplet ejection driven by superhydrophobic micro-nanostructures together provided a higher capability to reduce the icing likelihood on the superhydrophobic-treated hierarchical surface. The implementation of this strategy may break through the environmental limitations of the current superhydrophobic anti-icing technologies in a certain range, thereby further enhancing the anti-icing efficacy of superhydrophobic on a larger time and space scale."
} | 2,316 |
28775718 | PMC5517491 | pmc | 81 | {
"abstract": "Biofilms are dynamic habitats which constantly evolve in response to environmental fluctuations and thereby constitute remarkable survival strategies for microorganisms. The modulation of biofilm functional properties is largely governed by the active remodeling of their three-dimensional structure and involves an arsenal of microbial self-produced components and interconnected mechanisms. The production of matrix components, the spatial reorganization of ecological interactions, the generation of physiological heterogeneity, the regulation of motility, the production of actives enzymes are for instance some of the processes enabling such spatial organization plasticity. In this contribution, we discussed the foundations of architectural plasticity as an adaptive driver of biofilms through the review of the different microbial strategies involved. Moreover, the possibility to harness such characteristics to sculpt biofilm structure as an attractive approach to control their functional properties, whether beneficial or deleterious, is also discussed.",
"conclusion": "Conclusion Architectural plasticity of biofilm constitutes a central process to actively adapt to stress and to increase productivity and fitness of microbial communities in response to changing environmental conditions. Considering dynamics of biofilm structure is thus required to better understand the emergence of novel functional properties and to decipher the communal mechanisms underlying microbial behavior, from single cell to multicellular community. Although our ability to predict and manage the functional properties and adaptation strategies of these complex dynamic communities is yet limited, the increasing development of predictive modeling approaches and the improvement of integration of experiments and models should, in a near future, enable to better link composition, dynamic organization and function of microbial communities ( Widder et al., 2016 ). Recent technological advances in single-cell analytic methods have led to the generation of quantities of novel interesting data on individual microbial behaviors which still are to be exploited through individual-based modeling approach for instance, to provide insights into self-organized spatial patterns and to construct a realistic vision of biofilm at both the individual and community levels ( Hellweger et al., 2016 ).",
"introduction": "Introduction The traditional perception of microbes as unicellular life forms has deeply changed over the last decades with the collection of scientific evidences showing that microorganisms predominantly live in dense and complex communities known as biofilms. Biofilms are classically defined as aggregates of cells adhering to a surface or interface and often embedded in an extracellular matrix of polymeric substances. They constitute one of the most successful mode of life on Earth ( Flemming et al., 2016 ). They are consequently found in natural, industrial, medical, household environments and, from the human point of view, they can be either beneficial or detrimental. Indeed, microbial biofilms are involved in essential nutrient cycling or biotechnological processes as well as in severe chronic infections and biodeterioration phenomenon (for instance Beech and Sunner, 2004 ; Bjarnsholt, 2013 ; Berlanga and Guerrero, 2016 ). Positive or negative impacts directly result from the ability of microorganisms to express specific functions in these complex communities compared to the single planktonic state. The higher resistance of biofilm cells to antimicrobials compared to that of their planktonic counterparts is a telling example of such specific functional properties and should be relied to the structural characteristics of the community ( Bridier et al., 2011 ). Indeed, both the microbial growth and the production of matrix lead to the rise of a biological edifice offering progressively a protective structure to inhabitants able to hinder penetration and action of antimicrobials. The development of three-dimensional biofilm structure also generates physicochemical gradients and physiological heterogeneity with slow growth resistant phenotypes for instance ( Stewart and Franklin, 2008 ). Recently, Berleman et al. (2016) demonstrated the central role of multicellular bacterial community structure in the colonization of surface by Myxococcus xanthus . Indeed, the authors showed that extracellular polymeric substances (EPS) synthesis led to the creation of microchannels which govern both bacterial motility and cell-to-cell interactions and finally organize multicellular behavior during swarm migration. In contrast, a mutant lacking EPS showed a deficiency of cell orientation and poor colony migration. As biofilms are mostly complex associations of strains and/or species in our environments, spatial arrangement of genotypes within biofilms also governs strain interactions and the evolution of social phenotypes as immediate neighbors in the structure are more affected by the social behaviors ( Nadell et al., 2016 ). Spatial organization of genotypes and social interactions will thus govern the whole community architecture and functions ( Liu et al., 2016 ). Functional properties of a biofilm therefore emerge from the construction and shaping of the microbial structure like many of the emergent properties of natural communities relying on the creation of biogenic structures by habitat-forming organisms ( Flemming et al., 2016 ). The close relationships between the architecture of a biofilm and its functional properties emphasizes the need to better describe and understand cell behavior, from single cell to multicellular scale, during biofilm structure development and maturation. Recent technological advances in methodologies including imaging and microscopy, molecular techniques, and physico-chemical assays, enabled the development of novel approaches dedicated to biofilm studies ( Azeredo et al., 2017 ). The possibility to observe biofilm using high resolution and non-destructive methods now allows investigating the dynamics of multicellular structure development and the fate of each of its individual cellular components in parallel. For instance, the key architectural transitions and associated biophysical and genetic mechanisms supporting the developmental program of Vibrio cholerae biofilms have been recently disclosed using single-cell live imaging ( Drescher et al., 2016 ; Yan et al., 2016 ). This kind of observations has clearly improved our understanding of spatio-temporal development of biofilms and has finally increasingly supported the intimate connection between structural modulations and the emergence of functional features and survival strategies. Indeed, the ability of biofilms to adapt their structure in response to internal or external stimuli, called hereafter the architectural plasticity, appears as a key factor affecting the fitness of individuals within the whole microbial community. Interestingly, the role of plasticity in bacterial survival was already demonstrated at the cellular scale. Bacteria are able to alter their morphology and to produce specific morphotypes conferring survival advantages in hostile environments. This was showed for numbers of bacterial pathogens for which filamentation is essential in the resistance to phagocytosis and overall for persistence during infection ( Justice et al., 2008 ; Justice et al., 2014 ). In this review, we will discuss the central role of architectural plasticity in the emergence of functional properties of biofilms and as a communal bacterial response to many harsh conditions and external attacks. We will also deal with the various mechanisms developed by microorganisms to build and modify the three-dimensional community and, with the existing strategies for humans to sculpt biofilm architecture in order to control their function."
} | 1,968 |
39813343 | PMC11734718 | pmc | 82 | {
"abstract": "Coral persistence in the Anthropocene depends on interactions among holobiont partners (coral animals and microbial symbionts) and their environment. Cryptic coral lineages—genetically distinct yet morphologically similar groups—are critically important as they often exhibit functional diversity relevant to thermal tolerance. In addition, environmental parameters such as thermal variability may promote tolerance, but how variability interacts with holobiont partners to shape responses to thermal challenge remains unclear. Here, we identified three cryptic lineages of Siderastrea siderea in Bocas del Toro, Panamá that differ in distributions across inshore and offshore reefs, microbial associations, phenotypic traits of holobiont partners (i.e., phenomes), and skeleton morphologies. A thermal variability experiment failed to increase thermal tolerance, but subsequent thermal challenge and recovery revealed that one lineage maintained elevated energetic reserves, photochemical efficiency, and growth. Last, coral cores highlighted that this lineage also exhibited greater growth historically. Functional variation among cryptic lineages highlights their importance in predicting coral reef responses to climate change.",
"introduction": "INTRODUCTION Climate change is altering environments at unprecedented rates, resulting in warmer and increasingly variable conditions with more extreme events ( 1 ). An organism’s response to such rapid changes [e.g., through shifts in thermal limits; ( 2 )] is influenced by their genetic background, environmental history, and interactions between these two forces [GxE; ( 2 , 3 )]. Understanding and predicting the relative importance of these factors on fitness are fundamental as environments continue to change, species ranges shift, and localized extinctions occur ( 1 , 2 ). Coral reefs represent one of the most productive and economically valuable ecosystems ( 4 , 5 ) threatened by global (e.g., warming and acidification) and local (e.g., nutrient pollution and overfishing) stressors ( 6 ). These stressors have increased the frequency and severity of coral bleaching—loss of the coral’s obligate symbiotic algae ( 7 )—which is projected to worsen under current emissions trajectories ( 8 ). However, reef environments are not changing homogeneously, and although our understanding of which reefs and species are more bleaching resistant is advancing ( 9 ), predicting their future remains challenging due to complexities governing coral resilience, including environmental variation, host genetics, and associations with diverse algal and bacterial symbionts ( 10 ). Coral “holobionts” encompass complex symbioses among coral hosts, algal symbionts (Symbiodiniaceae), and a diverse array of microorganisms, all interacting to shape aggregated holobiont phenotypes (i.e., phenomes). Each member of the holobiont contributes to coral bleaching heterogeneity, including genetic variation of the host ( 11 ), algal symbiont communities ( 12 ), and bacterial communities ( 13 ). For coral hosts, genomic sequencing has revealed an unexpected level of cryptic diversity, including cryptic lineages [i.e., distinct genetic clusters previously characterized as one species; ( 14 , 15 )]. These lineages can also interact with a diversity of holobiont members to produce functionally distinct phenotypes. For example, a lineage in the Acropora hyacinthus species complex more frequently hosts Durusdinium algae and exhibits higher thermal tolerance, although it coexists with other lineages on the same reef ( 16 ). Together, interactions among these cryptic lineages and holobiont members likely play a role in determining bleaching outcomes. In addition to holobiont genetic variation, environmental heterogeneity can influence coral bleaching patterns ( 17 ) and a growing body of literature links coral thermotolerance to high frequency temperature variability, also termed diel thermal variability (DTV) [e.g., ( 18 , 19 )]. DTV is theorized to “prime” [i.e., beneficial acclimation hypothesis; ( 20 )] organisms to more effectively respond to and recover from heat stress ( 21 – 23 ). However, DTV is correlated with other environmental parameters that can also produce beneficial acclimatory effects [e.g., light and flow; ( 24 , 25 )], and it remains unclear whether this variability facilitates thermotolerance via priming and/or whether environmental selection in high DTV environments is selecting for more thermally tolerant individuals. For example, cryptic coral lineages are known to exhibit divergent spatial distributions across depths ( 26 ), which likely involves adaptations in photosynthetic pigment concentrations and skeletal traits that can have important effects on light-harvesting potential ( 27 , 28 ). In addition, lineages and their algal symbionts ( 29 ) have been shown to segregate across inshore-offshore gradients, where offshore habitats experience lower turbidity, higher flow, and more stable temperatures relative to inshore habitats [e.g., ( 30 )]. Together, these patterns of complex environmental heterogeneity likely produce adaptive phenotypes through a combination of acclimation and selection for unique holobiont combinations. To investigate the influence of coral holobiont diversity and environmental history on coral phenomes, we characterized holobiont genetic diversity of the reef-building coral Siderastrea siderea from three inshore and three offshore sites in the Bocas del Toro reef complex (BTRC), Panamá. We found three cryptic lineages with unique symbiotic associations that differed in their distributions across the seascape, as well as distinct baseline phenomes and skeletal morphology. Next, we conducted a 50-day DTV experiment, followed by thermal challenge and recovery, to test the hypothesis that exposure to DTV would increase coral resistance to thermal challenge. In contrast to our hypothesis, we found that cryptic lineages differed in their response to thermal challenge, whereas the effect of experimental DTV treatment was minimal beyond promoting growth. Lineages also differed in their growth, especially under thermal challenge. To determine whether these growth differences between lineages were conserved in situ, we used coral cores to contrast long-term growth trajectories between lineages. We found that these records were consistent with experimental patterns, with lineages differing in skeletal density and linear extension rates. Together, these data showcase the strong influence that cryptic lineages have in shaping coral distributions, symbioses, thermotolerance, and growth. Failure to appreciate this genetic diversity will ultimately lead to challenges when predicting coral responses to climate change.",
"discussion": "DISCUSSION Understanding how holobiont diversity is partitioned across the seascape is critical to predicting coral responses to climate change ( 11 , 12 ). Here, we identified three cryptic host lineages (L1, L2, and L3) in a S. siderea species complex across the BTRC that varied in baseline phenotypes including algal and bacterial symbiont communities, energetic reserves, and skeletal morphology relevant to their in situ light environments. In addition, L1 corals maintained elevated energetic reserves throughout a mesocosm experiment, as well as elevated photochemical efficiency and growth throughout subsequent thermal challenge and recovery. Last, coral core data sampled from lineages across the BTRC demonstrated that these distinct growth patterns observed between lineages under experimental conditions were consistent with patterns of historical growth. This work builds on the growing evidence for widespread cryptic diversity in corals ( 15 ), which has been detected at large [e.g., archipelago-wide; ( 33 )] and small spatial scales, including within reefs in Puerto Rico ( 26 ), the Florida Keys ( 34 ), and American Samoa ( 16 ). Depth has emerged as a common driver of lineage differentiation in corals, with relevant abiotic factors including temperature, seawater optical properties ( 26 , 34 ), and small-scale current patterns ( 35 ). Cryptic lineages have also been shown to exhibit divergence in ecologically relevant traits. For example, sympatric lineages of the Pachyseris speciosa species complex differ in skeletal morphology and holobiont physiology ( 27 ). In addition, lineages of Porites around Kiritimati exhibited differential mortality rates in response to a marine heat wave ( 36 ). Across larger spatial scales, only one of three A. hyacinthus cryptic lineages was able to occupy habitats along a range expansion front in Japan, which was attributed to divergence at loci associated with adaptation to temperate, seasonally fluctuating environments ( 33 ). Thermal variability has also been associated with differential distributions of A. hyacinthus cryptic lineages that exhibit distinct thermal tolerances in American Samoa ( 16 , 37 ). Here, we build on the emerging understanding of cryptic coral diversity and provide an in-depth analysis of functional variation among lineages. Specifically, we find that cryptic lineages vary in their distributions across sites that differ in thermal variability, exhibit distinct and ecologically relevant phenomes (i.e., energetic reserves, skeletal morphologies, and growth) including responses to thermal challenge, and differ in patterns of historical growth. We find evidence for environmental structuring of lineages and that unique skeleton morphologies could contribute to their success in these distinct environments. The cryptic host lineages identified here were structured across an inshore to offshore gradient in the BTRC, with L2 and L3 more prevalent inside Bahia Almirante (inshore), and L1 more prevalent outside the bay (offshore). Inshore BTRC sites are characterized by limited influence from the open ocean, riverine inputs that deliver nutrients, agricultural runoff and sewage to the bay, higher turbidity, and, most recently, hypoxic events that have altered coral communities ( 38 , 39 ). We find evidence that lineages exhibit unique skeleton morphologies that could contribute to their success in these distinct environments. Namely, L2 skeleton morphology suggests that this lineage is better adapted to the low light environments of inshore BTRC to promote algal photosynthesis. The ability of L2 skeletons to better amplify incoming light could also explain its lower symbiont densities and chlorophyll a concentrations compared to L1, which would serve as a mechanism to reduce pigment packing while “sensing more light” ( 28 , 40 ). These skeleton morphology differences persisted even when lineages coexisted in the same environment (CI), which suggests a genetic basis for this trait. Future explorations of this system would benefit from collecting precise colony depths in addition to diffuse attenuation coefficients for downwelling irradiance ( K d ) from each site to link in situ light environments with cryptic lineage assignment. This information, along with additional optical traits of the coral holobiont, particularly metrics taken with intact tissue, would provide more detailed information on how seawater optical properties drive lineage distributions in the BTRC [e.g., ( 41 )]. These cryptic lineages not only occupied distinct habitats, but they also exhibited unique symbioses. L1 and L2 hosted distinct communities of bacterial symbionts at baseline; however, these differences did not persist after a 50-day DTV treatment. Instead, bacterial communities of corals exposed to variability were distinct from corals under control conditions. Although previous work has highlighted that bacterial communities can contribute to coral thermal tolerance [e.g., ( 13 )], this does not seem to be the main driver here given that we observed thermal tolerance differences among corals with similar bacterial communities. In contrast, algal communities were strongly structured by lineage, with L1 associating with higher proportions of D. trenchii and hosting a unique Cladocopium (C3) DIV. Increased D. trenchii in L1 could contribute to their elevated Fv/Fm throughout thermal challenge relative to L2 as D. trenchii has been shown to confer thermal tolerance to some hosts [e.g., ( 12 , 42 )]. Rose et al. ( 16 ) also demonstrated that a more bleaching-resistant A. hyacinthus cryptic lineage hosted greater proportions of D. trenchii . Here, although some L2 colonies were dominated by D. trenchii , this only occurred in low frequencies at sites where both L1 and L2 were sampled (CI and BS). L2 appeared to show the greatest thermal resistance when it associated with C1; however, more statistically robust explorations of this pattern were limited by sample size. Additional sampling will be needed to better characterize the importance of these less common cryptic host lineage–symbiont associations on holobiont phenomes, growth, and thermal responses. Regardless, we posit that the elevated LEF of L2 skeletons coupled with their acclimation to lower light environments of inshore BTRC and lower symbiont densities could have led to reduced Fv/Fm during thermal challenge. Future work investigating the impact of acclimation to different light levels on heat tolerance in these lineages is warranted. In addition, as coral bacterial communities are strongly influenced by their host traits and environment ( 43 ), more thorough sampling across sites, cryptic lineages, and regions within a coral colony [e.g., ( 44 )] will be necessary to better identify any lineage-specific bacterial taxa. We initially hypothesized that DTV would shape coral phenomes to increase thermal resilience [e.g., ( 18 , 21 , 23 , 45 )]. Although an effect of experimental DTV on holobiont phenomes was observed, it was largely driven by differences in growth across treatments. DTV increased growth, suggesting that it represents a promising coral restoration tool to improve growth in nursery settings [as in ( 45 )]. However, experimental exposure to DTV did not facilitate the maintenance of Fv/Fm during subsequent thermal challenge. It is possible that the DTV treatment used here was insufficient to prime corals [reviewed in ( 46 )]. Unique DTV temperature regimes and timing of exposure have previously resulted in variable phenotypic outcomes for the corals Montipora capitata ( 21 ) and Acropora aspera ( 47 ). In M. capitata , only two of four variability pre-exposure profiles altered gene expression and resulted in improved thermal tolerance in stress tests 4 months later ( 21 ). In addition, in A. aspera , exposure to DTV in the month preceding a heat challenge had a larger, and ultimately deleterious, effect on heat tolerance compared to corals that experienced 1.5 years of preconditioning to variable temperatures ( 47 ). Previous work has also demonstrated that DTV can have negative effects when additional heat stress is present [e.g., ( 48 )], and the BTRC reached ~7° heating weeks in the year prior to coral collection (2015; NOAA Coral Reef Watch v3.1), which could have influenced responses to DTV. Elevated growth of L1 compared to L2 was evident not only during the mesocosm DTV experiment but also in coral cores from across the BTRC belonging to the same cryptic lineages. Century-scale data from cores of long-lived corals such as S. siderea allow investigation into the impact of long-term environmental conditions on coral growth and, ultimately, health. Here, cores sampled across inshore and offshore environments of the BTRC show that L1 maintained greater linear extension, lower skeletal density, and trended toward greater overall calcification compared to L2. Previous coring studies on S. siderea demonstrate that reef environments play an important role in shaping long-term growth trajectories. Specifically, in S. siderea sampled from the southern Mesoamerican Barrier Reef System, fore-reef corals exhibited long-term declines in skeletal extension rates whereas nearshore and back-reef coral extension rates were stable ( 49 , 50 ). The authors propose that resilience of nearshore and back-reef corals is linked to their exposure to greater diurnal and seasonal thermal variability ( 50 ). In contrast, more widespread sampling of S. siderea cores across the entire Mesoamerican Barrier Reef System revealed declines in skeletal extension rates only for nearshore corals, which was attributed to exposure to land-based anthropogenic stressors and ocean warming ( 51 ). Although it is possible that the lineage differences in historical growth found here are driven by reef environments, because lineage and environment were almost fully confounded in this coring sampling design, it is also possible that previous works on coral growth trajectories are complicated by the presence of S. siderea cryptic diversity across the Mesoamerican Barrier Reef System. Future work sampling cores from coexisting lineages would disentangle the role of environmental and genetic factors in determining long-term growth trajectories. A more thorough characterization of general environmental conditions at these sites (e.g., nutrient concentrations, pH, and dissolved oxygen) is also needed as sites where L1 and L2 are sympatric suggest other environmental characteristics could be driving differentiation and distributions of S. siderea cryptic lineages. S. siderea is a horizontally transmitting, gonochoric broadcast spawning coral, with colonies of separate sexes spawning gametes to produce aposymbiotic larvae that spend time in the water column before settling, leading to the potential for broad population connectivity across great distances [up to 1200 km; ( 52 )]. Although much more work is warranted, we propose that, in the BTRC, the distinct light environments across inshore and offshore reefs along with physical characteristics of the archipelago ( 39 , 53 ) result in spatially varying selection on cryptic lineages that are uniquely adapted to distinct light environments. Although few sites were found to host multiple lineages and no site hosted all three, sampling for this study was limited to <8 m and exact depths of corals were not recorded. Therefore, we hypothesize that additional sampling across depth within individual sites will reveal differential depth distributions of the lineages that reflect patterns observed across inshore and offshore environments (i.e., L1 associated with higher light and L2 associated with lower light). As was previously demonstrated by Quigley et al. ( 54 ), it is also likely that environmental pools of algae are much more diverse than communities hosted by adult corals, and therefore once recruits begin establishing symbiosis, algal symbionts likely compete through a “winnowing” process with dominance depending on local environmental conditions (i.e., light and depth) that are further shaped by coral colony and skeleton morphology ( 28 ). Surviving recruits of distinct lineages then develop associations with specific Symbiodiniaceae in environments that differ in temperature, light, and nutrients, likely resulting in further acclimation to local conditions. Together, these genetic and environmental factors interact to determine the patterns of responses observed here, where unique holobiont partners shape variation in phenomes, response to thermal challenge, and historical growth. Ultimately, reciprocal transplant experiments are needed to disentangle the relative roles of adaptation and acclimation in the observed phenotypes between lineages. In addition, whole-genome sequencing of the cryptic lineages in this system represents an important future goal, which will help uncover their evolutionary history and potentially identify the genomic basis of their distinct holobiont phenomes. This work highlights the importance of understanding cryptic coral diversity when determining species responses to future climate change and in restoration planning."
} | 5,001 |
29238288 | null | s2 | 83 | {
"abstract": "Printable and flexible electronics attract sustained attention for their low cost, easy scale up, and potential application in wearable and implantable sensors. However, they are susceptible to scratching, rupture, or other damage from bending or stretching due to their \"soft\" nature compared to their rigid counterparts (Si-based electronics), leading to loss of functionality. Self-healing capability is highly desirable for these \"soft\" electronic devices. Here, a versatile self-healing polymer blend dielectric is developed with no added salts and it is integrated into organic field transistors (OFETs) as a gate insulator material. This polymer blend exhibits an unusually high thin film capacitance (1400 nF cm "
} | 180 |
39812995 | PMC11735819 | pmc | 84 | {
"abstract": "Background Large-scale coral bleaching events have become increasingly frequent in recent years. This process occurs when corals are exposed to high temperatures and intense light stress, leading to an overproduction of reactive oxygen species (ROS) by their endosymbiotic dinoflagellates. The ROS buildup prompts corals to expel these symbiotic microalgae, resulting in the corals’ discoloration. Reducing ROS production and enhancing detoxification processes in these microalgae are crucial to prevent the collapse of coral reef ecosystems. However, research into the cell physiology and genetics of coral symbiotic dinoflagellates has been hindered by challenges associated with cloning these microalgae. Results A procedure for cloning coral symbiotic dinoflagellates was developed in this study. Several species of coral symbionts were successfully cloned, with two of them further characterized. Experiments with the two species isolated from Turbinaria sp. showed that damage from light intensity at 340 μmol photons/m 2 /s was more severe than from high temperature at 36 °C. Additionally, preincubation in high salinity conditions activated their endogenous tolerance to bleaching stress. Pretreatment at 50 ppt salinity reduced the percentage of cells stained for ROS by 59% and 64% in the two species under bleaching stress compared to those incubated at 30 ppt. Furthermore, their Fv’/Fm’ during the recovery period showed a significant improvement compared to the controls. Conclusions These findings suggest that intense light plays a more important role than high temperatures in coral bleaching by enhancing ROS generation in the symbiotic dinoflagellates. The findings also suggest the genomes of coral symbiotic dinoflagellates have undergone evolutionary processes to develop mechanisms, regulated by gene expression, to mitigate damages caused by high temperature and high light stress. Understanding this gene expression regulation could contribute to strengthening corals’ resilience against the impact of global climate change. Supplementary Information The online version contains supplementary material available at 10.1186/s40529-025-00451-5.",
"discussion": "Discussion Frequent high temperatures, often coupled with intense light, in surface seawater pose significant threats to coral reef ecosystems worldwide. As these ecosystems support approximately 25% of all marine species, large-scale coral bleaching has the potential to drive many of these species toward extinction, underscoring the importance of preventing coral bleaching to ensure sustainable reef protection and development. A primary cause of bleaching under high temperature and light conditions is the overproduction of reactive oxygen species (ROS), such as superoxide, hydrogen peroxide, and hydroxyl radicals, within the photosynthetic processes of symbiotic dinoflagellates in coral hosts. Reducing ROS production and enhancing ROS scavenging in symbionts are therefore essential strategies for preventing coral bleaching. Among the various approaches to cope with ROS, an effective method involves mutagenesis combined with artificial selection to develop stress-tolerant coral symbionts. The coral symbiont cloning method presented in this study offers a practical approach for coral research groups across different regions to participate in mutagenesis-artificial selection studies. The selected strains with enhanced thermal and high light tolerance could potentially be reintroduced to coral hosts, expediting natural selection, which is typically a very slow process. Additionally, we found that high salinity can increase thermal and light tolerance in coral symbionts. This finding suggests that coral symbiotic dinoflagellates may have evolved genetic mechanisms that confer tolerance to high temperature and high light stress. Understanding these tolerance-enhancing mechanisms could provide valuable insights for preventing coral bleaching. This intrinsic stress tolerance is reminiscent of that observed in some resilient corals in highly saline, warm waters, such as those in the Red Sea. Future studies should investigate whether the resilience of Red Sea corals is a result of high salinity. Key technical aspects of the symbiotic dinoflagellate cloning procedure are detailed below, as this is a novel protocol for coral symbiont cloning and no alternative reference available. Disrupting coral tissues to release their symbionts produced substantial amounts of organic debris and macromolecules. If not thoroughly removed, these organic materials could serve as nutrients for bacterial and fungal growth, potentially harming the symbionts. To remove larger organic matter, filtration was performed using Kimwipes paper on top of a 25 μm mesh nylon screen. To remove micro-debris and macromolecules, since their precipitation rates are slower than that of the symbionts, the supernatant containing these fine particles was removed by suction after most of the symbionts had settled at the bottom of the collection tube. Subsequently, the symbionts were diluted to reduce cell density to fewer than one eukaryotic cell per 200 µL, further decreasing the levels of organic matter in the suspension. This dilution also eliminated protozoan grazing, as no more than one cell from the suspension was transferred to each well. Recognizing that the symbionts were freshly released from their host cells and faced a major environmental change, we minimized potential stress by performing the cloning under low light and placing the symbionts in 96-well plates in the dark for 1 week to allow gradual adaptation to their new environment. During this dark incubation period, glycine was added to the medium as a source of nitrogen and carbon. This incubation in glycine-containing medium under dark conditions was essential for successful cell cloning. After 1 week in the dark, the cells were exposed to 20 µmol photons/m 2 /s light intensity (14/10-h day/night cycle) at 25 °C. It took 2–3 months for a single cell to proliferate into a visible population. Prior to the formation of visible colonies, the cells were observed under an inverted microscope, with minimal light exposure and short observation times to prevent stress. Methods for cloning coral symbionts have been presented in previous studies (Beltran et al. 2021 ; LaJeunesse and Parkinson 2012 ). In comparison, the procedure developed in this study is both simpler and more effective. This method has been successfully used to clone symbiotic dinoflagellates from Turbinaria sp., Acropora sp., Stylocoeniella sp., and Leptoria sp. in two independent laboratories. The observation that elevated salinity enhances the tolerance of two coral symbiont species to high temperature and intense light is consistent with the concept of cross-tolerance observed in plants. In this phenomenon, stress tolerance induced by one environmental factor confers protection against other stressors. For instance, drought-induced tolerance has been shown to protect plants from low-temperature stress, and vice versa (Pastori and Foyer 2002 ; Zhu 2000 ). Although the mechanisms underlying salinity-induced tolerance in coral symbionts remain poorly understood, they are likely linked to the upregulation of protective proteins. These may include reactive oxygen species (ROS)-scavenging enzymes, small antioxidant biosynthesis enzymes, heat shock proteins, molecular chaperones, and fatty acid desaturases. The identification of transcription factors that regulate the expression of these proteins is crucial, as advancements in gene editing technologies could be applied to coral symbionts in the future. Through the controlled manipulation of key transcription factors, it may be feasible to enhance stress resilience in these symbionts by modulating the expression of protective proteins, potentially mitigating coral bleaching in response to rising ocean temperatures."
} | 1,980 |
36338139 | PMC9626825 | pmc | 85 | {
"abstract": "The demand for non-petroleum-based, especially biodegradable plastics has been on the rise in the last decades. Medium-chain-length polyhydroxyalkanoate (mcl-PHA) is a biopolymer composed of 6–14 carbon atoms produced from renewable feedstocks and has become the focus of research. In recent years, researchers aimed to overcome the disadvantages of single strains, and artificial microbial consortia have been developed into efficient platforms. In this work, we reconstructed the previously developed microbial consortium composed of engineered Pseudomonas putida KT∆ABZF (p2-a-J) and Escherichia coli ∆4D (ACP-SCLAC). The maximum titer of mcl-PHA reached 3.98 g/L using 10 g/L glucose, 5 g/L octanoic acid as substrates by the engineered P. putida KT∆ABZF (p2-a-J). On the other hand, the maximum synthesis capacity of the engineered E. coli ∆4D (ACP-SCLAC) was enhanced to 3.38 g/L acetic acid and 0.67 g/L free fatty acids (FFAs) using 10 g/L xylose as substrate. Based on the concept of “nutrient supply-detoxification,” the engineered E. coli ∆4D (ACP-SCLAC) provided nutrient for the engineered P. putida KT∆ABZF (p2-a-J) and it acted to detoxify the substrates. Through this functional division and rational design of the metabolic pathways, the engineered P. putida - E. coli microbial consortium could produce 1.30 g/L of mcl-PHA from 10 g/L glucose and xylose. Finally, the consortium produced 1.02 g/L of mcl-PHA using lignocellulosic hydrolysate containing 10.50 g/L glucose and 10.21 g/L xylose as the substrate. The consortium developed in this study has good potential for mcl-PHA production and provides a valuable reference for the production of high-value biological products using inexpensive carbon sources.",
"conclusion": "4 Conclusion In this study, we reconstructed and optimized a previously developed microbial consortium consisting of engineered P. putida and E. coli , which can produce mcl-PHA from lignocellulosic hydrolysate. Heterologous expression of the gene encoding laccase from S. azureus resulted in the engineered E. coli ∆4D (ACP-SCLAC), which could produce 3.38 g/L of acetic acid using 10 g/L xylose as substrate, indicating that the acetic acid production capacity was not reduced by the metabolic burden imposed by the dual expression vector. At the same time, 0.67 g/L of FFAs was obtained, which was 48.89% higher than that in E. coli ∆4D. In the engineered strain P. putida KT∆ABZF (p2-a-J), the phaZ gene encoding PHA depolymerase and the yqeF gene were also knocked out in order to weaken fatty acid β-oxidation, resulting in a mcl-PHA titer of 3.98 g/L using 10 g/L glucose and 5 g/L octanoic acid as substrates, which was 1.75 times higher than that of the wild-type P. putida KT2440. The microbial consortium consisting of these above two engineered bacteria had good interspecies communication, and the consortium could efficiently produce an mcl-PHA titer of 1.64 g/L using a mixed carbon source consisting of 10 g/L glucose and 10 g/L xylose. Furthermore, the consortium could accumulate 1.02 g/L of mcl-PHA using lignocellulosic hydrolysate containing 10.50 g/L glucose and 10.21 g/L xylose, which had a competitive advantage ( Table 2 ). In all, the successful reconstruction of the microbial consortium based on the concept of “nutrient supply-detoxification” can be used as a basis for further process development to produce high value-added compounds such as mcl-PHA from lignocellulosic biomass. TABLE 2 Research on artificial consortia for PHA synthesis. Artificial consortia Substrate Type of PHA Titer (g/L) Yield (g/g) Production ratio (g/L/h) References \n P. putida KT2440 + E. coli MG1655 Glucose + Xylose mcl-PHA 1.64 ± 0.10 0.08 0.03 This study \n P. putida KT2440 + E. coli MG1655 Glucose + Pretreatments of corn stover mcl-PHA 1.02 ± 0.01 0.06 0.02 This study \n P. putida KT2440 + E. coli MG1655 Pretreatments of corn stover mcl-PHA 0.65 ± 0.01 0.05 0.01 The study \n P. putida KT2440 + E. coli MG1655 Glucose + Xylose mcl-PHA 1.32 ± 0.03 0.07 0.02 \n Zhu et al. (2021) \n \n P. putida KT2440 + E. coli MG1655 Glucose + Xylose mcl-PHA 0.54 ± 0.03 0.03 0.01 \n Liu et al. (2020) \n \n P. putida KT2440 + E. coli MG1655 Corn stover hydrolysate mcl-PHA 0.43 ± 0.02 0.02 0.01 \n Liu et al. (2020) \n \n S. elongatus cscB + P. putida cscAB CO 2 \n mcl-PHA 0.16 ± 0.04 N.A. <0.01 \n Loewe et al. (2017) \n \n C. necator DSM 428 + P. citronellolis NRRL B-2504 Apple pulp waste P (3HB) and mcl-PHA 1.85 ± 0.03 0.11 0.04 \n Rebocho et al. (2020) \n \n S. degradans 2-40 + B. cereus \n Xylan scl-PHA 0.27 N.A. <0.01 \n Sawant et al. (2017) \n \n R. eutropha H16 + B. subtilis 5119 Sugarcane sugar scl-PHA 2.30 0.08 0.02 \n Bhatia et al. (2018) \n \n A. hydrophila ATCC7966 + A. junii BP25 Acetic acid + butyric acid scl-PHA 2.64 N.A. 0.01 \n Anburajan et al. (2019) \n \n C. necator IPT 026 + X. campestris IBSBF 1867 Palm oil scl-PHA 6.43 0.05 0.05 \n Rodrigues et al. (2019) \n Values were calculated based on visible data of the original paper with unified to two decimal places. The yield is the ratio of the final PHA titer to the substrate concentration consumed. The production ratio is the titer of the PHA during a specific period. The data in the “Titer” column without the standard error added is not available in the original paper. N.A., not available in the original paper.",
"introduction": "1 Introduction The diminishing supply of fossil resources and ecological problems caused by petroleum-based plastics have led to a new wave of exploration of biodegradable polymers that can replace traditional petroleum-based plastics ( Koller and Mukherjee, 2022 ). Polyhydroxyalkanoates (PHAs) are natural polyesters that can be synthesized by microorganisms from renewable sources ( Chen et al., 2020 ). PHAs are considered a viable alternative to petroleum-based plastics due to their excellent biodegradability, biocompatibility, optical activity, gas separation, piezoelectricity, and general physiochemical properties similar to petroleum-based plastics ( Wang et al., 2015 ; Tarrahi et al., 2020 ). The material properties of PHAs depend on the type and distribution of the various monomeric structural units, which can be classified into three groups on the basis of the number of constituent carbon atoms, including short-chain-length polyhydroxyalkanoates (scl-PHAs) such as PHB, composed of 3-5 carbon atoms, medium-chain-length polyhydroxyalkanoates (mcl-PHAs), consisting of monomers with 6–14 carbon atoms, and long-chain-length polyhydroxyalkanoates (lcl-PHAs), consisting of monomers with more than 15 carbon atoms ( Xiang et al., 2020 ; Behera et al., 2022 ). There are different degrees of variation in the crystallinity and tensile strength of PHAs as the chain length changes ( Elmowafy et al., 2019 ). Mcl-PHAs have a lower glass transition temperature (Tg) and melting point (Tm) than scl-PHAs ( Rai et al., 2011 ; Gopi et al., 2018 ). This makes mcl-PHAs thermo-elastic, and once the temperature rises above the Tm value, mcl-PHAs exhibit characteristics such as softness, flexibility and amorphous behavior ( Grigore et al., 2019 ). Accordingly, mcl-PHA are considered true elastomers, whereas most scl-PHAs are hard and friable highly crystalline polymers ( Knoll et al., 2009 ). Due to this structural diversity, mcl-PHAs can meet the flexible demands of a wider range of engineering applications ( Chen and Wang, 2013 ; Zhang et al., 2015 ). However, the industrial production of mcl-PHAs is currently negligible compared to the production and commercialization of scl-PHAs ( Koller and Mukherjee, 2022 ). The current barriers limiting the large-scale industrial production of mcl-PHAs are mainly the process cost and production capacity ( Mannina et al., 2020 ; De Donno Novelli et al., 2021 ). The currently available microbial cell factories require the provision of expensive precursors such as FFAs and have limited production capacity. We therefore need to better understand which natural microorganisms are suitable for mcl-PHA production from a given substrate and how to modify natural microorganisms by means of metabolic engineering to develop efficient cell factories, which will enable us to reduce production costs by using inexpensive carbon sources. Mcl-PHAs can be produced naturally by a variety of microorganisms, especially species of Pseudomonas ( Możejko-Ciesielska and Kiewisz, 2016 ; Chio et al., 2019 ). For example, P. putida KT2440 can naturally produce mcl-PHA as an energy storage material under conditions of surplus FFAs. Moreover, it has a clear genetic background and diverse metabolism pathways. The synthesis of mcl-PHAs in P. putida KT2440 proceeds through the fatty acid β-oxidation pathway and is based on so-called related carbon sources such as FFAs. However, it can also use the fatty acid de novo pathway to produce mcl-PHAs from unrelated carbon sources such as sugars ( Prieto et al., 2016 ). There are two key enzymes that link the fatty acid β-oxidation pathway to mcl-PHA biosynthesis. The enzyme encoded by the phaJ gene converts the β-oxidation intermediate enoyl-CoA into 3-hydroxyacyl-CoA ( Belda et al., 2016 ), which is then polymerized into mcl-PHA by the enzyme encoded by phaC . Furthermore, in terms of metabolic pathways, direct provision of FFAs seems to be an effective strategy for enhancing mcl-PHA, but as mentioned above this also results in a significant increase of process cost, thus requiring us to look for less expensive carbon sources. A good candidate is lignocellulose, which is widely available and considered to be the most abundant organic raw material on earth ( Al-Battashi et al., 2019 ). In the past decades, the conversion of lignocellulosic biomass into value-added products has attracted increasing attention, and has been widely studied due to its low cost and other advantages. However, lignocellulose contains a mixture of sugars that make it difficult for a single strain to fully utilize the available carbon source, and although metabolic engineering may give them the ability to utilize multiple substrates ( Wang et al., 2018 ), it also increases the metabolic burden. On top of that, pure cultures have high maintenance costs, which runs counter to the starting point of using lignocellulose to reduce costs ( Chen et al., 2019 ). Considering these problems, researchers tried to produce mcl-PHA using natural microbial communities, but the functions of the individual strains in these communities could not be fully determined. Moreover, the time required to domesticate wild strains is long, and their transformation efficiency is generally low ( Morgan-Sagastume et al., 2015 ). In order to solve this problem, the targeted design and construction of artificial microbial consortia from the perspective of synthetic biology development can achieve tasks that cannot be accomplished by pure cultures of individual microorganisms. Compared with natural communities, artificial microbial consortia with simple composition and clear division of labor can reduce the metabolic burden of monocultures and theoretically achieve efficient production of mcl-PHA by regulating the interactions between bacteria ( Minty et al., 2013 ). However, the complexity of the design and construction of artificial microbial consortia has limited the achievable titers of mcl-PHA ( Kataria et al., 2018 ). In 2019, we strengthened the acetate assimilation pathway of P. putida KT2440 by overexpressing the acs gene encoding acetyl CoA synthase and constructing the acetate kinase-phosphotransacetylase pathway encoded by ackA-pta ( Yang et al., 2019 ). The engineered P. putida produced 0.674 g/L of mcl-PHA from acetate, suggesting that acetate may be a potential substrate for the production of mcl-PHAs. To avoid the limitations of single-strain cultures, we constructed a microbial consortium consisting of engineered E. coli and P. putida based on the concept of “nutrient supply-detoxification” in 2020. Based on the engineered P. putida overexpressing the acs gene, the ptsG and manZ genes were knocked out in E. coli to reduce substrate competition so that it could preferentially utilize xylose, while the atpFH and envR genes were further knocked out to promote the ability of engineered E. coli to synthesize acetate and FFAs as precursors ( Liu et al., 2020 ). When grown on a mixture of glucose and xylose, the consortium produced an increased mcl-PHA titer of 0.541 g/L, demonstrating its potential to utilize lignocellulose. However, the titer of mcl-PHA still needed to be improved due to the limited modification of engineered P. putida . To address this issue, we further optimized P. putida to improve the conversion efficiency of FFAs to mcl-PHA by knocking out the fadA and fadB genes in the FFAs β-oxidation pathway and overexpressing the phaJ gene ( Zhu et al., 2021 ). At the same time, fatty acid catabolism was downregulated by knocking out the fadD gene and expressing a heterologous gene encoding the acyl carrier protein thioesterase from E. coli . The substrate conversion efficiency and mcl-PHA titer of the reconstituted microbial consortium were significantly improved, with a maximum mcl-PHA titer of 1.32 g/L obtained from a mixture of glucose and xylose. However, the functional validation of lignocellulose utilization by the consortium was limited to the use of a mixed carbon source comprising glucose and xylose, while the substrate conversion efficiency and productivity still needed further improvement. In this study, we further enhanced the metabolic functions of the two strains in the artificial P. putida - E. coli microbial consortium ( Figure 1 ), allowing further improvement in the conversion of substrates to intermediate metabolites and final products. This was achieved by introducing dual vectors encoding the acyl carrier protein thioesterase gene ( ACP ) and the SCLAC gene encoding laccase from Streptomyces azureus to improve the ability of the engineered E. coli to secrete acetic acid and FFAs, which were the material basis for interspecies communication. In the engineered P. putida , we further knocked out the phaZ gene encoding PHA depolymerase to restrict PHA catabolism and further knocked out the yqeF gene to weaken the fatty acid β-oxidation pathway, thus enhancing PHA precursor synthesis. In addition, we overexpressed the phaJ gene encoding enoyl coenzyme A hydratase and the acs gene in the acetic acid assimilation pathway to improve the efficiency of acetic acid and FFAs conversion. Based on this design, we reconstructed the artificial consortium, then optimized its aerobic fermentation conditions, including the proportion of mixed sugars, the balance of nutrient restriction, and cell growth. Finally, we tested the ability of the consortium to produce mcl-PHA using lignocellulosic hydrolysate. FIGURE 1 Schematic diagram of the metabolic engineering protocol for the production of mcl-PHA by the P. putida - E. coli microbial consortium. Acronyms: X5P, Xylulose 5-phosphate; G3P, Glyceraldehyde-3-phosphate; Acetyl-P, Acetyl Phosphate; G6P, Glucose-6-Phosphate. The red words in the biosynthetic pathway indicate the deletion of corresponding genes. The blue words indicate the overexpression of corresponding genes. The related genetic engineering represented by i∼xii was conducted in this work.",
"discussion": "3 Results and discussion 3.1 Metabolic engineering of P. putida for increased medium-chain-length polyhydroxyalkanoate production The design and construction of the artificial microbial consortium was based on the “nutrient supply-detoxification” inter-bacterial relationship with acetic acid and FFAs as intermediate metabolites. In this approach, the targeted enhancement of metabolic functions in different chassis cells can help enhance the capacity of product synthesis in the microbial consortium. Based on the metabolic pathway of P. putida KT2440 to synthesize mcl-PHA, we applied various metabolic engineering strategies with the goal of producing mcl-PHA from lignocellulosic hydrolysate in the artificial microbial consortium. These included: 1) elimination of mcl-PHA solubilization ( Salvachua et al., 2020a ), 2) weakening of the fatty acid β-oxidation pathway to reduce the flux of intermediates to FFAs degradation ( Liu et al., 2011 ), 3) inhibition of competing pathways ( Borrero-de Acuña et al., 2014 ), and 4) enhancement of the conversion of β-oxidation intermediates to synthesize mcl-PHA precursors ( Zhang et al., 2021 ). To eliminate mcl-PHA depolymerization, the phaZ gene encoding PHA depolymerase was knocked out in the wild-type strain P. putida KT2440 as well as the engineered KTΔAB strain in which the fadA and fadB genes had been knocked out previously, resulting in the engineered strains KT2440ΔZ and KTΔABZ, respectively. In order to reduce the degradation of the β-oxidation intermediate 3-hydroxyacyl-CoA, the yqeF gene encoding acetyl coenzyme A acetyltransferase and the paaJ gene encoding β-ketohexanoyl coenzyme A thiolase were knocked out in KTΔABZ to produce the engineered strains KTΔABZF and KTΔABZFJ. In addition, the tctA gene encoding a tricarboxylic acid transport protein was knocked out in strain KTΔABZFJ to produce KTΔABZFJT, which can be used to disrupt the tricarboxylic acid transport system aiming to divert energy toward PHA accumulation. To further increase the conversion efficiency of acetic acid and FFAs in these knockout strains, thus contributing to the availability of mcl-PHA precursors, the acs gene encoding acetyl coenzyme A synthase and the phaJ gene encoding enoyl coenzyme A hydratase were overexpressed to produce the strains KTΔABZF (p2-a-J) and KTΔABZFJ (p2-a-J). To enhance the polymerization of the PHA precursor 3-hydroxyacyl-CoA to mcl-PHA, the phaC1 and phaC2 genes encoding PHA polymerase ( Li et al., 2021 ) were expressed using the T3 promoter or Tac promoter, to produce strains KTΔABZF (p2 T3 -C1C2) and KTΔABZF (p2 Tac -C1C2), respectively. Furthermore, the acs and phaJ genes were overexpressed using the T3 promoter, while the phaC1 and phaC2 genes were overexpressed using the Tac promoter, resulting in the strain KTΔABZF (p2-a-J-C1C2). In order to overcome the inhibitory effect of lignin in the lignocellulosic hydrolysate, the SCLAC gene encoding laccase derived from S. azureus was expressed heterologously using the T3 promoter or Tac promoter, resulting in the strains KTΔABZF (p2 T3 -SCLAC) and KTΔABZF (p2 Tac -SCLAC), respectively. 3.1.1 Engineering the endogenous metabolism of P. putida to increase the production of medium-chain-length polyhydroxyalkanoate To test the mcl-PHA production capacity of the different metabolically engineered P. putida strains introduced above, we used 10 g/L glucose and 5 g/L octanoic acid as substrates for the synthesis of mcl-PHA in a two-step fermentation lasting 60 h. After the engineered P. putida as incubated with glucose for 24 h, octanoic acid was added to the medium as a precursor of mcl-PHA. The wild-type strain KT2440 and the starting engineered strain KTΔAB were included as controls. As shown in Figure 2A , the intracellular mcl-PHA content of all the engineered strains increased to some extent compared to the wild-type strain KT2440, as well as the initial engineered strain KTΔAB. As expected, the intracellular content of mcl-PHA in the engineered KT2440ΔZ and KTΔABZ with knockout of the phaZ gene were enhanced compared to their respective parental strains. This led to a significant increase of PHA production, which was consistent with the literature ( Salvachúa et al., 2020b ). This suggests that knocking out the phaZ gene encoding PHA depolymerase to eliminate PHA depletion is an effective strategy to increase PHA production. The mcl-PHA content of the KT∆ABZF strain with the yqeF gene knocked based on KT∆ABZ reached 3.62 g/L, which was 1.59 times higher than that of the original strain KT2440 and 2.76 times higher than that of the initial engineered strain KT∆AB. Acetyl coenzyme A acetyltransferase is encoded by the yqeF gene ( Zhuang and Qi, 2019 ), and the deletion of this gene resulted in a large accumulation of acetyl-CoA, weakening the fatty acid β-oxidation cycle. Further knockout of the paaJ gene resulted in a strong blockage of the fatty acid β-oxidation cycle, as shown in Supplementary Table S3 , and although a high intracellular content of PHA was accumulated (85.89 wt%), the strain’s growth was significantly inhibited because the carbon flux from the weakened fatty acid β-oxidation was mainly redirected toward the synthesis of PHA, which reduced the intracellular energy supply, resulting in poor biomass accumulation of the engineered strain KT∆ABZFJ, as evidenced by a significant decrease of cell dry weight (CDW). By contrast, KT∆ABZFJT with a further knock out of the tctA gene encoding the tricarboxylic acid transport protein based on KT∆ABZFJ, did not perform as well as reported for a similar strain in a previous study ( Tao et al., 2017 ), probably because the simultaneous blockage of the tricarboxylic acid cycle and severe inhibition of the fatty acid β-oxidation cycle resulted in the absence of reducing coenzymes essential for cell growth, which could not complete oxidative phosphorylation to provide the necessary energy for cell growth ( Zhuang et al., 2014 ; Tao et al., 2017 ). FIGURE 2 The ability of the engineered P. putida by genetic modifications to produce mcl-PHA. (A) The engineered P. putida by knocking out corresponding genes. (B) The engineered P. putida by overexpression corresponding genes. (C) The engineered P. putida by heterologous expression of laccase. The error bars indicate the standard deviation of triplicate experiments. * p < 0.05; ** p < 0.01; *** p < 0.001. The mcl-PHA production capacity of the engineered strains was characterized under the same conditions. To exclude the effect of the introduced plasmid pBBR1MCS2 on the engineered bacteria, the KT∆ABZF (p2) strain containing the empty vector was used as a control. The content of intracellular PHA in engineered KT∆ABZF (p2-a-J) overexpressing the acs and phaJ genes reached 91.96 wt% ( Supplementary Table S3 ) and the mcl-PHA titer was 3.98 g/L ( Figure 2B ), which was 3.18 times higher than that of the control strain KT∆ABZF (p2). However, similar co-expression of both phaJ and fabG in P. putida KCTC1639 did not yield satisfactory results in a previous study ( Vo et al., 2008 ). Nevertheless, the cell growth and mcl-PHA accumulation of the overexpression strain KT∆ABZFJ (p2-a-J) were improved compared with its parental strain KT∆ABZFJ, indicating that overexpression of acs could promote the production of acetyl coenzyme A, thus alleviating the cell growth limitation of KT∆ABZFJ due to the severe weakening of fatty acid β-oxidation. At the same time, the phaJ gene enhanced the conversion efficiency of FFAs into mcl-PHA precursors via the β-oxidation pathway. In addition, overexpression of phaC1 and phaC2 using different promoters resulted in significant differences of cell growth and mcl-PHA accumulation in the engineered bacteria ( Supplementary Table S3 ). Accordingly, the intracellular content of mcl-PHA reached 96.77 wt% in KT∆ABZF (p2 Tac -C1C2) using the Tac promoter, while the growth of KT∆ABZF (p2 T3 -C1C2) using the T3 promoter was significantly inhibited. This might be due to the fact that T3-mediated expression did not achieve the expected effect of polymerizing PHA precursors and the strain could not effectively utilize FFAs to accumulate mcl-PHA, while still incurring a metabolic burden due to the overexpression vector. Additionally, the mcl-PHA production of the co-overexpression strain KT∆ABZF (p2-a-J-C1C2) was not further enhanced, probably due to the increased metabolic burden. The monomer fractions of PHA accumulated by these metabolically engineered strains are summarized in Supplementary Table S3 . The PHA contained all even medium-chain-length monomers from C6-C14, whereby 3-hydroxyoctanoate (3HO) accounted for the largest fraction. For example, strain KT∆ABZF (p2-a-J) accumulated 86.69 mol% of 3HO. As expected, the utilization of octanoic acid as precursor caused the accumulation of C8 monomers due to the disruption of FFAs degradation. 3.1.2 Engineering P. putida by heterologous expression of laccase to increase inhibitor tolerance Since this study aimed to produce mcl-PHA from lignocellulosic hydrolysate, it was necessary to decrease the potential inhibitory effect of lignin on the engineered P. putida . Accordingly, the SCLAC gene encoding laccase from S. azureus was introduced into the engineered strain KT∆ABZF and heterologously expressed using the T3 and Tac promoters, respectively. To examine the effect of heterologous expression of laccase on the accumulation of mcl-PHA by the engineered P. putida , the engineered bacteria were incubated in M9 medium with a final glucose concentration of 10 g/L for 24 h, followed by the addition of 2 g/L of the lignin-derived monomer p-coumaric acid (p-CA) as a substrate until 60 h ( Figure 2C ). The medium with only 10 g/L glucose was used for comparison, while the KT2440, KT∆ABZF and KT∆ABZF (p2) strains were included as controls. In all engineered strains, the CDW was significantly lower in the medium with added p-CA compared to glucose only ( Supplementary Table S3 ), indicating that the biomass accumulation was significantly inhibited, which was caused by the cytotoxicity of p-CA itself ( Sarlaslani, 2007 ). In contrast to the CDW, the intracellular content of PHA was increased in all the engineered strains in the medium supplemented with p-CA, which was attributed to the fact that P. putida KT2440 has a native ability to break down aromatic compounds, and convert p-CA into mcl-PHA via protocatechuic acid and the β-KA pathway to complete the lignin depolymerization and synthesis of FFAs ( Linger et al., 2014 ). In particular, KT∆ABZF (p2 T3 -SCLAC) and KT∆ABZF (p2 Tac -SCLAC) both showed a significant increase in the intracellular content of mcl-PHA, with KT∆ABZF (p2 Tac -SCLAC) accumulating close to 50 wt% mcl-PHA. At the same time, the mcl-PHA titer of both engineered strains was increased compared to KT∆ABZF (p2), especially when p-CA was added as a substrate, whereby the PHA titer of KT∆ABZF (p2 Tac -SCLAC) was higher (1.0 g/L) than that of the corresponding strain with only glucose as substrate. This indirectly indicates a positive effect of heterologous expression of laccase on the potential utilization of lignocellulosic hydrolysate to produce mcl-PHA by the engineered P. putida . Unlike the previous use of octanoic acid as a precursor, the engineered strain utilizing only glucose accumulated PHA with predominantly 3-hydroxydecanoate (3HD) monomers, similar to the engineered bacteria in the medium supplemented with p-CA, with KT∆ABZF (p2 Tac -SCLAC) accumulating the highest 3HD fraction of 56.75 mol% ( Supplementary Table S3 ). 3.1.3 Production of medium-chain-length polyhydroxyalkanoate by the engineered P. putida using mixed substrates When the artificial microbial consortium produces mcl-PHA from lignocellulosic hydrolysate, the potential carbon sources utilized by engineered P. putida include glucose, acetic acid, FFAs and aromatic compounds. In order to test the production of mcl-PHA by engineered P. putida using mixed substrates, we fermented the engineered P. putida in medium containing 10 g/L glucose, 5 g/L acetic acid, 3 g/L octanoic acid and 2 g/L p-CA as mixed substrates for 60 h. The original strain KT2440, the parental strain KT∆ABZF and the engineered strain KT∆ABZF (p2) containing the empty vector pBBR1MCS-2 were also fermented under the same conditions as controls. As shown in Figure 3 , the engineered strains KT∆ABZF (p2 Tac -C1C2) and KT∆ABZF (p2-a-J) showed superior performance in terms of intracellular content and final titer of mcl-PHA. These two engineered strains accumulated comparable amounts of mcl-PHA, with similar titers of 1.96 and 1.80 g/L, respectively. The remaining engineered strains showed a decrease of CDW, intracellular PHA content ( Supplementary Table S3 ) and final PHA titer compared to KT∆ABZF, indicating that acetic acid and p-CA may also decrease cell growth ( Sarlaslani, 2007 ; Ludwig et al., 2013 ) in addition to the effect of the vector itself. In the mcl-PHA synthesized by all strains, 3HO still accounted for the highest percentage of monomers, indicating that engineered P. putida could efficiently convert octanoic acid for the accumulation of mcl-PHA. FIGURE 3 The performance of mcl-PHA production by the engineered P. putida using 10 g/L glucose, 5 g/L acetic acid, 3 g/L octanoic acid and 2 g/L p-CA as mixed substrates. The engineered P. putida were incubated with 10 g/L glucose, 5 g/L acetic acid, 3 g/L octanoic acid and 2 g/L p-CA as mix substrates for 60 h. The error bars indicate the standard deviation of triplicate experiments. * p < 0.05; ** p < 0.01; *** p < 0.001; ns, no significance. In summary, the results showed that knockout of the phaZ gene encoding PHA depolymerase and moderate weakening of the fatty acid β-oxidation pathway could effectively improve the mcl-PHA production, where by the product titer of the engineered strain KT∆ABZF reached 3.62 g/L using 10 g/L glucose and 5 g/L octanoic acid as substrates, which was 2.76 times higher than that of the initial engineered strain KT∆AB. The engineered strain KT∆ABZF (p2-a-J), with an additionally enhanced acetic acid assimilation pathway and improved conversion of fatty acid β-oxidation intermediates, exhibited a further increase of the mcl-PHA titer to 3.98 g/L, which was 3.04 times higher than that of KT∆AB. In addition, heterologous expression of laccase also had a positive effect on the ability of the engineered P. putida to use the lignin derivative p-CA for mcl-PHA production, whereby the engineered strain KT∆ABZF (p2 Tac -SCLAC), which heterologously expressed SCLAC using the Tac promoter, accumulated 49.39% of mcl-PHA, which was 142.82% higher than in the wild-type strain KT2440. Finally, we tested the mcl-PHA production capacity of the engineered bacteria in lignocellulosic hydrolysate using simulated mixed substrates, and the results showed that the engineered strain KT∆ABZF (p2-a-J) produced 1.80 g/L mcl-PHA, which was 73.08% higher compared to the wild-type strain KT2440. 3.2 Metabolic engineering of E. coli for enhanced substrate supply for the engineered P. putida to produce medium-chain-length polyhydroxyalkanoate Based on our previous studies, engineered E. coli has been shown to preferentially utilize xylose in media containing both glucose and xylose, while slowly metabolizing glucose and secreting the intermediate metabolites acetate and FFAs ( Liu et al., 2020 ; Zhu et al., 2021 ). In this study, we showed that engineered P. putida can use FFAs as a related carbon source to increase the supply of precursors for mcl-PHA production and improve the final titer. In addition, it has been shown that overexpression of relevant genes in the FFAs metabolic pathway of E. coli is an effective strategy to improve FFAs production ( Lennen et al., 2010 ; Zhang et al., 2012 ). Therefore, increasing the FFAs and acetic acid production of E. coli is essential for efficient synthesis of mcl-PHA from lignocellulosic hydrolysates using artificial microbial consortia. In this sturdy, we used E. coli ∆4D as the starting strain and overexpressed the tesA gene ( Steen et al., 2010 ) encoding thioesterase, the fabZ gene ( Yu et al., 2011 ) encoding hydroxylated acyl-ACP dehydratase and the fabD gene ( Zhang et al., 2012 ) encoding acyltransferase to produce the engineered E. coli ∆4D ( tesA ), E. coli ∆4D ( fabZ ) and E. coli ∆4D ( fabD ), respectively. Among them, the enzyme encoded by the tesA gene catalyzes the last step of FFAs synthesis and can release free FFAs from acyl ACP, while the enzymes encoded by the fabZ and fabD genes enhance the synthesis of acyl ACP. To further enhance the supply of precursors for FFAs synthesis, the thioesterase gene of the ricin acyl carrier protein ( Hernández Lozada et al., 2018 ) was co-expressed with the above three genes in different combinations to produce the engineered strains E. coli ∆4D (AD), E. coli ∆4D (AAD), E. coli ∆4D (AZD) and E. coli ∆4D (AAZD), respectively. In addition, to achieve the secretion of FFAs and acetic acid by engineered E. coli grown on lignocellulosic hydrolysate, we added another vector containing the SCLAC gene to E. coli ∆4D (ACP) and constructed the engineered E. coli ∆4D (ACP-SCLAC). 3.2.1 Engineering E. coli to produce acetic acid and free fatty acids from xylose The ability of the engineered E. coli strains to produce acetic acid ( Figure 4A ) and FFAs ( Figure 4B ) was assessed in shake-flask fermentations with 10 g/L xylose as substrate for 64 h. The starting strains E. coli ∆4D and E. coli ∆4D (ACP) were included as controls. As shown in Figure 4A , the engineered strains that individually overexpressed the tesA , fabZ and fabD genes in the fatty acid synthesis pathway, did not have an enhanced ability to produce acetic acid, whereas the engineered strains that overexpressed these genes in different combinations produced more acetic acid. Specifically, the engineered E. coli ∆4D (AZD) and E. coli ∆4D (AAZD) respectively produced 3.67 and 3.55 g/L acetic acid, representing 22.74 and 18.73% increases compared to E. coli ∆4D. By contrast, the production of FFAs was increased in the engineered strains with individual overexpression of tesA , fabZ and fabD , as shown in Figure 4B . However, the FFAs production of E. coli ∆4D (A) overexpressing the tesA gene was reduced to 0.65 g/L, compared to 0.86 g/L produced by the engineered E. coli ∆4D (ACP). This indicated that the overexpression of the gene encoding thioesterase had a significant effect on FFA secretion in engineered E. coli , in agreement with the literature ( Lee et al., 2013 ). Compared to the control strain E. coli ∆4D, the acetate and FFAs production of E. coli ∆4D (A) increased by 91.11% and 44.44%, respectively, indicating that overexpression of the tesA gene had a positive effect. However, no significant increase in FFAs production was observed in the co-overexpression strain compared to E. coli ∆4D. In terms of metabolic pathways, co-overexpression may have led to excessive flux through the fatty acid synthesis pathway, which may have resulted in the depletion of the acetyl-CoA pool in the engineered E. coli ( Lu et al., 2008 ), thus leading to reduced xylose utilization ( Supplementary Figure S1 ). In addition, the dual-vector strain E. coli ∆4D (ACP-SCLAC) secreted up to 3.38 g/L acetic acid and 0.67 g/L FFAs. The FFAs production capacity of E. coli ∆4D (ACP-SCLAC) was reduced compared to E. coli ∆4D (ACP), which was ascribed to the increased metabolic burden from the dual plasmid vector. By contrast, both acetic acid and FFAs production were significantly higher compared to E. coli ∆4D, and no residual xylose was detected in the medium after 64 h ( Supplementary Figure S1 ). Therefore, establishing a balance between product accumulation and strain growth is necessary to increase the production of acetic acid and FFAs in engineered E. coli . FIGURE 4 The production of acetic acid and FFAs by the engineered E. coli using different substrates. (A) The production of acetic acid by engineered E. coli using 10 g/L xylose as substrates for 64 h. (B) The production of FFAs by engineered E. coli using 10 g/L xylose as substrates for 64 h. (C) The production of acetic acid by engineered E. coli using lignocellulosic hydrolysate as substrates for 64 h. (D) The production of FFAs by engineered E. coli using lignocellulosic hydrolysate as substrates for 64 h. The error bars indicate the standard deviation of triplicate experiments. * p < 0.05; ** p < 0.01; *** p < 0.001; ns, no significance. 3.2.2 Engineering E. coli to utilize lignocellulosic hydrolysate for medium-chain-length polyhydroxyalkanoate production In view of the above results on the production of acetic acid and FFAs by engineered E. coli using xylose, as well as the design of the reconstructed artificial microbial consortium for the production of mcl-PHA using lignocellulosic hydrolysate, we selected the three engineered strains E. coli ∆4D, E. coli ∆4D (ACP) and E. coli ∆4D (ACP-SCLAC) for fermentation experiments using lignocellulosic hydrolysate obtained through acid pretreatment of corn straw. The lignocellulosic hydrolysate used in this study contained 1.63 g/L glucose, 21.39 g/L xylose and 3.13 g/L arabinose, which was in agreement with a previous study using a similar pretreatment method ( Avci et al., 2013 ). The lignocellulosic hydrolysate and M9 medium were mixed at a 9:1 (v/v) ratio and used as the culture medium ( Salvachua et al., 2015 ). As shown in Figures 4C,D , the acetic acid titer of the engineered E. coli ∆4D (ACP-SCLAC) reached 2.32 g/L, which was 2.76 times higher than that of the starting strain E. coli ∆4D. Although the increase in FFAs production of E. coli ∆4D (ACP-SCLAC) compared to E. coli ∆4D (ACP) was not significant, the heterologous expression of laccase enhanced the ability of the engineered E. coli ∆4D (ACP-SCLAC) to use lignocellulosic hydrolysate to some extent. In particular, the engineered E. coli ∆4D (ACP-SCLAC) secreted 0.69 g/L FFAs, which was close to the titer obtained on xylose. Combined with the changes in the concentration of major sugars in the medium of the lignocellulosic hydrolysate ( Supplementary Table S4 ), it was found that engineered E. coli did not significantly deplete the arabinose that was also available in the medium, while the initial small amount of glucose in the medium was completely consumed, along with almost 5 g/L of xylose. Theoretically, E. coli can use glucose, xylose as well as arabinose, and although the strains used in this study were genetically engineered to prefer xylose and slowly metabolize glucose, carbon catabolite repression (CCR) was not completely released. Similar studies have shown that this appears to be due to residual activity of the glucose PTS (phosphotransferase system) ( Xiao et al., 2011 ). In conclusion, the engineered E. coli ∆4D (ACP-SCLAC) exhibited improved acetic acid and FFA production capacity using xylose or lignocellulosic hydrolysate as substrate and was selected as one of the functional strains to construct the artificial microbial consortium. 3.3 Reconstruction and optimization of the artificial microbial consortium using mixed sugars Artificial microbial consortia are a novel and promising platform for the biosynthesis of target products that can distribute the metabolic burden among individual strains ( Li et al., 2019 ; Roell et al., 2019 ). In this study, the “nutrient supply-detoxification” principle was used to reconstruct a synthetic microbial consortium consisting of two engineered strains belonging to different species ( Figure 1 ). First, the best-performing engineered E. coli ∆4D (ACP-SCLAC) was selected as the nutrient supply module for the microbial consortium, while the most promising strain for mcl-PHA production, P. putida KT∆ABZF (p2-a-J), was selected to perform the task of utilizing the complex components of lignocellulosic hydrolysate while detoxifying acetic acid and FFAs for increased productivity. One of our goals in designing the artificial microbial consortium was to use lignocellulosic hydrolysate as an inexpensive substrate for the production of mcl-PHA. We therefore first tested the ability of the artificial microbial consortium to produce mcl-PHA using glucose and xylose, the main monosaccharides found in lignocellulosic hydrolysate, as mixed carbon sources. The construction of an artificial microbial consortium that can fully coordinate the interactions and synergistic growth of bacteria is the key to optimizing the co-culture strategy ( Kong et al., 2018 ). Our previous studies have demonstrated that the inoculation ratio of consortia has a small effect on the titer of mcl-PHA, while the inoculation order between engineered bacteria has a decisive effect on strain dominance in the co-culture ( Zhu et al., 2021 ), which in turn affects the titer of mcl-PHA. We therefore added E. coli ∆4D (ACP-SCLAC) and P. putida KT∆ABZF (p2-a-J) at a 1:2 ratio, whereby the engineered P. putida was added to the medium 12 h before the engineered E. coli ∆4D (ACP-SCLAC). Moreover, due to the preference of the two species for glucose and xylose, as well as the significant effect of the intermediate metabolites (acetic acid and FFA) on mcl-PHA synthesis, adjusting the ratio of glucose and xylose in the mixed sugar medium is likely to affect the final titer of mcl-PHA. Consequently, we tested three different ratios of mixed sugar substrates to test the ability of the artificial microbial consortium to produce mcl-PHAs under nutrient-limited conditions with 1 g/L NH 4 Cl. As shown in Figure 5A , the amount of mcl-PHA accumulated in the medium was similar at glucose-xylose ratios of 1:1 and 3:1, with the highest titer reaching 1.30 g/L in the former. By contrast, only 0.71 g/L mcl-PHA was detected in the medium with a glucose-xylose ratio of 1:3. This was because the engineered P. putida KT∆ABZF (p2-a-J) first used glucose to rapidly accumulate biomass, after which the engineered E. coli ∆4D (ACP-SCLAC) subsequently supplied acetic acid and FFAs to the engineered P. putida KT∆ABZF (p2-a-J) to produce mcl-PHA. For the engineered P. putida KT∆ABZF (p2-a-J), which was inoculated first, the carbon source available in the medium consisting of glucose and xylose (1:3) was inadequate to accumulate sufficient biomass, which was positively correlated with the synthesis of mcl-PHA, so that the biomass deficiency affected the subsequent mcl-PHA synthesis. This is also evidenced by the utilization of mixed sugars in the artificial microbial consortium ( Supplementary Figure S2 ). FIGURE 5 The production of mcl-PHA by the P. putida - E. coli microbial consortium using M9 medium with added glucose and xylose. (A) The mcl-PHA titer in the microbial consortium with different mixed sugar ratios and the total concentrations of mix sugars were 20 g/L. (B) The mcl-PHA titer in the microbial consortium at different nitrogen concentration. (C) The residual contents of different sugars by the microbial consortium. (D) The monomer composition of mcl-PHA produced by the microbial consortium. The error bars indicate the standard deviation of triplicate experiments. * p < 0.05; ** p < 0.01; *** p < 0.001; ns, no significance. However, under nitrogen-limited conditions, the post-inoculated E. coli ∆4D (ACP-SCLAC) did not seem to be affected by the CCR and it could continuously consume xylose ( Supplementary Figure S2 ), but the overall xylose utilization rate was low and there was still a small amount of residual xylose in the medium at the end of fermentation. This indicates that the nitrogen-limited conditions affected xylose utilization by E. coli ∆4D (ACP-SCLAC) in the co-culture system. In order to further improve the utilization of xylose by the engineered E. coli , the nitrogen source concentration had to be increased, but nitrogen source limitation promotes the accumulation of mcl-PHA, so exploring the appropriate nitrogen source concentration to achieve a balance between the two engineered bacteria is crucial for the production of mcl-PHA ( Sangani et al., 2019 ). We therefore increased the concentration of the nitrogen source (NH 4 Cl) to 3 g/L, while using glucose and xylose at a 1:1 ratio as the carbon source. As shown in Figure 5C , glucose was rapidly depleted as before, but the increase of nitrogen source caused xylose to be depleted at 52 h of fermentation. In addition, the two-species consortium accumulated more mcl-PHA under these conditions. As shown in Figure 5B , the mcl-PHA titer reached 1.64 g/L, which was 26.15% higher than before. Similar to mcl-PHA produced by engineered P. putida in pure culture in a previous study ( Yang et al., 2019 ), the mcl-PHA produced by the consortium in this study mainly contained medium-length monomers, with C10 as the most abundant monomer ( Figure 5D ), accounting for 50.30% of the total mcl-PHA content. This phenomenon was previously shown to be related to the fact that the main substrate utilized by the engineered P. putida was glucose ( Liu et al., 2020 ). 3.4 Production of medium-chain-length polyhydroxyalkanoate from lignocellulosic hydrolysate using the artificial microbial consortium To explore the productivity of the consortium in the actual lignocellulosic hydrolysate, we used corn straw as raw material and pretreated it with acid to obtain a pretreatment solution rich in xylose and a small amount of glucose (21.39 g/L xylose, 1.63 g/L glucose, and 3.13 g/L arabinose). The straw solids that remained after acid pretreatment were further treated with cellulase and hemicellulose degrading enzymes, resulting in an enzymatic hydrolysate contained mainly 9.29 g/L glucose and 8.59 g/L xylose according to HPLC. The ability of the artificial microbial consortium to produce mcl-PHA in shake flasks was tested using the above acid pretreatment solution and enzymatic digest as substrates, respectively. As shown in Figure 6 , the consortium performed best in the medium based on the pretreatment solution with added glucose (M1-1+), accumulating 1.02 g/L mcl-PHA. However, the addition of glucose seemed to have no significant effect on the synthesis of mcl-PHA by the consortium using a mixture of pretreatment solution and M9 medium at a ratio of 9:1, resulting in similar mcl-PHA production. Specifically, in the pretreatment solution medium without glucose addition (M9-1), the consortium accumulated 0.65 g/L mcl-PHA. Although the initial glucose concentration in the medium was only 2.03 g/L, the engineered P. putida KT∆ABZF (p2-a-J) which was seeded first nevertheless accumulated a certain amount of biomass, and subsequently used the intermediate metabolites acetic acid and FFAs for growth and mcl-PHA accumulation. By contrast, the consortium only accumulated 0.30 g/L mcl-PHA in the pretreatment medium without the addition of glucose (M1-1), which was mainly attributed to insufficient initial biomass. As shown in Supplementary Table S5 , the two-species consortium was less efficient in utilizing sugars from the enzymatic hydrolysate compared to the acid pretreatment, with residual glucose and xylose still remaining at the end of fermentation. The decrease of substrate utilization in turn affected the product titer, and the consortium accumulated 0.45 g/L of mcl-PHA, presumably due to the lack of essential elements for microbial growth in the enzymatic hydrolysate. In addition, the high concentration of citrate contained in the hydrolysate may also have an inhibitory effect on microbial growth. At the same time, the chemicals from the pretreatment process have a strong inhibitory effect on the enzymatic process ( Qin et al., 2016 ). However, the monomer composition of mcl-PHA was dominated by C8 and C10 ( Table 1 ), while most recent studies using lignocellulose as a substrate produced scl-PHA ( Linger et al., 2014 ; Liu et al., 2017 ; Arreola-Vargas et al., 2021 ). FIGURE 6 The production of mcl-PHA using lignocellulosic biomass by the P. putida - E. coli microbial consortium. M1-1: The ratio of acid-pretreatment solution of corn straw and M9 medium was 1:1; M1-1+: The ratio of acid-pretreatment solution of corn straw and M9 medium was 1:1 and 10 g/L glucose was added. M9-1: The ratio of acid-pretreatment solution of corn straw and M9 medium was 9:1; M9-1+: The ratio of acid-pretreatment solution of corn straw and M9 medium was 9:1 and 10 g/L glucose was added. TABLE 1 The monomer compositions of mcl-PHA produced by the microbial consortium using from lignocellulosic hydrolysate. Strains Monomer compositions (%) 3HHX (C6) 3HO (C8) 3HD (C10) 3HDD (C12) 3HTD (C14) M1-1 7.61 ± 0.19 52.35 ± 1.14 20.54 ± 0.38 13.44 ± 0.66 6.06 ± 0.08 M1-1+ 27.50 ± 1.69 28.25 ± 2.46 24.71 ± 2.58 12.81 ± 1.18 6.73 ± 0.39 M9-1 2.47 ± 0.44 66.51 ± 0.97 16.73 ± 1.76 10.83 ± 0.40 3.46 ± 0.82 M9-1+ 16.33 ± 0.02 34.69 ± 1.40 22.04 ± 1.34 18.56 ± 1.48 8.38 ± 1.40 Enzymatic digest solution 8.15 ± 0.53 15.27 ± 1.21 57.83 ± 0.06 11.74 ± 0.51 7.00 ± 0.12 The main difference between the consortium reported in this manuscript and the previous studies is the target substrates. The main objective of our previous studies were to design and validate the ability of the microbial consortia to produce mcl-PHA from mixed sugars ( Liu et al., 2020 ; Zhu et al., 2021 ), whose results demonstrated the good potential of these consortia, of which one could also produce mcl-PHA from lignocellulosic hydrolysate (0.434 g/L) ( Liu et al., 2020 ). Nevertheless, the engineered E. coli ∆4D (ACP-SCLAC) in this study introduced heterologously expressed laccase aimed at increasing the tolerance and utilization of lignocellulosic hydrolysate by the consortium, while other genetic modifications were also made to increase the precursor supply of mcl-PHA, weaken the competitive pathway and reduce the depolymerization of mcl-PHA. Specifically, after various genetic modifications, the engineered E. coli ∆4D (ACP-SCLAC) preferentially used xylose to secrete fatty acid and acetic acid, while the engineered P. putida KTΔABZF (p2-a-J) could use glucose for its own growth and reproduction, and then used fatty acid and acetic acid to achieve the production of mcl-PHA. At the same time, the toxic effect of acetic acid on the E. coli cells was relieved to a large extent. Experimental results verified this design: as shown in Figure 3 of this study, P. putida KT∆ABZF (p2-a-J) produced 1.80 g/L mcl-PHA using mixed substrates (10 g/L glucose, 5 g/L acetic acid, 3 g/L octanoic acid and 2 g/L p-coumaric acid), which indicated that the engineered P. putida KT∆ABZF (p2-a-J) could produce mcl-PHA in the presence of acetic acid and fatty acids and also demonstrated its detoxification effect; Figures 4C,D in the study demonstrated the ability of E. coli ∆4D (ACP-SCLAC) to secrete acetic acid and fatty acids from lignocellulosic hydrolysate with the titer of 2.32 and 0.69 g/L, respectively, indicating that it could supply P. putida KT∆ABZF (p2-a-J) nutrition. The purpose of constructing the microbial consortium in this study was to use a cheap carbon source for mcl-PHA production, and the results showed ( Figure 6 ) that the consortium was able to produce 1.02 g/L mcl-PHA using lignocellulosic hydrolysate, which was comparable with the titer from mixed sugars and was increased by 135.02% from lignocellulosic hydrolysate than our previously constructed consortium ( Liu et al., 2020 ). Petroleum-based plastics are already produced on a very mature scale, with a minimum selling price of about $1 445 per ton ( Kim et al., 2022 ). Currently, the cost of the production of mcl-PHA by microorganisms is still higher than the conventional petroleum-based plastics. However, the use of microbial plants for mcl-PHA production has a broader prospect in view of the increasing crude oil prices and a series of environmental and health problems coming with petroleum-based plastics. There are few published literatures on the techno-economic analysis of mcl-PHA as far as we know, but there are some reports discussing scl-PHA, which has already been industrially produced and is sold at a minimum price of about $4 000 per ton currently ( Wang et al., 2022 ). Meanwhile, current production of mcl-PHA is mainly on laboratory scale because of its higher production cost as well as lower production, compared with scl-PHA. The substrate is considered to be the key factor in the final production cost in the fermentation production of PHA, accounting for about 50% of the total cost ( Kim, 2000 ; Nitzsche et al., 2016 ). The main substrates reported to be used for the production of mcl-PHA with microorganisms include p-coumaric acid, glucose, propionate, nonanoic acid (one of the fatty acids), and xylose, etc. ( Davis et al., 2015 ; Zhuang and Qi, 2019 ; Salvachúa et al., 2020b ; Zhu et al., 2021 ), all of which are high-cost substrates. In contrast, the lignocellulosic hydrolysate used in this study is generated from waste corn stover by pretreatment, thus would significantly reduce the substrate cost, which is also considered to be an excellent substrate for many other products ( Salvachúa et al., 2016 ). As noticed, there is still space for the production improvement of the two-species microbial consortium, which requires improvement of corn stover pretreatment methods, further modification of engineered strains, and optimization of fermentation conditions to achieve the purpose of mass production of high value-added products with low substrate."
} | 13,331 |
29666244 | PMC5939081 | pmc | 87 | {
"abstract": "Significance Quorum sensing is a communication system that allows bacteria to coordinate their activities, and these systems are critical for virulence in several bacteria, including Pseudomonas aeruginosa . There is a significant gap in knowledge about how quorum sensing proceeds during infection, particularly how spatial organization of the infecting microbial community impacts signaling. Using a model that recapitulates the biogeographical properties of P. aeruginosa infection of the cystic fibrosis lung, we discovered that communication primarily occurs within P. aeruginosa aggregates and that communication between aggregates is only observed for very large aggregates containing ≥5,000 cells. This study identifies a critical role for spatial distribution and bacterial phenotypic heterogeneity in bacterial signaling during infection, and provides a platform for future ecological and evolutionary studies.",
"discussion": "Discussion One of the key findings of this study is that SCFM2 aggregates are mostly clonal and display a range of sensitivities to 3OC12-HSL ( Fig. 3 , Table 1 , Table S1 , and Datasets S1 and S2 ). From a population level, the response to increasing 3OC12-HSL is graded with a maximum of ∼80% of aggregates initiating QS when exposed to a 5-µM signal ( Fig. S4 A ). The lack of complete (100%) response to added signal could be biologically relevant or a result of the stringent parameters we used to classify cells as QS + . We suggest the former, because overexpression of lasR increased the sensitivity of P. aeruginosa to 3OC12-HSL levels below 5 µM, but did not increase the maximum response ( Fig. S4 B ). Thus, it is likely that those cells may simply be “blind” to biologically relevant levels of signal (between 0.5 and 1.0 µM). Of course, these nonresponsive cells could be dead or dormant “persisters,” although they are at least grossly structurally intact as they can be identified by the presence of intracellular mCherry, and at least some are metabolically active as we observed nonresponding planktonic cells swimming during microscopic examination. These data indicate that aggregate response is not simply binary, but instead a dynamic response, in which the response of subpopulations of cells lies within a gradient, similar to that observed for single cells ( 28 , 29 ). In addition, the volumes of individual nonresponsive and QS + aggregates were similar ( Table S2 ), indicating that there is no threshold volume for a positive response to quorum signals in SCFM2. Our discovery of a range of aggregate sensitivities to 3OC12-HSL likely has important implications for P. aeruginosa fitness and evolution during chronic infection, as well as the emergent properties of P. aeruginosa communities that are impacted by aggregate interactions. The observation that 2-pL–sized producer communities do not induce QS in neighboring aggregates as close as 5 µm ( Fig. 3 , Table 1 , and Datasets S1 and S2 ) contradicted our previous studies that trapped communities of this size induced QS in a responder population 8 µm away ( 11 ). The primary difference in these studies was that Connell et al. ( 11 ) used a common laboratory media (LB), while in this study we used SCFM2. In light of the significant impact of SCFM2 on diffusion ( Fig. 1 B ), it is not surprising that interaggregate communication is impacted by this new growth environment. Mucin had the most significant impact on diffusion and viscosity ( Fig. 1 ), and it has previously been shown to reduce QS induction in vitro ( 13 , 30 – 32 ), thus the lack of interaggregate signaling observed in SCFM2 is likely due to the presence of this polymer. It should also be noted that 2-pL P. aeruginosa aggregates are twice as large as the maximum size observed in CF lung tissue (∼1,000 cells) ( 5 ), suggesting that most aggregates within the CF lung are not of sufficient size to engage in interaggregate signaling in the CF lung. This finding has important consequences for understanding the evolution and ecology of P. aeruginosa in the CF lung, as social interactions including cooperation and cheating have been implicated as important ( 33 ), yet our data indicate that these interactions may be largely confined to individual aggregates. One caveat to our study is that the autoinduction (signal amplification) loop was removed in the responder cells as the goal was to assess 3OC12-HSL–mediated interactions originating from a single aggregate. This likely has a profound effect on aggregate response and calling distance as bacteria that are situated on the outer edges of an aggregate have the potential to respond first and QS induction of interior cells (and more distant aggregates) may be initiated by this response and feedback. Studying autoinduction is challenging experimentally, and future studies likely will require new experimental approaches to fully answer this question. Finally, while one of the primary strengths of our system is the ability to study a single signal-producing aggregate, a CF sputum sample containing 10 8 cells will have ∼10 5 signal producing aggregates. Thus, in addition to size, the geographical location of aggregates in relation to one another will no doubt impact communication. While our results support a model in which P. aeruginosa 3OC12-HSL signaling is primarily an intraaggregate phenomenon in CF based on the aggregate sizes observed in CF lung tissue, aggregate size may of course vary with ecology. Indeed, an examination of P. aeruginosa aggregate sizes in a murine surgical wound model during coinfection with Staphylococcus aureus revealed that P. aeruginosa aggregates are considerably larger than 2,000 cells (∼7,000 cells) and formed distinct monospecies aggregates, indicating that aggregate sizes sufficient for interaggregate communication in vitro are observed in some infection models ( Fig. S5 ). While it is possible that the physical characteristics of each infection setting (e.g., viscosity, chemical composition) may impact aggregate communication in unique ways, our aggregate size data suggests that the role of QS in coordinating activities may be different in some ecological settings. In the future we plan to adapt the methodology used here to study intra- and interaggregate QS signaling in animal models of P. aeruginosa infection. Our observation of a 3OC12-HSL “calling distance” that was similar for all trapped communities indicates that producer traps greater than 5 pL are able to signal to aggregates at distances (120–180 µm) that are at the outermost perimeters of our experimental system (200 µm). We were unable to define a calling distance more precisely, as it was necessary to bin the data into 60-µm intervals for statistical analysis ( Fig. 3 and Dataset S2 ). It should be noted that Fig. 3 does not imply a defined point at which communication stops, as responding aggregates were found throughout the 120- to 180-µm distance interval ( Fig. S3 and Dataset S2 ), although we can say with confidence that the outermost aggregate that responded in our experimental set-up was a single aggregate in one replicate at 176-µm outside a 20-pL trap. We predict that larger traps would signal at greater distances within the 120- to 180-µm region, and our results provide benchmark data for future experiments and modeling studies. It is important to point out that the CF lung presents a vast landscape in which populations of cells can develop, and considering the numbers of cells commonly retrieved from an expectorated sputum sample, it is not unfounded to consider that aggregates may be distributed hundreds of microns apart ( 34 ), thus communication over this distance is likely biologically important in the CF lung."
} | 1,941 |
40109662 | PMC11920157 | pmc | 88 | {
"abstract": "In view of the growing volume of data, there is a notable research focus on hardware that offers high computational performance with low power consumption. Notably, neuromorphic computing, particularly when utilizing CMOS-based hardware, has demonstrated promising research outcomes. Furthermore, there is an increasing emphasis on the utilization of emerging synapse devices, such as non-volatile memory (NVM), with the objective of achieving enhanced energy and area efficiency. In this context, we designed a hardware system that employs memristors, a type of emerging synapse, for a 1T1R synapse. The operational characteristics of a memristor are dependent upon its configuration with the transistor, specifically whether it is located at the source (MOS) or the drain (MOD) of the transistor. Despite its importance, the determination of the 1T1R configuration based on the operating voltage of the memristor remains insufficiently explored in existing studies. To enable seamless array expansion, it is crucial to ensure that the unit cells are properly designed to operate reliably from the initial stages. Therefore, this relationship was investigated in detail, and corresponding design rules were proposed. SPICE model based on fabricated memristors and transistors was utilized. Using this model, the optimal transistor selection was determined and subsequently validated through simulation. To demonstrate the learning capabilities of neuromorphic computing, an SNN inference accelerator was implemented. This implementation utilized a 1T1R array constructed based on the validated 1T1R model developed during the process. The accuracy was evaluated using a reduced MNIST dataset. The results verified that the neural network operations inspired by brain functionality were successfully implemented in hardware with high precision and no errors. Additionally, traditional ADC and DAC, commonly used in DNN research, were replaced with DPI and LIF neurons, resulting in a more compact design. The design was further stabilized by leveraging the low-pass filter effect of the DPI circuit, which effectively mitigated noise.",
"conclusion": "4 Conclusion In order to overcome the limitations of power consumption that are inherent to the traditional von Neumann architecture, which is characterized by a bottleneck, ASIC systems have been proposed. Among these, there is a particular need for research on the hardware accelerator in order to address the issue of bottlenecks. We put forth the proposition of SNN edge computing, wherein memristors are employed in a PIM capacity. We provide a detailed account of the operational and utilitarian aspects of the requisite components, including memristors, transistors, TIAs, and LIF neurons. In particular, we elucidated the distinctions between MOS and MOD configurations when integrating memristors and transistors into a 1T1R structure, emphasizing the challenges associated with conventional resistors and their categorization according to their set and reset behavior when employed as memristors. By delineating the selection criteria for suitable transistors for our memristors, we enhanced the comprehension of 1T1R configurations and furnished practical directives for implementation. Following a comprehensive examination of the attributes of individual devices, we devised a SPICE hardware simulation to emulate PyTorch simulations, thereby demonstrating that devices exhibiting conductance levels of 3 bits or more do not exhibit notable discrepancies in accuracy. Moreover, we addressed the implementation challenges posed by tuning errors, demonstrating that a tolerance within 5% enhances feasibility. The cumulative effect and low-pass filter functionality of the DPI circuit mitigated the intrinsic noise, allowing for up to 7% noise without significantly affecting accuracy. In addition to the superior hardware design, the benefits of SNNs, such as latency coding and reduced load due to temporal operation, were also leveraged. The elimination of the necessity for ADC and DAC resulted in a notable reduction in power consumption and enhanced resilience to noise. While PIM has not yet supplanted traditional computing, ongoing research into high-quality hardware and software technologies is anticipated to facilitate the deployment of memristor-based SNN analog computing.",
"introduction": "1 Introduction The exponential growth of data requires the development of efficient hardware systems that consume minimal power while operating at high processing speeds. The von Neumann bottleneck, which is characterized by the separation of processing units and memory, results in a considerable increase in power consumption due to the constant transfer of data between these two components. To address this limitation, a plethora of research has been conducted into and developments have been made in technologies such as ASICs and processing-in-memory (PIM) with the aim of enhancing operations within the von Neumann architecture. Nevertheless, in order to resolve these issues in a fundamental manner, there is an increasing necessity for research into neuromorphic computing, which represents a paradigm shift from the traditional von Neumann architecture ( Kemp, 2024 ). The deployment of these novel computational architectures presents a number of challenges in relation to throughput, latency and power budget when applied to existing hardware. It is therefore necessary to design specific hardware. The majority of research in this field is based on CMOS technology and can be broadly categorized into two main areas: studies focusing on artificial neural network (ANN) and studies focusing on spiking neural network (SNN). In research based on ANN, the technique of gradient descent is typically employed to adjust the loss, with backpropagation being the primary method for training. Hardware accelerators, specifically neural processing units (NPUs), are developed based on artificial neural networks (ANNs). Functioning between the CPU and memory, NPUs perform parallel processing and large-scale data handling, enabling the rapid processing of bottleneck data and significantly enhancing overall system performance. The development of these accelerator units has reached a point where they are not only utilized in commercial smartphones but also incorporated into laptops ( Tan and Cao, 2023 ; Feng et al., 2024 ). In contrast, research based on spiking neural networks (SNN) is primarily concerned with the development of processors that emulate the functionality of the human brain through processing of spatiotemporal spike patterns. Notable advancements have been documented, including the introduction of Intel’s Loihi chip and IBM’s TrueNorth chip. Both chips are designed to include over 1,000,000 neurons and more than 120,000,000 synapses per chip, representing an attempt to replace traditional computing architectures on a fundamental level ( Vogginger et al., 2024 ). Consequently, there have been continuous efforts to advance and implement neuromorphic computing leveraging CMOS technology. This technology involves the utilization of CMOS-based neuron circuits and synapses, typically implemented using SRAM or DRAM as the foundational synapse elements. However, in terms of area efficiency (GOPS/mm 2 ) and energy efficiency (GOPS/W), CMOS based neuromorphic technology generally offers a performance improvement of about one order of magnitude (approximately 10 times) compared to systems driven by GPUs based on conventional von Neumann architectures. The mean values reported in the literature indicate an area efficiency of approximately 300 GOPS/mm 2 and an energy efficiency of around 400 GOPS/W for the CMOS based neuromorphic systems ( Zhang et al., 2020 ). Reports indicate that the human brain contains over 10 13 synapses in the neocortex ( Tang et al., 2001 ). The synapse activity is estimated to occur between 10 13 and 10 16 times per second ( Merkle, 2007 ). When this activity is divided by the brain’s power consumption of approximately 25 W, the result is an energy efficiency of around 400,000 GOPS/W (based on 10 15 operations per second). Further research is required to enhance the area efficiency and achieve power efficiency at the level of the human brain, as well as to reduce volume through stacking. The current CMOS synapse-based approach to neuromorphic computing is characterized by high power consumption compared to emerging synapse-based neuromorphic computing. Additionally, it requires significant additional circuitry (e.g., ADCs, DACs), and most synapse devices are implemented using SRAM-based designs, which require at least six transistors, leading to limitations in terms of area ( Vogginger et al., 2024 ; Zhang et al., 2020 ). To overcome these limitations, studies exploring the use of emerging devices for both neurons and synapses have also been reported. A widely adopted approach involves replacing synapses, which account for a significant portion of area and power consumption, with emerging synapse devices, while neurons are commonly implemented using simplified CMOS, several studies on neuromorphic systems based on emerging synaptic devices have demonstrated area efficiencies exceeding 4,000 GOPS/mm 2 and power efficiencies surpassing 3,000 GOPS/W ( Zhang et al., 2020 ; Mochida et al., 2018 ; Xue et al., 2019 ). Memristors can be classified into several categories, including phase change memory (PCM) ( Fong et al., 2017 ; Burr et al., 2010 ), magnetic random-access memory (MRAM) ( Burr et al., 2010 ; Tehrani, 2006 ), ferroelectric random-access memory (FeRAM) ( Chen et al., 2020 ), and resistive random-access memory (RRAM) (of which there are several subcategories, including interface-type RRAM, VCM, and ECM) ( Chen, 2020 ; Ryu et al., 2020 ). RRAM is distinguished by its stable operation, on/off ratio, speed, and high compatibility with complementary metal-oxide semiconductor (CMOS) technology ( Raghavan, 2014 ; Kim et al., 2021 ). RRAM offers a number of advantages over traditional DRAM or SRAM, including low power consumption, high operational speed, the ability to store multiple bits of data, and the elimination of the need for refresh, which allows for the construction of large-scale matrices ( Dogan, 2013 ; Perez and De Rose, 2015 ). Memristors are typically organized in crossbar arrays, wherein each memristor represents a weight value in the matrix. However, crossbar arrays are susceptible to sneak path currents due to Kirchhoff’s law, which has the potential to compromise the accuracy of the network. In order to mitigate the impact of sneak path currents, it is common practice to employ 1T1R structures incorporating transistors ( Youssef et al., 2021 ; Pan et al., 2024 ). Memristor-based artificial neural networks (ANN) have been widely documented as hardware accelerators for the recognition and inference of MNIST patterns ( Mochida et al., 2018 ; Xue et al., 2019 ; Adam et al., 2016 ; Prezioso et al., 2015 ). Extensive validation by numerous researchers has also reported the fabrication and validation of memristor chips that are capable of being applied to real-world tasks, including speech recognition, image classification, and motion control ( Zhang et al., 2023 ; Ambrogio et al., 2023 ). In contrast, memristor-based spiking neural networks (SNN) concentrate on the implementation of innovative neuron structures with the objective of further reducing system power consumption, with the ultimate goal of developing highly efficient and applicable hardware. The application of research on memristor-based SNN chips has been constrained, with the majority of efforts only achieving MNIST inference ( Valentian et al., 2019 ). This highlights the necessity for further investigation into the circuitry architecture and algorithms associated with the relevant hardware ( Bouvier et al., 2019 ). A significant number of studies employ transistors for the purpose of suppressing sneak paths and acting as selectors. It is therefore essential to exercise caution when selecting transistors for use with memristors, considering the memristor’s operating voltage, resistance characteristics, and the transistor’s current characteristics and on-resistance. The operational characteristics of the memristor are dependent upon whether it is attached to the transistor on the source side (memristor-on-source, MOS) or the drain side (memristor-on-drain, MOD). This must be considered when designing the circuit. The 2T2R structure consists of two supply voltage lines (each connected to an electrode of the memristor), two gate lines, and a shared source terminal. In this configuration, weights can be implemented with greater flexibility, allowing for the representation of both positive and negative weights. Specifically, one memristor in the 2T2R structure is designated to represent positive weights, while the other represents negative weights. As a result, during weight evaluation, the combined weight is obtained by summing the values of both memristors. In typical implementations, the currents corresponding to positive and negative weights are processed through differential amplifiers or similar circuits, leading to power consumption from both currents. However, the 2T2R structure leverages the opposing directions of the net current flow resulting from the combined weights, allowing the net current to flow directly. This characteristic provides a power-saving advantage over the 1T1R structure, particularly in large-scale neural network implementations ( Zhang et al., 2023 ). Nevertheless, modifying and operating the weights of individual devices in the 2T2R structure requires more complex algorithms. As a result, state-of-the-art research often adopts a hybrid approach, utilizing 2T2R structures for large-scale networks or computationally simple operations and 1T1R structures for regions requiring precise operations. Such hybrid implementations have been reported in recent studies ( Zhang et al., 2023 ). Implementing neural network arrays with memristors involves numerous considerations, and research focused on the design of transistor-memristor interactions is essential to address these challenges effectively. The operation of a well-designed transistor-memristor array is influenced by several factors, including the accuracy of memristor conductance mapping, which significantly affects the final results as tuning error. In addition to programming errors, intrinsic noise is a major factor that reduces accuracy in neural networks ( Zeng et al., 2023 ; Huang et al., 2023 ; Park et al., 2021 ). It is therefore imperative that circuit design techniques which serve to minimize the influence of these factors are employed. As the initial step in optimizing the design of a complete transistor-memristor synapse, it is essential to consider the operating voltage levels of individual memristors and transistors. Consequently, the objective was set to design CMOS devices, known for their higher technological maturity, to align with the operational requirements of memristors. To achieve this, a methodology was proposed to optimize the 1T-1R configuration by designing transistors with variable W/L ratios, thereby enabling the adaptation of transistor characteristics to meet the specific needs of the memristor-based system. An additional consideration involves determining the optimal orientation for attaching the memristor to the transistor. This methodology was utilized to examine the differences and impacts between MOS and MOD configurations, providing insights into the most effective design approach. To further investigate these effects, a compact model was developed with characteristics identical to those of the fabricated memristor. This was used in conjunction with a design of SNN hardware, including a 1T1R array, a differential-pair-integrator (DPI) synapse circuit, and a leaky-integrate-fire (LIF) neuron, to implement an inference accelerator in a circuit level. SNN simulations by SPICE were conducted on the designed memristive neural networks of a small scale (8 × 8), considering both tuning errors and intrinsic noise. The findings of this study thus lead to the proposal of an optimal design for noise-tolerant memristive-SNN hardware, and to the demonstration of the advantages of using SNN for high-efficiency computing in comparison to ANN.",
"discussion": "3 Results and discussion 3.1 Multi-level memristor device Figure 2A shows a cross-sectional view of a single unit memristor device illustration intended for use in the memristive neural network and its resistive switching. When a positive bias is applied to the top electrode, Cu migrates toward the bottom electrode to form a filament. Conversely, when a negative bias is applied to the top electrode, the Cu filament dissolution occurs and migrates back toward the top electrode. The existence of the Te interfacial layer restricts the Cu migration path, enabling stable switching behavior ( Goux et al., 2011 ). In addition, the alloy of Cu and Te efficiently suppresses excessive Cu migration, improving endurance ( Tseng et al., 2018 ). The gradual resistance change characteristic is essential for implementing multi-conductance levels. In a multi-level state, each resistance state must be precisely defined. Gradual switching characteristics allow fine modulation of resistance through gradual resistance changes, rather than abrupt resistive switching characteristics, making it easy to set a specific desired resistance state and implementing various intermediate resistance states. Therefore, IGZO based Cu:Te device exhibits gradual resistance change behavior because it exhibits multi-weak filament characteristics rather than strong single filament. The number of pulses required to transition from the initial conductivity state to the minimum and maximum conductivity. During the electro-forming process of the device, the presence of IGZO, a buffer layer, creates a heat confinement effect so multi filaments are induced within the switching layer ( Gao et al., 2017 ). At this time, multiple filaments are formed sequentially, resulting in gradual switching behavior. If only a partial reset is achieved by reducing the reset voltage rather than fully resetting (−2 V), the conductance level can be modulated by controlling the number of multi filaments. Figure 2 (A) Schematic illustration of single memristor device, and I–V characteristic result for each reset voltage sweep range. The dash lines are fully set 2 V or fully reset −2 V sweep curves, respectively. The blue to red lines are sweep curves formed by increasing the reset stop range after set. (B) Conductance level distribution at read voltage 0.05 V for each reset voltage sweep range. (C) Retention and noise behavior at 0.05 V read voltage in 100 s for each conductance level. The colors of each conductance level correspond to the V reset stop colors in (A) . (D) Relative standard deviation of the (C) result. In order to show analog multi-level characteristics in IGZO-based Cu:Te devices, applying negative voltage ( V reset ) sweep up to specified conductance values and presented in Figures 2A , B , respectively. Due to the gradual reset behavior of the device, conductance can be modulated at various negative voltages. The multi-level formation process is as follows. First, a −0.5 V reset was performed to form the initial level ( G 0 ), and then a reset operation was performed by adding δV to the current reset voltage when forming the next level ( G 1 ). At this time, the initial δV value was −1 mV, and to clearly distinguish between levels considering the C2C and D2D variations when forming the next level, a 3-s read operation was performed. If the conductance difference between the current level ( G n ) and the next level ( G n + 1 ) was less than 0.2 μS (ΔG = G n, min – G n + 1, max < 0.2 μS) during the 3-s read operation, the reset process was performed again by increasing −1 mV from the current δV. This level formation process was performed until the fully reset voltage of 2 V was reached without a set process. Figure 2C show the multi-level behavior obtained in Figure 2A through a read operation (0.05 V) for 100 s. As shown in Figure 2C , 23 multi-level states were formed in 5.7–200 μS due to various reset stop sweep operations (−0.5 V to −2 V). Each conductance level obtained through various reset stop sweep operations (−0.5 V to −2 V) was maintained constant without degradation, and the interval between levels was modulated to be at least 0.2 μS. Figure 2D shows the relative standard deviation (RSD) of the multi-levels obtained through the read operation. As the reset stop voltage increases and the conductance level decreases, the RSD tends to increase. This means that random telegraph noise (RTN) intervention is different for each conductance level and the number of multi-level states is modulated by IGZO-based Cu:Te devices ( Veksler et al., 2013 ). Therefore, if the conductance level is low, the number of conductive filaments in the switching layer and the number of Cu atoms in the constriction area of each filament will decrease. Therefore, the probability of electrons being trapped/de-trapped by the charged instable filament around the conductive filament will increase ( Belmonte et al., 2014 ; Rao et al., 2023 ). 3.2 Design of 1T1R synapse and its operation for multilevel conductance tuning The attachment of a memristor to a transistor through BEOL processing results in the formation of the structures shown in Figure 3A , wherein the resistor is situated either at the transistor’s drain or source ( Ghenzi et al., 2018 ; Maheshwari et al., 2021 ). These configurations are designated as “memristor on drain” (MOD) and “memristor on source” (MOS), respectively, as illustrated in Figures 3B , C . Figure 3 Panel (A) presents the schematic of a memristor fabricated through the BEOL process after transistor design. Panel (B) shows the 1T1R schematic in the MOD set and MOS reset conditions. Panel (C) illustrates the 1T1R schematic in the MOD reset and MOS set conditions. The transistor, acting as a selector, ideally must transmit the full supply voltage ( V DD ) to the memristor. Additionally, it must possess an off-resistance greater than that of the memristor to effectively suppress sneak path currents. While a transistor is, in theory, capable of functioning as a switch without resistance, a number of practical considerations must be taken into account. These include the operating regions of the transistor, which may be either a triode or in saturation. One such factor is the attachment orientation of the memristor. In the case of a MOD configuration with a positive supply voltage, where the bottom electrode of the memristor is attached to the NMOS’s drain terminal, the supply voltage V DD is divided into both of the transistor and the memristor depending on their resistances. From the memristor’s point of view, this results in a loss of the supply voltage due to the voltage drop across the transistor. Accordingly, a transistor with an appropriate on-resistance should be selected based on the current required for memristor operation. The memristor, a variable resistor whose resistance changes with voltage, is represented as a resistive element in the 1T1R structure illustrated in Figure 3A . The memristor, while often simplified as a fixed resistor, is in fact a bipolar device with two terminals: an anode and a cathode. The 1T1R operation conditions depend on both the memristor attachment configuration and the bias polarity, rather than being a simple transistor-resistor relationship. Specifically, the set behavior of the memristor in the MOD configuration and the reset behavior in the MOS configuration exhibit structural and operational symmetry as illustrated in Figure 3B . Similarly, the reset behavior in the MOD configuration and the set behavior in the MOS configuration also demonstrate identical voltage application methods and structural characteristics, as depicted in Figure 3C . Effective memristor operation requires facilitation of both set and reset operations. In the fabrication of 1T1R devices, achieving a monolithic structure is crucial to eliminate unnecessary auxiliary circuits and simplify the design. Since the orientation of the memristor (MOD or MOS) is predetermined during manufacturing, careful selection of the attachment orientation is essential to avoid operational interference ( Liu et al., 2024 ; Bengel et al., 2023 ). Before explaining the differences caused by the attachment direction, it is important to note that in a typical 1T1R configuration, the voltage drop across the memristor typically exceeds the voltage drop over the transistor (V DS ) when the transistor operates in the triode region. The following discussion is based on this condition. In the MOD configuration, the applied V DD voltage is applied to the memristor with a negligible transfer loss if the gate-to-source (V GS ) is over the transistor threshold voltage (V th ). Conversely, in the MOS configuration where the memristor is attached to the source of the transistor, applying a high voltage to the V DD pad and grounding the BE (bottom electrode) of the memristor (MOS set case) results in the transistor’s source voltage (V S ) equating to the memristor’s top electrode voltage (V TE ). Consequently, V GS becomes V G − V memristor , requiring a higher gate voltage to turn on the transistor. When the gate voltage is V DD (V G = V DD ), the maximum voltage drops across the memristor are V DD – V th . These relationships differ depending on whether a memristor is in the set or reset state, as summarized in Table 1 . If the memristor requires a higher set voltage than a reset voltage, the MOD configuration, which minimizes a voltage transfer loss across the transistor during the set operation, is advantageous. Conversely, if a higher reset voltage than a set voltage is required, the MOS configuration, which minimizes a voltage loss during the reset operation, is preferred. Therefore, research groups designing and fabricating 1T1R arrays must determine whether to use the MOS or MOD configuration in advance. For our Cu:Te-based memristor devices, which exhibit a gradual set characteristic, a set voltage of up to 2 V is required. As demonstrated in Figures 2A , C , analog states were achieved under full set conditions by adjusting the reset stop voltage. Hence, fully setting the memristor is an essential requirement. To meet this condition, we adopted the MOD configuration, which minimizes the voltage transfer loss over the transistor during the set operation. After determining the 1T1R configuration, the operation conditions analysis for the memristor switching was performed based on the W/L ratio and the resistance of the memristor. The detailed analysis results are provided in Supplementary Figure S4 . Table 1 The table provides a summary of the maximum voltage that can be applied to the memristor ( V memristor ), determined by the 1T1R configuration. Voltage on memristor ( V memristor ) SET RESET Memristor on drain (MOD) V DD V DD – V th Memristor on source (MOS) V DD – V th V DD Considering these findings, the MOD configuration was selected as the optimal operating mode, and for ease of set and reset operations, the device was designed with a W/L ratio of at least 65. To confirm the correct operation of the 1T1R device, we constructed it by depositing the BE on the drain pad, followed by the switching layer, and finally the TE pad. The optical microscope (OM) images of the fabricated device are presented in Figures 4A , B . Figure 4A depicts the overall view, while Figure 4B shows a magnified view of the switching layer. The fabricated device was tested by sweeping V TE while the gate voltage was set to 3 V for both set and reset operations. As illustrated in Figure 4C , when the W/L ratio is at least 65, the set behavior is comparable to that of the memristor alone, and the reset operation is also performed smoothly, achieving an on/off ratio of approximately 87.5% in comparison to the unit device. The 10% loss can be attributed to the on-resistance ( R on ) value of 120 Ω for a width-to-length (W/L) ratio of 65, which results in a 12% voltage drop across the transistor relative to the memristor’s low-resistance state (LRS) of 1 kΩ. It is therefore proposed that a transistor with an on-resistance within 10% of the memristor’s resistance will facilitate optimal operation. In light of the considerable increase in size for a larger transistor model (W/L ratio of 330) and the flexibility of operating the transistor in triode and saturation modes with a W/L ratio of 65, the decision was made to select a transistor with a W/L ratio of 65. The IV results for the memristor on source configuration and the on/off ratio variations with W/L changes are presented in the Supplementary Figure S5 and Supplementary Table S2 . Supplementary Figure S6 illustrates the analogous IV curve obtained through the utilization of the compact model for 1T1R measurements, encompassing both MOS and MOD configurations. Figure 4 Panel (A) shows an optical microscope (OM) image of the monolithic 1T1R device fabricated between the Drain pad and V DD pad through the BEOL process, with a scale bar representing 100 μm. Panel (B) presents a magnified OM image of the region where the top electrode (TE) and bottom electrode (BE) of the memristor intersect. The switching layer and buffer layer are marked in orange and green, corresponding to TaOx and IGZO, respectively. The scale bar represents 50 μm. Panel (C) displays the V DD voltage vs. Drain current IV curve for the monolithic 1T1R device, as previously shown in the OM images, across varying W/L ratios. The gate voltage is fixed at 3 V during the set and reset measurements. The structure of the NMOS and memristor used for conductance tuning is shown in Figure 5A . The gate voltage was fixed at 3 V, the source voltage was grounded, and the top electrode voltage was varied during the process. A closed-loop conductance tuning procedure was conducted using the 1T1R cell in the MOD configuration with a W/L ratio of 65. The fundamental algorithmic process is depicted in Figure 5B . By applying an initial voltage of 2 V to fully set the memristor and then adjusting the initial reset voltage based on the unit memristor’s conductance, the target conductance was successfully identified. With a ΔV of 0.01 V and an error range of 3%, nine conductance tunings were achieved within 55 pulses, as illustrated in Figure 5C . However, when the error range was decreased to 1%, the system was unable to identify the target conductance, exhibiting a repetitive switching between set and reset states, as illustrated in Figure 5D . This outcome suggests that the intrinsic noise of the unit memristor, estimated to be approximately 2%, may have hindered the successful identification of the target value within the specified 1% error range. Furthermore, even if the target value were to be successfully identified, the intrinsic noise would likely introduce instability, necessitating repeated loop executions. Figure 5 Panel (A) shows the 1T1R schematic of the MOD structure for conductance tuning, where the gate voltage is fixed at 3 V, and the V DD voltage is adjusted. Panel (B) illustrates the voltage profile applied to V DD for conductance tuning. A voltage of 0.05 V is used to read the current conductance state, which is then compared with the target value. Based on this comparison, either a potentiation pulse or a depression pulse is applied, and the process is repeated. The amplitude of the potentiation and depression pulses is continuously varied by Δ V (0.01 V) until the target conductance is reached. The initial potentiation pulse voltage is 2 V, while the depression pulse voltage is set according to the conditions in Figure 2A based on the target conductance. Panel (C) demonstrates that the closed-loop conductance tuning algorithm successfully achieved 9 target conductance values within a 3% error margin using 60 pulses or fewer. Panel (D) shows that when the error margin is tightened to 1%, the algorithm fails to achieve the target conductance, even after more than 150 pulse iterations. 3.3 Circuit-level design of a neuron-synapse-neuron unit for spiking neural networks In accordance with the previously established 1T1R configuration, a verification process was undertaken at the unit level of neuron-synapse-neuron (N-S-N) prior to the execution of the comprehensive network simulation. As illustrated in Figure 6 , the single N-S-N network circuit was implemented using a 1T1R synapse, a TIA circuit, a DPI circuit and a LIF neuron circuit. The DPI circuit was introduced to emulate a dynamic synapse current behavior. Biological neurons transmit information through electrical or chemical synapses ( Petersen, 2016 ). Electrical synapses directly connect and allow current flow between two neurons through gap junctions. However, the most common synaptic mechanism is chemical synapses ( Pereda, 2014 ). In chemical synapses, the generated spike signal antidromically propagates to the axon terminal, triggering synaptic vesicle exocytosis and subsequently release neurotransmitters. When the released neurotransmitters cross the synaptic cleft and bind to postsynaptic receptors, postsynaptic ion channels such as AMPA or GABA receptors open. This alters the ionic permeability such as Na + , Ca 2+ or Cl − , and subsequently depolarizing or hyperpolarizing the membrane potential of the dendrites forming synapse. Since these steps are highly dynamic due to chemical diffusion and reaction of neurotransmitters, a synaptic response model that evokes postsynaptic current using a unit function input without considering any synaptic current cannot accurately describe the synaptic response in the postsynaptic neuron. Moreover, even when employing a synaptic current model of a single exponential model that only considers the decay phase of postsynaptic current fails to fully capture the rising dynamics synaptic current ( Rothman and Silver, 2014 ). Therefore, in most studies, postsynaptic responses are commonly described using a double exponential, where one exponential for the rising phase and another for the decay phase of the synaptic response ( Beniaguev et al., 2021 ; Jang et al., 2020 ; Tikidji-Hamburyan et al., 2023 ). The double exponential synapse dynamic behavior can be emulated in numerical simulation using the following equations of discrete forms: \n (1) \n Normalize _ factor N F = τ s y n decay τ s y n decay − τ s y n rise \n \n (2) \n I s y n _ new rise = τ s y n rise · I s y n rise + S t · W n · N F \n \n (3) \n I s y n _ new decay = τ s y n decay · I s y n decay + S t · W n · N F \n \n (4) \n I s y n _ total = I s y n _ new decay − I s y n _ new rise \n \n (5) \n V m e m _ new = τ m e m · V m e m + I s y n _ total \n Figure 6 The circuit structure that mimics biological behavior in a Neuron-Synapse-Neuron configuration consists of four parts: the Memristor part, TIA part, DPI synapse part, and Post-neuron part. The Memristor part features a 1T1R structure that converts the input spike ( V input, spike ) into a current spike based on the memristor’s weight. In the TIA part, the current spike is converted into a voltage spike signal. The DPI synapse part processes the voltage spike, incorporating the weight, and converts it into synaptic current ( I syn ) following the double exponential rule through the charging and discharging of C syn , resulting in changes in V syn . Finally, in the Post-neuron part, I syn drives the charging and discharging of C mem , leading to the membrane potential of the neuron, represented as V membrane . The application of the double exponential model requires the definition of the normalize_factor in accordance with the specifications set forth in Equation 1 . The calculation of the rise and decay synaptic currents is performed by considering τ syn(rise) and τ syn(decay) through Equations 2 , 3 , respectively. Equations 2 , 3 are composed of different terms, including 𝜏 syn , S(t), and W n and normalize_factor. S(t) represents the spike generated when the membrane potential exceeds the threshold, indicating the occurrence of a spike under that condition. The W n value corresponds to the weight, and it can be observed that a higher weight leads to a larger current flow, even for the same spike, depending on the equation. The total synaptic current, 𝐼 syn_total , is then obtained using Equation 4 , which accounts for the time constants of both the rise and decay phases. As outlined in Equation 5 , the membrane potential rises in response to the synaptic current and decays in accordance with the membrane time constant, 𝜏 mem . To achieve emulation of the double exponential synapse behavior in the neuron-synapse-neuron (N-S-N) network, adjustments were made to the tau-related gate voltages in the DPI circuit, ensuring that the current variation corresponding to the memristor’s state remained consistent. Upon the application of input spikes to the gate of the NMOS transistor, the current flows in accordance with the conductance of the memristor, with V read set to 50 mV. The current is then amplified by the TIA using operational amplifier 1, and the inverted output voltage is applied to the gate of the MN3 transistor in the DPI circuit via operational amplifier 2. In 1989, Mead put forth a circuit that emulates a pulsed current source synapse, whereby synapse current is conducted when pulse signals are applied. This circuit has undergone continuous improvement, with the current form of the DPI proposed by Bartolozzi and Indiveri (2007) , fabricated, and verified using foundry processes. The voltage input to the MN3 transistor in the DPI circuit, which includes the memristor’s weight value, generates the total current, I tot . This current is the result of the discharge of C syn , with the amount of discharge varying in accordance with the magnitude of the signal at the gate of MN3. A change in the voltage applied to the MP2 transistor, V syn , will result in a corresponding alteration of the gate voltage, which in turn will affect the current flowing through the synapse, I syn . The decay time (τ) of I syn and the charging time of V syn can be modified by adjusting the V syn,tau value at the gate of the MP1 transistor, thereby controlling I tau . A portion of the generated I syn flows to the ground through MN5, while the remainder charges C mem . The voltage across C mem ( V membrane ) represents the post-neuron’s membrane potential, and the membrane τ can be adjusted by controlling the gate voltage V mem,tau of MN5. Consequently, when a spike signal occurs in the 1T1R structure, a voltage signal incorporating the memristor’s weight value is generated through the utilization of the TIA. This signal is then converted into the I syn current via the DPI circuit, which discharges C mem , thereby altering the V membrane signal. From a mathematical perspective, the synapse exhibits both a charge time constant and a discharge time constant. The membrane potential rises in accordance with I syn , excluding the influence of the discharge current I D0 caused by the MN5 transistor. The operation of the circuit can be mathematically described by Equations 6 , 7 , which account for the charge and discharge of the synapse. Ultimately, the change in membrane potential is expressed by Equation 8 . \n (6) \n I s y n t = I gain I tot I τ 1 − e − t − t i τ s y n + I s y n e − t − t i τ s y n \n \n (7) \n I s y n t = I s y n e − t − t i τ s y n \n \n (8) \n I gain = I 0 e − k V t h r − V d d u T at P M O S ’ s subthreshold \n \n (9) \n Membrane : d V C d t = I s y n C − I D 0 C exp V C − V G − | V t h | k u T \n In practice, factors such as the subthreshold slope factor ( k ) and the thermal voltage ( u T ) were employed due to the use of actual transistors ( Streetman and Banerjee, 2000 ). The aforementioned set of equations indicates that the dynamic rise and fall behaviors of the synapse current, as described by Equations 6 , 7 , and the membrane potential Equation 9 in the form of the LIF neuron, can provide a similar operational output to that simulated in PyTorch, provided that the appropriate parameters synapse current tau ( τ syn ), membrane potential tau (according to I D0 ), V th , C mem are selected. 3.4 Implementation of memristive spiking neural networks for inference As stated above, the dynamic spiking neural network behavior can be simulated in PyTorch using the set of the equations ( Equations 1–5 ) and emulated in SPICE using the hardware shown in Figure 6 , respectively. With the proper choices of the parameters of the circuit, the spiking neural network behavior of the circuit can be matched to that of PyTorch simulation. Input spike signals fired at 10, 110, 150, and 200 ms as shown in Figure 7A were used both of the PyTorch and SPICE simulations. In PyTorch, the time step is defined in 1 ms increments, resulting in an impulse-like firing structure. In SPICE, the signal was generated as a pulse with a rise time of 100 μs, a fall time of 100 μs, and a pulse width of 900 μs. When these pulses were applied, the changes in I syn were observed as shown in Figure 7B . In PyTorch, τ rise was set to 0.5 ms and τ fall was set to 2.0 ms. To mimic this in SPICE, V syn,tau was set to 1.44 V and C syn was set to 260 pF. Lastly, Figure 7C is provided to verify the accuracy of the following pattern, the membrane potential in PyTorch used τ mem of 15 ms, while in SPICE, V mem,tau was set to 0.1 V and C mem was set to 260 nF. The result showed a time error of approximately 2.1% relative to the maximum potential in the membrane potential, achieving a satisfactory match. Figure 7 The following results were obtained from both PyTorch and SPICE simulations, showing the voltage spike signal, synaptic current, and membrane potential. Panel (A) illustrates the shape of the voltage spikes that occur at 10, 110, 115, and 120 ms in both PyTorch and SPICE simulations. Panel (B) shows the changes in synaptic current in response to the spike events in the PyTorch simulation, where 𝛕 rise was set to 0.5 ms and 𝛕 fall was set to 2.0 ms. Panel (C) illustrates the resulting membrane potential graph, with 𝛕 membrane set to 15 ms (converted to arbitrary units for relative comparison). The prepared components were employed in the performance of pattern recognition, which is an exemplar of edge computing. An 8 × 8 handwritten digit image with 16 intensity levels, as illustrated in Figure 8A , was employed, and the values were transformed through latency coding. \n (10) \n Latency coding : t max · ln x x − x t h r \n Figure 8 (A) The input data is a reduced 8×8 MNIST image with 16 intensity levels representing digit patterns. (B) Based on the intensity of the input image, the spike firing time is calculated using the Equation 10 , and the neurons that fire at each time step are plotted. (C) To verify the accuracy of MNIST pattern recognition in PyTorch, the signals are converted using latency coding. These signals are fed into an input neuron layer with 64 inputs, 640 weights, and 10 output neurons. The accuracy is then assessed based on the membrane potential of the output neurons. (D) In SPICE, the spike signals from 64 input neurons are applied to a 1T1R gate. The current, modulated by the read voltage and the memristor conductance, is converted to synapse current through a DPI circuit, and the final output neuron’s membrane potential is used to evaluate the accuracy. The equation for latency coding is presented in Equation 10 . Latency coding employs t max as the maximum value, with the firing time determined when the pixel intensity, x , surpasses the threshold ( x thr ). Upon applying a t max of 20 and a thr of 0.3 to the image in Figure 8A , the neurons firing at each time step are determined, as illustrated in Figure 8B . Subsequently, the latency-coded spike signals were introduced as input into the PyTorch neurons, as illustrated in Figure 8C . Subsequently, the signals were conveyed through a 64 × 10 configuration of weights to the output neurons. By examining the membrane potential signals within the output neurons, we were able to verify whether the neuron corresponding to the correct label exhibited the highest membrane potential. Upon completion of the PyTorch simulation, the 64 × 10 weights were extracted and input into the normalized state variable parameters (ranging from 0 to 1) of the SPICE memristor model. Upon inputting the latency-coded spikes as gate voltages to the 1T1R devices, a net current was generated by V read , set at 50 mV. The current was then converted into a voltage signal by the TIA and subsequently input as the gate voltage to the DPI circuit. Based on the input signal, I syn was generated, and ultimately, the maximum value of the membrane potential of the LIF neuron over a 100 ms time period was employed as the criterion for verifying accuracy. The SPICE implementation structure is depicted in Figure 8D . As illustrated in Figure 9A , the simulation outcomes demonstrate the precision outcomes as a function of the number of bits. The results demonstrate that for both PyTorch ANN and PyTorch SNN, as well as SPICE SNN, the accuracy reaches a saturation point at 3 bits or more. Notably, both PyTorch SNN and SPICE SNN exhibit a saturation accuracy of 90%. This finding is consistent with other literature on bit dependence, indicating that a certain number of bits beyond a threshold are necessary, but not unlimited. The discrepancy in accuracy between PyTorch SNN and SPICE SNN at 2 bits and 3 bits can be attributed to the differing sizes of the training datasets. The PyTorch SNN utilized 1,124 images from the training data set, whereas SPICE simulations were conducted on only 100 images due to time constraints, resulting in a sampling bias. In light of the fact that the conductance results obtained from the 1T1R devices yielded nine distinct states, it may be reasonably assumed that accuracy should not differ significantly with more than three bits of conductance states. Figure 9B illustrates that, although not observed in PyTorch SNN, real hardware implementation demonstrated a range of tuning errors due to the inability to achieve the target conductance with precision. The aforementioned tuning error affects the accuracy of the system, with a tolerance of up to 5% exhibiting no significant decline in accuracy. However, beyond a 10% tuning error, there is a notable reduction in the accuracy of the system. Although the absolute accuracy is lower for a one-bit binary representation, it demonstrates a higher tolerance to tuning errors. Figure 9C examines the influence of intrinsic noise on accuracy without constraints on bit precision or tuning error, with a 3% tuning error from our device. In the SPICE model, noise was defined as a time-dependent function, with values ranging from 0 to 7% relative to a 0% noise baseline. Although intrinsic noise has a slight impact on accuracy, the overall system demonstrates tolerance, as the cumulative effect of the DPI circuit and its function as a low-pass filter effectively suppresses the noise ( Bartolozzi et al., 2006 ). Figure 9 We summarize the accuracy and sparsity of pattern recognition based on the PyTorch network and SPICE circuit simulations. (A) The accuracy was evaluated as a function of the synapse bits. (B) A graph was generated by performing SNN simulations in SPICE, accounting for tuning error (i.e., mapping accuracy) and synapse bits. (C) PyTorch and SPICE simulations were conducted, considering the combined effects of intrinsic noise, a 3% tuning error, and varying synapse bits. Our experimental results from measured device data, with 3-bit precision, 3% tuning error, and 5% intrinsic noise, achieved 90% accuracy. (D) A comparison table between latency-coded SNN and DNN highlights the spike firing sparsity. While variations exist depending on the dataset, all three representative examples show spike sparsity within 40%. These results are in line with the study of chemical synapses, which show that despite the presence of intrinsic noise, chemical synapses which our DPI circuit mimics enhance the system’s coherence through the selective reduction of unnecessary correlations, thereby suggesting more robust and reliable information processing compared to electrical synapses ( Balenzuela and García-Ojalvo, 2005 ). The LIF neuron exhibits low-pass filter behavior due to the cumulative effects of the RC circuit. However, parameter tuning can suppress the neuron’s operation. To address this, we implemented a tunable and stable low-pass filter for noise attenuation using a differential pair integrator (DPI) circuit. Additional data, presented in Supplementary Figure S7 and Supplementary Table S3 , demonstrates the impact of intrinsic noise on the membrane potential. Despite a noise level of 7%, the peak difference in membrane potential is approximately 3%. Consequently, as shown in Figures 2D , 6 , the proposed Cu device exhibits intrinsic noise within 5% and a tuning error around 3%, allowing for the implementation of analog behavior with a precision exceeding 3 bits. Additional detailed results for bit count and tolerance are presented in Figure 9C . As a result, a final inference accuracy of 90% can be achieved. In terms of intrinsic noise and tuning error, the array using only memristors demonstrated greater stability compared to traditional configurations ( Park et al., 2022 ). In previous DNN-based research, significant accuracy degradation was observed due to system tuning errors and the intrinsic noise of the devices (such as RTN noise). Under conditions similar to ours, with 5% intrinsic noise and 3% tuning error, systems with conductance below 80 μS showed an accuracy decrease of over 20% ( Park et al., 2022 ). However, in our research, we utilized a high-performance memristor that suppresses intrinsic noise to within 2% even in the conductance range below 100 μS. Even when assuming 5% noise in simulations, we constructed a noise-tolerant inference system by leveraging the cumulative effects of the DPI circuit to attenuate noise. Additionally, we employed a 1T1R configuration to suppress sneak path currents, thereby preventing the overlap of errors and noise during actual operation ( Youssef et al., 2021 ). This study builds upon the design of SNN inference accelerator for power efficiency, extending it to the implementation of SNN edge computing functionalities. This is demonstrated through the reduction of MNIST 8 × 8 simulations. In a SNN-based edge computing system employing latency coding, spikes from low-intensity pixels below the threshold are not processed. The spike sparsity of latency coding is demonstrated in Figure 9D . Spike sparsity refers to the average number of neurons that fire per image. When the operation of all neurons in a DNN is considered 100%, the SNN, due to latency coding, shows a spike sparsity of less than 40% across datasets such as Reduced MNIST, MNIST, and Fashion MNIST, although the exact sparsity varies depending on the dataset. Furthermore, the analysis process and results were reflected in Supplementary Figure S8 through power analysis of the SPICE circuit. Although the circuit does not exhibit the highest power efficiency, we compared its power efficiency with that of research from other groups outside the state-of-the-art level. Although the actual simulation was conducted using a single-layer neural network, the power consumption was calculated based on a more complex multilayer structure. The network for the Reduced MNIST (8 × 8) dataset comprised 64 input neurons, 21 hidden neurons, and 10 output neurons. The network was composed of 28 × 28 input neurons, 256 hidden neurons, and 10 output neurons for the MNIST and Fashion MNIST datasets. In memristive neural networks, power consumption is primarily attributed to the access of memristors by spikes. With sparsity levels within 40% in the input layer and within 6% in the hidden layer, the efficiency increases as the number of layers grows. Previous studies have also reported a reduction in spikes in multilayer structures ( Chowdhury et al., 2022 ; Dampfhoffer et al., 2022 ). Furthermore, SNNs benefit from sparse input signals, which reduces the current burden on the driving circuit and enables temporal operation. This allows for intermittent inference and lower idle power consumption due to event-driven operation. From a power and circuit perspective, memristor-based deep neural network research has revealed considerable power consumption and noise susceptibility in analog-to-digital converter (ADC) and digital-to-analog converter (DAC) components ( Moro et al., 2022 ). In contrast, the use of DPI and LIF neurons eliminates the necessity for an ADC and a DAC, thereby offering a distinct advantage ( Li et al., 2023 )."
} | 13,326 |
28644896 | PMC5482491 | pmc | 89 | {
"abstract": "Bacteria in a biofilm colony have the capacity to monitor the size and growth conditions for the colony and modify their phenotypical behaviour to optimise attacks, defence, migration, etc. The quorum sensing systems controlling this involve production and sensing of diffusive signal molecules. Frequently, quorum sensing systems carry a positive feedback loop which produces a switch at a threshold size of the colony. This all-or-none switch can be beneficial to create a sudden attack, leaving a host little time to establish a defence. The reaction-diffusion system describing a basal quorum sensing loop involves production of signal molecules, diffusion of signal molecules, and detection of signal molecules. We study the ignition process in a numerical solution for a basal quorum sensor and demonstrate that even in a large colony the ignition travels through the whole colony in a less than a minute. The ignition of the positive feedback loop was examined in different approximations. As expected, in the exact calculation the ignition was found to be delayed compared to a calculation where the binding of signal molecules was quasistatic. The buffering of signal molecules is found to have little effect on the ignition process. Contrary to expectation, we find that the ignition does not start when the threshold is reached at the center—instead it allows for the threshold to be approached in the whole colony followed by an almost simultaneous ignition of the whole biofilm aggregate.",
"conclusion": "Conclusion We have modelled a generic single-loop quorum sensing system with positive feedback. The primary goal of the study has been to study of the space-time structure of the ignition of the switch produced by the positive feedback in the quorum sensor. The exact solution exhibits a delayed response compared to both the quasi-static and static approximations, as it is limited by the time it takes the regulator stages to react and build up concentration of activated regulator. Inclusion of buffering terms produces no further retardation of the system. Using a 3D representation to depict both the spatial and temporal dependency of the activated regulator concentration, the collective behaviour of the colony could be studied. A slow build-up over five minutes to the ignition concentration of activated regulator followed by a quick ignition was observed. The process exhibits the desired behaviour, as the entire colony is either in an on- or off-state. These observations indicate that a partial ignition is difficult to achieve, even for slow systems. The model demonstrates that even the largest naturally occurring biofilm aggregates in chronic infections [ 41 ] ignite fully in less than a minute and, truly, can be said to produce a surprise attack.",
"introduction": "Introduction Quorum sensing (QS) is a biological regulation process utilised by bacteria to control behaviour in accordance with size, density, and growth-rate of a bacterial population [ 1 ]. The process is based on diffusible signal molecules, produced by the bacteria at a background level. The signal molecules are able to bind to regulator molecules within the bacteria, thereby activating the regulator [ 2 ]. The collective behaviour regulated by QS was reported in Vibrio fischeri where it regulates camouflage light in large cell colonies in the host [ 3 – 5 ]. Since then, QS systems have been reported in many bacteria, e. g. Aeromonas hydrophila [ 6 – 9 ], Agrobacterium tumefaciens [ 10 ], and Pseudomonas aeruginosa [ 11 – 13 ]. The presence of QS in colonies of bacteria appears to be the rule rather than the exception. Frequently, the array of gene expressions acting under the control of activated QS regulators includes signal molecule synthetase. This positive feedback leads to a size-sensitive switch which can be used to control collective behaviour [ 1 – 5 , 14 ]. The switch makes it possible to maintain an invisible state until the sudden QS regulated attack sets in. Recently, a proper measure of the “size” of a spherical biofilm aggregate was established as the cell density multiplied by the squared radius of the colony [ 15 , 16 ]. The establishment of the size measure was based on the observation that the concentration of the activated regulator\n r a = [ R 2 S 2 ] (1) \nmay be interpreted as the intrinsic measure of how quorate the state of the colony is and controls the quorum sensing feedback as well as QS regulated genes [ 15 ]. In small colonies the signal molecules are produced at a low background level. In larger colonies, the diffusive signal molecules accumulate, activate the transcriptional regulator, and induce transcription of the signal molecule synthetase at an increased level. The dimer form of the activated regulator implied in Eq (1) allows for a fully developed switch in the ignition of the quorum as observed in Gram negative bacteria [ 3 – 5 , 14 ]. The dimeric form is typical to quorum sensors and has been confirmed in main QS loops in a range of Gram negative bacteria [ 1 , 12 , 17 – 25 ]. In the present study we will study the time course of the ignition of the size-dependent switch in the most basic form of a single quorum sensitive switch."
} | 1,298 |
37287087 | PMC10246149 | pmc | 90 | {
"abstract": "Background Reef-building corals, the foundation of tropical coral reefs, are vulnerable to climate change e.g. ocean acidification and elevated seawater temperature. Coral microbiome plays a key role in host acclimatization and maintenance of the coral holobiont’s homeostasis under different environmental conditions, however, the response patterns of coral prokaryotic symbionts to ocean acidification and/or warming are rarely known at the metatranscriptional level, particularly the knowledge of interactive and persistent effects is limited. Using branching Acropora valida and massive Galaxea fascicularis as models in a lab system simulating extreme ocean acidification (pH 7.7) and/or warming (32 °C) in the future, we investigated the changes of in situ active prokaryotic symbionts community and gene expression of corals under/after (6/9 d) acidification (A), warming (H) and acidification–warming (AH) by metatranscriptome analysis with pH8.1, 26 °C as the control. Results A, H and AH increased the relative abundance of in situ active pathogenic bacteria. Differentially expressed genes (DEGs) involved in virulence, stress resistance, and heat shock proteins were up-regulated. Many DEGs involved in photosynthesis, carbon dioxide fixation, amino acids, cofactors and vitamins, auxin synthesis were down-regulated. A broad array of new DEGs involved in carbohydrate metabolism and energy production emerged after the stress treatment. Different response patterns of prokaryotic symbionts of massive G. fascicularis and branching A. valida were suggested, as well as the interactive effects of combined AH and persistent effects. Conclusions The metatranscriptome-based study indicates that acidification and/or warming might change coral’s in situ active prokaryotic microbial diversity and functional gene expression towards more pathogenic and destabilized coral-microbes symbioses, particularly combined acidification and warming show interactive effects. These findings will aid in comprehension of the coral holobiont’s ability for acclimatization under future climate change. Supplementary Information The online version contains supplementary material available at 10.1186/s40793-023-00505-w.",
"conclusion": "Conclusion Based on the RNA-based diversity and gene expression comparison in this study, A, H and AH stressors change coral’s in situ active microbial community structure and functional gene expression profiles, whatever during and after these stresses, which probably cause shifts in coral prokaryotic symbionts from healthy mutualistic relationships to those that are more pathogenic and detrimental to the coral host. DEGs comparison indicates that these stressors could increase the abundance of microbial genes involved in virulence, stress resistance, and heat shock proteins. Many DEGs involved in photosynthesis, carbon dioxide fixation, amino acids, cofactors and vitamins, auxin synthesis are down-regulated, indicating the destabilized or destroyed coral-microbes symbioses. In particular, compared with single stress, synergetic effects of combined AH on microbial diversity and gene expression are indicated. Massive G. fascicularis with higher microbial diversity shows more tolerance than branching A. valida , indicating coral species-dependent microbial response patterns. Further physiological and biochemical studies are required to ascertain the consequences of these functional microbial diversity and gene expression shifts caused by acidification and /or warming since this study was only carried out using metatranscriptome strategy. The different response patterns of sensitive and tolerant corals indicated in this study highlight the importance of coral prokaryotic symbionts in assessing the impacts from thermal and/or acidification, which is helpful for us to predict the consequences of global climate change for coral reef ecosystems.",
"discussion": "Discussion A, H and AH stressors change coral’s in situ active microbial community towards potentially pathogenic bacteria Taxonomic shifts of microbes of corals under environmental stress have been observed [ 17 – 20 , 24 , 38 – 40 ]. Corals in lowered pH had higher microbial diversity compared with the control [ 41 ]. Higher bacterial diversity was represented in bleached sample compared to health sample during a bleaching event [ 19 ]. Such changes could result in the disequilibrium of the coral microbial community structure. Compared with the study using 16 S rDNA, 16 S rRNA could reflect exactly the ‘active’ fraction of microbial community. In this study, using a RNA-based sequencing approach we were able to detect significant differences in the in situ prokaryotic symbionts under different stresses. Vezzulli et al. [ 42 ] reported that destabilization of the coral holobionts was concomitant with a microbial community shift towards opportunistic microorganisms or potential pathogens, such as Vibrio spp. due to thermal stress. In this study, based on RNA-based diversity analysis of coral prokaryotic symbionts, coral G. fascicularis hosted a significantly higher relative abundance of Desufobacterales in the thermal- or acidified groups. Besides, some potentially coral pathogens, e.g. Clostridiales, Vibrionaceae , Flavobateriaceae , Rhodobacteraceae , and Desulfobacteraceae were found in the corals under stress. Particularly significant increase of Rhodobacteraceae and Vibrionaceae in A. valida under AH stress was detected. Besides the diversity change, another primary driver of microbial variation was the relative abundance change, for example reduced relative abundance of Caldithrixales, Pseudanabaenaceae ( Cyanobacteria; Synechococcophycideae ) in G. fascicularis , putatively endosymbiotic Endozoicimonaceae in A. valida , and increased abundance of Oscillatoriales ( Cyanobacteria ; Oscillatoriophycideae ) for both coral species were detected in this study. Compared with the control, coral A. valida showed significant higher relative abundance of Cyanobacteria in all H and AH groups, the relative abundance of Chroococcales and Oscillatoriales increased in both bleached corals, indicating Cyanobacteria’s replacement potential for Symbiodiniaceae under stress. This is probably a strategy for coral holobionts try to maintain the carbon utilization from destroyed photosynthesis of Symbiodiniaceae. \nBased on our results, coral species-dependent microbial community change with thermal/acidification stress was suggested. Though A, H and AH stressors could change both corals’ in situ active microbial diversity, H showed most negative effects on G. fascicularis microbial composition, while H, AH on A. valida microbial composition (Table 1 a, Additional file 1 : Table S2). The DEGs number in A. valida . was bigger than G. fascicularis and most of the DEGs were down-regulated in A. valida . Totally A. valida prokaryotic symbionts was much more sensitive than G. fascicularis prokaryotic symbionts, indicating coral-species dependent response and recover patterns (Table 1 ; Additional file 1 : Table S1; Figs. 2 and 3 ). Based on the weighted unifrac distance analysis, although the microbial community structure of G. fascicularis was not completely recovered when the stress was removed, its recoverability was better than that of A. valida (avg. weighted unifrac distance: 0.360–0.651 vs. 0.607–0.876). Accordingly, we hypothesized that coral species with higher microbial diversity may have a greater probability of the presence of stress-resistant taxa or more complex interaction-network and develop functional redundancy that could maintain stability of community structure and even confer stress tolerance to their holobionts. This hypothesis is in consist with the research of Yu et al. [ 43 ], they also suggested that the higher tolerance of Pavona decussata compared with that of Acropora pruinosa might result from a complex biological process caused by higher symbiotic bacterial diversity, different dominant bacteria, higher host immune and stress resistance responses, and lower metabolic rate. Recently study showed that bacterial diversities associated with massive coral were generally higher than those associated with branching corals [ 44 ]. The thickness of tissue layer influences the ability of microbes to colonize [ 45 ], G. fascicularis with a thicker tissue may provide a higher complex niche for microbial colonization which is supported by the fact that massive G. fascicularis hosted higher bacterial diversity than A. valida in all treatment and control groups. Interestingly, OTUs affiliated with Caldithrix were only detectable in the prokaryotic symbionts of G. fascicularis . This bacterial phylum has been recently recognized as a new independent phylum-level clade, and were identified from the sponges [ 46 ], ascidians [ 47 ] and marine hydrothermal vents [ 48 ]. Genomic analysis of the representative lineage Caldithrix abyssi , demonstrated the ability to synthesize nucleotides, most amino acids and vitamins as well as contain genes encoding proteins that confer O 2 tolerance, suggested that such a flexible metabolism has the potential to help C. abyssi to adapt to changing conditions [ 48 ]. The results from Grottoli et al. [ 49 ] also indicted that temperature-stress tolerant corals have a more stable microbiome and propose that coral with a stable microbiome were also more physiologically resilient and thus more likely to persist in the future. Ocean acidification as a result of increased anthropogenic CO 2 can cause a shift in coral-associated microbial communities of p CO 2 -sensitive corals. But in the case of p CO 2 -tolerant coral, massive Porites spp. showed a high degree of tolerance to OA [ 50 ]. A, H and AH stressors change the coral microbial metabolism and destabilize the coral-microbes symbioses Thurber et al. [ 51 ] evaluated the structural and functional changes in coral microbial communities of Porites compressa under increased temperature, reduced pH stressors. They found that stressors increased the abundance of microbial genes involved in virulence, stress resistance, sulfur and nitrogen metabolism, motility and chemotaxis, fatty acid and lipid utilization, and secondary metabolism. Ziegler et al. [ 22 ] found that functional profiles of microbial communities changed with thermal stress, several functions related to carbohydrate metabolism were enriched. The study of Rädecker et al. [ 52 ] also indicated that heat stress destabilizes symbiotic nutrient cycling in corals, which will reduce the coral holobiont health. Ocean acidification could result in the increase of virulence-associated gene expression and shifts in the community composition, e.g. increases in opportunistic pathogens such as Vibrionaceae and Alteromonadaceae and a loss of putatively symbiotic Endozoicomonas spp. [ 43 ]. Thus, the diversity and relative abundance changes of coral microbiota caused by stress could alter microbial metabolism, profoundly shift the health status of the coral holobionts. In this study, H and AH stressors could change both corals microbial gene expression obviously (Table 1 b; Figs. 3 , 4 and 5 ). For both corals tested, the DEGs mainly include Carbohydrates, Cofactors, Vitamins, Prosthetic Groups, Pigments, Protein Metabolism, Amino Acids and Derivatives, and Stress Response, indicating the broad influences of A, H and AH stressors on coral microbial metabolisms. Particularly, functional gene analysis demonstrated that stressors increased the abundance of microbial genes involved in virulence, stress resistance, sulfur and nitrogen metabolism, motility and chemotaxis, fatty acid and lipid utilization, and secondary metabolism. For example, Vibrionaceae, Flavobateriaceae, Rhodobacteraceae , and Desulfobacteraceae showed up-regulated expression patterns (Fig. 3 ). In addition, persistent effect on coral microbial metabolisms was also suggested even after the stress was removed for 9 d, since some DEGs exhibited in exposure groups were difficult to recover to that of the control level, especially after H stress for G. fascicularis and AH, H for A. valida . Meanwhile, a broad array of new DEGs emerged after the stress, suggesting a continued effect on the coral microbial metabolisms. A, H and AH show great impacts on coral-microbes symbioses since the expression of nutrient metabolism related genes’ expression was mostly down-regulated (Fig. 5 ), e.g. many DEGs involved in photosynthesis, carbon dioxide fixation, amino acids, cofactors and vitamins, auxin synthesis were detected), indicating the blocked microbial metabolism especially in branching A. valida (Fig. 5 ). Interestingly, plant hormone gene expression was also down-regulated under these stresses, suggesting the possible corrupted relationship of microbes with Symbiodiniaceae (Figs. 5 and 6 ). It is known that bacterial auxin can cause interference with plant developmental processes [ 53 ]. Indole-3-acetic acid (IAA), the major naturally occurring auxin, was detected in marine sediments, and was proposed producing by various marine bacteria, may affect algal growth in natural environments [ 54 ]. IAA synthesis related gene expression changes may affect the growth and reproduction of symbiotic algae, thus, affect the entire symbiotic system in corals, e.g. the interaction between bacteria and algae in corals could be reduced. In this study, microbial endemic gene IPA (iso-propyl alcohol) decarboxylase gene (the whole reaction rate-limiting step) was down-regulated in each treatment group. The bacterial auxin synthesis and adhesion genes were significantly down-regulated after stress removal. Synergetic interaction of combined A and H on the coral prokaryotic symbionts Synergetic effects could either offset each other (i.e. antagonistic effect, where two stressors interact to produce an effect that one stressor reduces/mitigates the level of another) or aggravate it through an accumulation of stress (acceleration effect). Prada et al. [ 25 ] showed a synergistic adverse effect on the mortality rates of corals ( Balanophyllia europaea , Leptopsammia pruvoti , and Astroides calycularis )(up to 60%), suggesting that high seawater temperatures may have increased their metabolic rates which, in conjunction with decreasing pH, could have led to rapid deterioration of cellular processes and performance. Pitts et al. [ 24 ] also suggested that ocean acidification partially mitigates the negative effects of warming on the larval development of Orbicella faveolata . Agostini et al. [ 26 ] investigated the effects of elevated temperature and high pCO 2 on G. fascicularis with oxygen and pH microsensors and found that, under a combination of high temperature and high CO 2 , the photosynthetic rate increased to values close to those of the controls indicating an antagonistic effect on the photosynthesis of G. fascicularis holobiont. In the case of coral microbiome’s response to warming and acidification, Webster et al. [ 39 ] explored the microbiome response of corals e.g. Acropora millepora , Seriatopora hystrix to near-future climate change conditions. An interactive acceleration effect between stressors was also identified, with distinct communities developing under different pCO 2 conditions only evident at 31 °C. In this study, the interactive effect from A and H on coral prokaryotic symbionts was indicated by the comparison between combined AH and A or H alone. The DEGs patterns shifted the most obvious under AH stress based on the distribution of DEG fold change (Fig. S3). In the case of branching coral A. valida , AH showed acceleration effect on both the in situ active microbial community and gene expression since the microbial diversity and DEGs number were the highest under AH stress compared with A or H (Table 1 ; Figs. 4 and 5 ), and this acceleration effect from AH still existed when AH was removed. The acceleration effect of acidification and heat stress on A. valida prokaryotic symbionts in this study is consist with the suggestion from webster et al. [ 39 ]. For massive coral G. fascicularis , AH showed acceleration effect on microbial gene expression compared with single A or H stress because AH caused the most DEGs (Fig. 4 ), which was also supported by PERMANOVA analysis in Table 1 ( pseudo -F 5.046 in AH treatment bigger than pseudo -F 4.806 in A treatment) (Fig. 4 ; Additional file 1 : Fig. S3). But, the acceleration effect on microbial gene expression disappeared when AH was removed. Interestingly, AH showed antagonistic effect on microbial community structure in stress treatment (T) and removal groups (P) because A was able to mitigate the effect of H on the microbial community structure change. Different from A. valida , G. fascicularis recover better when AH stress was removed and the acceleration effect was not on-going. The different interactive effect of AH on coral G. fascicularis from A. valida indicates the coral species-dependent response to stress, which may result from their different microbial communities."
} | 4,293 |
39967391 | PMC11938035 | pmc | 91 | {
"abstract": "Abstract Inspired by nature's ability to master materials for performance and sustainability, biomimicry has enabled the creation of bioinspired materials for structural color, superadhesion, hydrophobicity and hydrophilicity, among many others. This review summarizes the emerging trends in novel sustainable fluorocarbon‐free bioinspired designs for creating superhydrophobic and superoleophobic surfaces. It discusses methods, challenges, and future directions, alongside the impact of computational modeling and artificial intelligence in accelerating the experimental development of more sustainable surface materials. While significant progress is made in superhydrophobic materials, sustainable superoleophobic surfaces remain a challenge. However, bioinspiration and experimental techniques supported by computational platforms are paving the way to new renewable and biodegradable repellent surfaces that meet environmental standards without sacrificing performance. Nevertheless, despite environmental concerns, and policies, several bioinspired designs still continue to apply fluorination and other environmentally harmful techniques to achieve the required standard of repellency. As discussed in this critical review, a new paradigm that integrates advanced materials characterization, nanotechnology, additive manufacturing, computational modeling, and artificial intelligence is coming, to generate bioinspired materials with tailored superhydrophobicity and superoleophobicity while adhering to environmental standards.",
"conclusion": "5 Conclusion The development of superhydrophobic, underwater superoleophobic, and superamphiphobic surfaces has led to environmental concerns due to the accumulation of fluorinated compounds. However, research is underway to create sustainable alternatives inspired by Nature. While successful methods such as using nanotechnology, 3D printing, and templating exist for superhydrophobicity and underwater superoleophobicity, limited success has been achieved in creating sustainable in‐air superoleophobic surfaces. To overcome the challenges associated with in‐air superoleophobicity, biomimicry holds great promise. Researchers can learn from nature‐based strategies employed by organisms such as Leafhoppers and Collembola to develop more effective solutions. In addition to experimental developments, computational studies and tools play a crucial role in understanding the structure–property relationships of these materials. Computational methods like DFT, MD, and FEM can provide valuable insights into atomistic‐level phenomena. The integration of these computational methods with AI techniques, such as ML, offers exciting opportunities for optimizing and predicting material structures. This review has showcased the recent advancements of superhydrophobic and superoleophobic surfaces, aided by bioinspiration. A summary of the bioinspired materials is shown in Table \n 2 \n . Bioinspiration, particularly from the Springtail, is providing a route to durable superoleophobicity without the use of fluorocarbon functionalities. Table 2 A comparative summary of the key properties of the bioinspired materials. Inspiration WCA [°] WSA [°] OCA [°] OSA [°] Fluorocarbon functionality Durability Refs. Rove beetles and water striders 160 9.8 N/A N/A No \n Thermostable \n pH stable \n UV stable \n Abrasion resistant \n [ 110 ] Conch shells 154.4 3.8 N/A N/A No \n Abrasion resistant \n Ductile \n [ 112 ] Diving bell spiders 146 N/A N/A N/A No N/A [ 113 ] Honeycomb >150 N/A Oleophilic N/A No \n Thermostable \n Ductile \n Flame‐retardant \n Stable in water and seawater \n pH stability \n [ 115 ] Marine mussels 151.9 11.3 N/A N/A No \n Washable \n Vulnerable to abrasion \n [ 116 ] Lotus leaf 164 2.6 Oleophilic N/A No N/A [ 111 ] Lotus leaf ≈160 <5 N/A N/A No N/A [ 117 ] Lotus leaf 153.9 8 N/A N/A No \n pH stable \n UV stable \n O 2 plasma stable \n Flame‐retardant \n [ 118 ] Goose feathers 153 N/A N/A N/A No N/A [ 119 ] Leather 161.5 N/A 152.2 N/A Yes \n Thermostable \n UV stable \n [ 126 ] Fish skin <5 N/A 160–170 (underwater) N/A No Oil/water separation efficiency unaffected by reuse; slight decrease in flux [ 127 ] Nacre Hydrophilic N/A 160 (underwater) N/A No \n pH stable \n Solvent resistance \n Pressure stable \n High Young's modulus and hardness \n [ 128 ] Manis pentadactyla N/A N/A ≈115 N/A No Impact resistant [ 130 ] Springtail 156 N/A 130 N/A Yes \n Ductile \n Pressure stable \n Elastic \n [ 122 ] Springtail >150 ≈15 >150 ≈10 Yes \n Abrasion resistant \n Lifespan >6 months \n [ 123 ] Springtail 154.7 0.2 >150 <4 Yes \n Pressure stable \n Elastic \n [ 124 ] Springtail >150 ≈2 >140 <10 Yes \n Thermostable \n pH stable \n Ultrasonic cleaning resistant \n Scratch resistant \n Lifespan >120 days \n [ 125 ] Springtail 161.5 N/A >150 N/A No \n O 2 plasma stable \n Ductile \n Elastic \n Impact resistant \n [ 133 ] Springtail 158.6 N/A >150 N/A No \n O 2 plasma stable \n Ductile \n Elastic \n Impact resistant \n [ 134 ] John Wiley & Sons, Ltd. Furthermore, computational tools can successfully predict the wetting properties of surfaces and provide atomistic‐level information with respect to time, which is not possible with experimental methods. These tools can be used to predict the contact angle of the surfaces for different liquids in the experimental range and to observe the liquid behavior on the surfaces, such as diffusion or adsorption. They can also virtually observe changes in inter‐ and intramolecular interactions, such as hydrogen bonds, throughout the simulation, which affect the observable experimental properties of the materials, such as form, stiffness, and water‐related properties. Computational tools allow us to follow specific atoms’ behavior over time, and this provides information regarding functional groups interacting with water, while identifying the forces governing this interaction. Understanding the governing interactions gives researchers the opportunity to control the synthesis processes and to make more accurate designs for their experiments. Additionally, AI and ML methods are accelerating material design. 5.1 Future Directions and Outlook Future directions should be based on successful methods thus far, whilst investigating alternatives to the energy extensive processes. Up to now the only avenue to produce the precise, nanoscale re‐entrant structures required for effective superoleophobic surfaces without incorporating fluorocarbons involves the use of photolithography. In the cases seen, this is used with a templating method which somewhat reduces the energy intensity of the process, as the molds created through photolithography can be reused many times. However, even the templating technique is costly, energy intensive, and usually uses harmful chemicals for etching. More systematic studies of the direct effect of dimension sizes on oleophobicity to ascertain the essential requirements are required. When the specifications are known, other potential methods of production can be identified and investigated. 3D printing is promising especially as technological advancements allow for easier creation of smaller feature sizes. Nevertheless, it is still expensive and time‐consuming to produce large volumes, particularly when a combination of techniques is required. This is the case for potential superoleophobic materials. Macroscale printing is used for the main substrate of the material, while micro‐ and nanoscale printing are used to produce hierarchical, re‐entrant structures. 3D printing could be implemented to print molds, as discussed with photolithography. The molds could then be used repeatedly to produce the final structured materials, if demolding is achievable, thus reducing cost and time per output made, and making the method more feasible for industrial scale. Nanosecond lasering is another prospective method as it can be used to produce nanoscale features on a range of substrates including metals, plastics, glass, and ceramics. Subsequent steps still need to be evaluated to validated nanosecond lasering on biobased polymer materials. The features produced with this technique are less precisely tuneable, but consistent re‐entrant structures across a surface have been achieved, as discussed. Compared to femtosecond lasers, nanosecond lasers are more cost‐effective with simpler technology, the processing time is faster and so they are more suitable for large scale production. Femtosecond lasering is also a potential option, with higher precision for nanostructures but is more costly and time consuming per unit of material produced. Again, evaluations into the compulsory feature dimensions would provide further insights, if the intensity of femtosecond lasers is required, or if nanosecond lasers would be sufficient. Shrinking techniques such as thermoresponsive hydrogels or pyrolysis methods are being used to reduce the dimensions of features. These reduce the strenuousness of the initial fabrication method, as the size restrictions are not so stringent. As seen in this review, shrinking was only applied to structures made through photolithography. Therefore, it would be formative to research this further with features made by other processes and determine if the shrinking effect can be applied to a more cost and energy efficient initial method. For example, with lasering techniques, using a shrinking technique could allow for femtosecond lasering‐like structures to be produced through nanosecond lasering and shrinking, reducing the energy intensity of the process whilst maintaining the final result. Electrospinning increases the roughness of a material, which enhances its inherent wettability properties. This is pertinent in the case of a hydrophobic polymer, which can often become superhydrophobic after electrospinning, but this method has not been able to induce superoleophobicity. This is likely because the material needs to already have oleophobicity before electrospinning, meaning the initial property comes from chemistry rather than structure, and so it is difficult to achieve superoleophobicity without the use of fluorocarbons. Furthermore, post processing is required to produce a usable film or coating from the electrospun polymer, although this can reduce the roughness and so the liquid repellency of the material. Despite this, if a superrepellent film or coating is successfully fabricated, it could then be applied to an independent substrate. Then, a sustainable material could be chosen such as cardboard, and the repellent properties of the film or coating would be donated to the cardboard, although this again mandates another fabrication step of applying the coating to its substrate, which would also require research into the most effective, efficient, and scalable application method. SAMs allow exact control of the surface chemistry, with functional groups added relatively easily depending on chemical compatibility. The thickness of the layers is also directly controllable, and they form spontaneously so the process is energy efficient and feasibly scalable. As the layers are formed through chemical bonding, they are usually stable and durable. The method is already used industrially in microelectronics, biosensors, and more, and has proven sufficient for sustainable superhydrophobic materials. However, the repellency properties are mostly a consequence of surface chemistry, so the resultant materials may never be able to achieve superoleophobicity without the use of fluorocarbons as the topography is not explicitly controllable. To further assist the development of the materials, computational tools need to be furthered integrated into the design process. Nevertheless, to fully harness the potential of computational tools, future studies should focus on combinations of simulation methods (DFT, MD, etc.) with ML algorithms. The role of experimental findings on the application of ML is also significant, because the training of ML models depends on the accuracy of the experimental data. Therefore, future studies need more collaboration between researchers with expertise in computational and experimental methods to gain better insight into surface design parameters. Another point discussed for using computational tools in surface design is the test methods used for the analysis of the simulations. Although computer‐based methods give more control over test environments and conditions, differences in simulation parameters and methods for the analysis of simulation data, as seen in contact angle and diffusion analysis, make more challenging the comparability of computational results across studies. While fully standardizing the set‐up of the simulations is extremely challenging given the differences in systems, test conditions and computational method, some degree of standardization is needed to enable more accurate comparisons. In summary, although the current simulation methods are quite helpful in understanding the atomistic level behavior of the surfaces and materials, they have still limitations in the length and time scale of the models and simulations. With the development of more advance high throughput, GPU‐enabled, and quantum computing technologies, and with the implementation of more accurate ML models for material discovery and property prediction, it will be possible to achieve larger models and longer simulations. These improvements will allow faster scanning of design options, which will contribute to a better understanding of governing parameters in surface design, reduce the time of computational‐to‐experimental data validation.",
"introduction": "1 Introduction Nature has more than 3.8 billion years of experience in material design at the pinnacle of performance and sustainable technologies. Only the most effective designs and efficient procedures have managed to propagate and stand the test of time. Since antiquity, humanity has taken inspiration from biology to advance contemporary technologies. Tools and weapons, clothing, and endless examples of aircrafts, helicopters, drones, and even wind turbines or urban planning have been developed with biomimicry. [ \n \n 1 \n , \n 2 \n , \n 3 \n , \n 4 \n , \n 5 \n \n ] \n Nanostructured surfaces have been very specific sources of inspiration. Geckos possess spatula surface nanostructure on their feet, which maximizes surface contact and van der Waals forces to adhere to walls. This has resulted in the development of Gecko‐inspired coatings and adhesives used in a range of fields including medical and electronic. [ \n \n 5 \n \n ] Velcro was invented after a Swiss engineer noticed burrs sticking to his dog's fur due to their microscale hooks clinging to the substrate. [ \n \n 6 \n \n ] Even fireflies known for their ability to emit light benefit from high‐transmission surface nanostructures that enhance light output. This resulted in firefly‐inspired LEDs with improved illumination and reduced energy consumption, as well as a longer lifespan than traditional bulbs. [ \n \n 7 \n \n ] Natural photonic crystals and structural color have provided inspiration for a huge range of applications such as displays, sensors, optics, medical fields, solar cells, camouflage, and paints. [ \n \n 8 \n , \n 9 \n \n ] The Kingfisher's splash‐minimizing dive is a result of the aerodynamics of the beak design, which was mimicked by engineers to develop the bullet train, reducing the noise pollution and pressure waves, as well as improving the speed and energy efficiency. [ \n \n 10 \n \n ] Undoubtedly, nature has a myriad of innovations, developed to promote survival and success in withstanding the dynamics of the perpetually changing environment, which hold elite solutions to aid in the design of advanced materials. Fascinated by these Natural designs, modern technology, with both computational and experimental techniques, has allowed humanity to mimic biota. There have been effective recent reviews regarding experimental techniques to produce superhydrophobic materials, superoleophobic materials, and bioinspired superhydrophobic materials. In this review we combine superhydrophobicity, superoleophobicity, bioinspiration, and sustainability, to provide an up to date and critical examination of the research field and current techniques encompassing all aspects from performance to environmental effects using both experimental and computational approaches. Nowadays, there are tools to observe, characterize, and recreate some natural designs from the nanoscale to the macroscale. Some examples include electrospinning, 3D printing, molding, hot embossing, lasering, lithography, oxygen plasma treatments, pyrolysis, sol–gel process, magnetron sputtering deposition, and even the use of nanomaterials or photonic crystals. Electrospinning increases the surface roughness to enhance the wetting properties of the material. [ \n \n 11 \n , \n 12 \n , \n 13 \n \n ] 3D printing creates the necessary topography required for superhydrophobic/oleophobic materials, although it is limited to the microscale, preventing the creation of nanostructured designs as intricate as those in nature. [ \n \n 14 \n , \n 15 \n , \n 16 \n , \n 17 \n \n ] Molding and templating have been used to create biomimetic structures with re‐entrant topography capable of repelling low‐surface‐tension liquids. [ \n \n 18 \n \n ] Similarly, hot embossing is able to create micro‐ and nanostructures on polymeric surfaces with superhydrophobic properties. [ \n \n 19 \n \n ] Lasering creates surface textures on metals, nonmetals, and composite materials. [ \n \n 20 \n , \n 21 \n \n ] Lithography and etching provide highly geometrically controllable surfaces of organized arrays through various size scales, tuneable for desired properties. [ \n \n 22 \n \n ] Oxygen plasma treatment oxidizes the surface to create hydrophilic polar groups. [ \n \n 23 \n , \n 24 \n \n ] Pyrolysis increases liquid repellency of a material, while sol–gel processes create xerogels and aerogels with hydrophobic/oleophobic properties. [ \n \n 25 \n , \n 26 \n , \n 27 \n , \n 28 \n , \n 29 \n , \n 30 \n \n ] Magnetron sputtering deposition can also provide hydrophobic and oleophobic thin films, although they are often lacking stability and durability. [ \n \n 31 \n \n ] \n Nanomaterials are also utilized to increase the surface roughness and to enhance the wettability and surface area to volume ratio of material surfaces, with different properties from the corresponding bulk material. [ \n \n 32 \n , \n 33 \n , \n 34 \n , \n 35 \n \n ] Thus, the use of superhydrophobic hybrid nanocomposites is an emerging field with potential to overcome stability and durability issues by directing the nanostructures downward creating nanopores or similar. [ \n \n 36 \n \n ] Also, photonic crystals have complex surface geometries and chemical compositions which can be tuned toward the desired wettability properties. [ \n \n 37 \n \n ] Layer‐by‐layer (self‐)assembled materials and self‐assembled monolayers (SAMs) can also display hydro/oleophobicity. [ \n \n 38 \n , \n 39 \n , \n 40 \n \n ] \n Additionally, developments in computational hardware and software alongside advancements in machine learning and artificial intelligence is accelerating the design of superhydrophobic and superoleophobic materials. Thus, a new paradigm that integrates advanced materials characterization, nanotechnology, additive manufacturing and computational modeling and artificial intelligence is coming to generate bioinspired materials for optimal performance, while adhering to environmental standards. 1.1 Hydrophobicity in Nature Many intricate designs in the biosphere have been perfectly fine‐tuned for superhydrophobicity. For example, the fogstand beetles ( Onymacris unguicularis ) and other Namib beetles ( O. laeviceps , Stenocara gracilipes , Physasterna cribripes ), which reside in the Namib Desert have very little access to water. [ \n \n 41 \n , \n 42 \n , \n 43 \n , \n 44 \n \n ] To combat this, they have developed a unique ability to obtain water from humid air due to the chemistry and structure of their backs ( Figure \n \n 1 a ). [ \n \n 42 \n , \n 45 \n \n ] They have an abundance of microscale bumps/ridges on the surface of their backs, which are uncoated and hydrophilic. [ \n \n 45 \n , \n 46 \n \n ] The areas which are not bumps are wax coated and therefore hydrophobic. This means water from fog can collect on the hydrophilic peaks, and drip down to the hydrophobic zone, where in combination with the head down stance of the beetle, the water can roll freely over the hydrophobic surface toward the beetle's mouth where it can drink the collected water. This has been a notable source of interest for synthetic fog harvesting opportunities. [ \n \n 47 \n , \n 48 \n \n ] Another example of hydrophobicity in nature is butterfly wings, showing self‐cleaning and anti‐icing properties (Figure 1b ). [ \n \n 49 \n \n ] Many butterfly species have chitin‐based wings composed of periodic microsized rectangular scales adorned with a secondary structure of nanoscale longitudinal ridges and lateral bridges as well as nanostrips atop the secondary structure, resulting in a three‐phase hierarchical micro/nanostructure. [ \n \n 50 \n , \n 51 \n \n ] As a result of this surface structure, butterflies have been identified to have high water contact angles (WCAs), sometimes even in the superhydrophobic region. [ \n \n 50 \n , \n 52 \n \n ] An abundance of plant species such as rose petals and rice leaves exhibit (super)hydrophobic surfaces due to either surface structures, cuticle waxes and coatings, or a combination of both. [ \n \n 53 \n , \n 54 \n \n ] Water‐repellent leaves possess a wax layer containing crystals from 0.5 to 20 µm in length and varying from large observable wax crystals clusters to small consistent crystals appearing as a smooth uniform layer. [ \n \n 55 \n \n ] Furthermore, leaves with wax covered trichromes are seen to be exceptionally hydrophobic. [ \n \n 55 \n , \n 56 \n \n ] Depending on the species of the plant, leaves can have considerably varying surface structures as well as wax types. Overall, the most consistently hydrophobic leaf types are those with low carbonyl species present, ordered platelet light structures with a high degree of roughness and high polar and Lewis base free energies. [ \n \n 57 \n \n ] The most prominent example is the lotus leaf ( Nelumbo nucifera ). Lotus leaves have WCAs over 150° and water sliding angles (WSAs) <5° endowing them with supreme superhydrophobic and self‐cleaning properties, leading to widespread recognition of the term the lotus effect to describe self‐cleaning surfaces (Figure 1c ). [ \n \n 55 \n , \n 58 \n , \n 59 \n \n ] The leaves have micro/nanoscale hair‐like structures along with a layer of wax crystals which significantly reduces the contact area between the water droplets and the surface of the leaves. [ \n \n 60 \n , \n 61 \n \n ] The upper epidermis of the lotus leaf provides the leaf with its superiority over other plant leaves with regard to stability, durability, and consistency as a result of its unique hierarchical structure paired with agglomerated wax tubules. [ \n \n 60 \n , \n 62 \n \n ] Water striders ( Gerris remigis ) are another example of natural water‐repellency due to hydrophobic legs as well as surface tension mechanisms, elastic momentum transfer, and wax secretion. [ \n \n 63 \n , \n 64 \n , \n 65 \n , \n 66 \n \n ] The WCA of the water strider's leg has been measured as >167°, and the WCA of the wax secreted by the water strider is only 105°, supporting that the surface structure must play a pivotal role in the superhydrophobicity. [ \n \n 66 \n \n ] The water strider's legs consist of needle‐like hairs of hundreds of nm to 3 µm in diameter and ≈5 µm in length, protruding at an angle of 20°, and enriched with nanoscale valleys. This creates the hierarchical micro/nanosurface structure, as shown in Figure 1d , that works synergistically with the wax secretion to create superhydrophobicity. Figure 1 Hydrophobicity in nature. a) A Namib Beetle ( Stenocara gracilipes ) and an SEM image of its surface features. Reproduced with permission. [ \n \n 44 \n \n ] Copyright 2010, Nørgaard and Dacke. b) A butterfly and an SEM image of the primary rectangular scale structure of its wing. Reproduced with permission. [ \n \n 51 \n \n ] Copyright 2014, Science China Press. c) A lotus leaf and an SEM image of its microstructure. Reproduced with permission. [ \n \n 60 \n \n ] Copyright 2011, Beilstein‐Institut. d) A water strider and an optical microscope image of its leg, showing the bristles, setae, and microtrichia microstructures. Reproduced with permission. [ \n \n 66 \n \n ] Copyright 2004, Springer Nature Limited. 1.2 Underwater Oleophobictiy in Nature The most abundant examples of oleophobicity in nature occur underwater, as the affinity for water facilitates the repelling of oils and low surface tension liquids. Hence, the lower part of the lotus leaf shows underwater superoleophobicity. [ \n \n 67 \n \n ] The surface of seaweed displays underwater superoleophobicity that can be sustained even in high salinity and high ionic strength water environments. [ \n \n 68 \n , \n 69 \n \n ] The surface consists of pores and ridges providing a hierarchical structure ( Figure \n \n 2 a ), which works in combination with the water bonding ability of the polysaccharide surface chemistry to create superoleophobicity. [ \n \n 69 \n , \n 70 \n \n ] Fish skin and scales are often underwater superoleophobic, giving them the ability to swim through oil‐polluted waters and remain contaminant‐free. The Filefish possess oriented microscale spines resembling hooks protruding from their skin (Figure 2b ) resulting in an anisotropic underwater oleophobic surface. [ \n \n 71 \n \n ] The ordered structures encourage the flow of oil droplets in the head‐to‐tail direction but pin them in the tail‐to‐head direction, avoiding the advance of oil toward the head of the fish. Crucian carps have a mucus layer across a hierarchical scale structure consisting of rows of ridges (50–70 µm) across the scales decorated with small tubercles (2–3 µm) to produce underwater superoleophobicity. [ \n \n 72 \n \n ] Clam shells are another example of underwater superoleophobicity, with the rough inside region using a combination of the CaCO 3 shell composition and micro/nanohierarchical structures (Figure 2c ). [ \n \n 68 \n , \n 73 \n \n ] The hydrophobic shark skin provides antifouling ability which allows the animals to swim faster as drag is reduced. [ \n \n 74 \n \n ] The shark skin is seen to contain overlapping denticles with microscale riblets (Figure 2d ) as well as a mucus coating. [ \n \n 75 \n , \n 76 \n \n ] \n Figure 2 Oleophobicity in nature. a) Seaweed and an SEM image of its surface. Reproduced with permission. [ \n \n 69 \n \n ] Copyright 2015, Wiley‐VCH GmbH & Co. KGaA, Weinheim. b) A Filefish and an SEM image of the hook‐like spines on its skin surface. Reproduced with permission. [ \n \n 71 \n \n ] Copyright 2013, Wiley‐VCH GmbH & Co. KGaA, Weinheim. c) A Clam shell and an SEM image of the oleophobic region of its surface. Reproduced with permission. [ \n \n 73 \n \n ] Copyright 2012, Wiley‐VCH GmbH & Co. KGaA, Weinheim. d) A Shark and an SEM image of its skin structure. Reproduced with permission. [ \n \n 74 \n \n ] Copyright 2013, Wiley‐VCH GmbH & Co. KGaA, Weinheim. 1.3 Amphiphobicity in Nature In air superoleophobicity and superamphiphobicity are the most demanding of the nonwetting materials. The requirements to achieve such repellency are much more strenuous and convoluted than for superhydrophobicity or underwater superoleophobicity. Accordingly, the examples in nature are much more limited. The carnivorous pitcher plant operates through a system of zones with varying wettability's: the lid, the peristome, the waxy surface of the slippery zone, and the glandular surface of the digestive zone. The lid surface has contact angles comparable to the leaves of the plant with relatively high surface free energies, with the key function of these parts being insect attachment through adhesive forces. The peristome and glandular zones are hydrophilic and oleophilic, with a very high surface energy and large polar component. The hydrophilic film causes lubricating effects, ceasing the insect adhesion. The waxy zone traps and retains the pray. The hydrophobicity of the zones has been well studied, and the hydrophilicity of the peristome and hydrophobicity of the slippery zone are understood. [ \n \n 77 \n , \n 78 \n , \n 79 \n , \n 80 \n \n ] However, the oleophobicity of the zones remains largely unexplored, and to the best of our knowledge is only discussed by Gorb and Gorb. They report the slippery waxy zone of the Nepenthes alata to have high repellency to both water and some lower surface energy liquids, namely, diiodomethane and ethylene glycol. [ \n \n 81 \n \n ] In contrast to the lotus leaf effect of using microstructure cavities to block out liquids, the pitcher plant uses its microcavities to store a repellent liquid, resulting in a film over the surface and removing the air cushion, which proves ineffective against low surface tension droplets. [ \n \n 82 \n \n ] When insects try to stand on this surface, the oils on their feet are repelled causing the insects to be trapped within the plant. The pitcher plants have been a fundamental influence on the design of SLIPS (slippery liquid‐infused porous surfaces). [ \n \n 83 \n \n ] \n Leafhoppers (Insecta, Hemiptera, Cicadellidae) have proteins arranged in the shape of hollow spheres, known as brochosomes, ≈200–700 nm in diameter creating an overall porous honeycomb‐like structure ( Figure \n \n 3 \n ). [ \n \n 84 \n \n ] This results in repellency toward water with a W of up to 172° as well as some lower surface tension liquids including ethylene glycol and diiodomethane but not ethanol. The brochosomes are comprised of proteins and are released as colloidal suspensions which the insects spread over their skin with their legs. After the liquid has dried, the brochosomes are spread further during grooming and ultimately the insect integument becomes saturated with a thin pruinose layer of brochosomes. Figure 3 Amphiphobicity in nature. a) A Leafhopper (Athysanus argentarius). b) SEM images of agglomerated brochosomes from a Leafhopper across scales and model representations of a single brochosome and cross‐section. Reproduced with permission. [ \n \n 84 \n \n ] Copyright 2012, The Royal Society of Chemistry. An exceptional demonstration of natural superamphiphobicity is the Springtail (Collembola) ( Figure \n \n 4 \n ). Springtails are skin‐breathing arthropods that live in a range of soil environments, including largely polluted areas, and the intricate skin structure and coatings provide them with antifouling ability and nonwetting characteristics facilitating their survival and ease of movement in such environments. [ \n \n 85 \n , \n 86 \n , \n 87 \n \n ] The topography consists of an array of micro/nanoscale hexagonal or rhombic granules with overhanging re‐entrant structures. [ \n \n 88 \n \n ] The structures are composed of a porous chitin base with an epicuticular layer of proteins and a homogeneous lipid layer containing groups such as steroids, esters, triglycerides, and terpenes. The cavities provide a stable Cassie–Baxter state with a high energy barrier and are nonwettable toward water, methanol, ethanol, hexadecane, and tridecane, but not dodecane or hexane. [ \n \n 86 \n \n ] \n Figure 4 Amphiphobicity in nature. Cuticle patterns of different life forms and different patterns of Collembola. a–a″) Habitus and hexagonal structures of epedaphic Entomobryomorpha species. b–b″) Habitus and hexagonal structures of hemiedaphic Isotomidae (Entomobryomorpha). c–c″) Habitus and secondary granules with basic hexagonal structures of euedaphic Poduromorpha. d–d″) Habitus and cuticle of hemiedaphic Symphypleona. Reproduced with permission. [ \n \n 88 \n \n ] Copyright 2012, Springer. 1.4 Wetting Models Wetting models have been defined to explicitly categorize materials based on their wetting behavior. For a material to be classified as superhydrophobic/superoleophobic, it must exhibit a water/oil contact angle (W/OCA) >150°. [ \n \n 89 \n , \n 90 \n \n ] In the case of superamphiphobicity, the material must demonstrate WCAs and OCAs >150°, signifying its exceptional resistance to both polar and nonpolar substances. [ \n \n 91 \n \n ] Another crucial aspect of a material's wettability is the sliding angle (SA), which describes the angle of rotation of a surface at which the liquid droplet slides off. Ideally, for superrepellent surfaces the SA is below 10°, and the lower the better. For liquid–solid systems, the solid surface energy (s) and liquid tension (γ) control the strength of interaction between the two phases, which dictates the extent of spreading and wetting of liquids over solids. The wetting behavior of materials is further classified by the wetting states: the Wenzel state and the Cassie–Baxter state. [ \n \n 92 \n , \n 93 \n , \n 94 \n \n ] These derive from the Young's equation (Equation ( 1 )), which describes the behavior of a droplet on a completely smooth, ideal surface. [ \n \n 95 \n \n ] The equation is expressed in terms of surface tensions, where γ sv is the solid–vapor surface tension, γ sl is the solid–liquid surface tension, γ lv is the liquid–vapor interface energy, and θ Y is the Young's contact angle \n (1) \n γ s v = γ s l + γ l v cos θ Y \n \n Based on this, the theory can be developed to produce the Wenzel state. [ \n \n 96 \n \n ] This describes liquid interactions with rough surfaces and can be defined by Equation ( 2 ) where θ W is the Wenzel contact angle, r is the roughness parameter—the ratio of actual solid‐liquid contact area to the expected planar area\n \n (2) \n cos θ W = r cos θ Y \n \n Or, in terms of surface tensions (Equation ( 3 )) can be used\n \n (3) \n cos θ W = r γ sv − γ sl γ lv \n \n The Cassie–Baxter state characterizes rough structured surfaces on which the droplets can situate upon the features, trapping air/another material pockets below, usually resulting in strong hydrophobicity due to the upward force on the droplet provided by the trapped material in the cavities. The Cassie–Baxter state can be specified by Cassie's law shown in Equation ( 4 ). Where θ CB is the Cassie–Baxter contact angle and f \n s is the fraction of the total surface area that is wet. In cases where there are different surface structures, each has its own fraction, f i \n , where the sum of f i \n = 1\n \n (4) \n cos θ CB = r f s cos θ Y + f s − 1 \n \n The Cassie–Baxter state can also be portrayed with respect to the surface tension as shown in Equation ( 5 ). This equation represents systems containing air, rather than another material, between the surface and the liquid. Where γ \n i \n indicates the respective fraction for which the corresponding energy applies\n \n (5) \n cos θ CB = ∑ i n f i γ i , sv − γ i , sl γ lv \n \n 1.5 Applications The applications of liquid repellent materials span across a vast number of domains. Packaging is one of the largest areas for consumption of such materials. This can include packaging for food and drink, electronics, general transportation, as well as medicinal products. Also in the medical industry, surgical tools, hearing aids, and biomedical devices including catheters, stoma bags, and tubing mandate repellent properties to prevent contamination and maintain cleanliness and performance capability. Emulsifications and suspensions are also used in the medical industry as well as in cosmetics. Another application area is in electronic coatings, microfluidics, and moisture barriers to prevent short circuits. The agricultural industry relies on large scale superamphiphobic materials such as bale wrap, poly tunnels, feed bags, irrigation systems, and machinery coatings. In the case of textiles, clothing, outdoor equipment, and uniforms for hazardous environments all require superhydrophobic or superamphiphobic materials to perform their intended purpose of keeping the user dry and safe. In construction and infrastructure, bridges, building exteriors, sealants, and concretes can necessitate superamphiphobic materials to protect against weathering, water runoff, and pollutants. Smartphones, ATMs, all other displays, as well as vehicle interiors, handles, and common touch points can use oleophobicity to reduce smudging and contamination. Optoelectronics are another area where smudge resistance and antifouling are of particular interest. Self‐cleaning properties are induced by superamphiphobic materials, when liquid droplets bead and roll off a surface, they can remove debris and contaminants in the process. This is paramount for solar panels, windows, vehicles, aircraft, and spacecraft, to avoid fouling which can reduce the transparency and efficiency of these products as well as providing anti‐icing properties to enhance performance in extreme conditions. Materials with an affinity toward either oil or water and repellency toward the other are used in oil spill clean ups and environmental protection through oil/water separation, as well as coatings on marine vehicles and underwater sensors to provide antifouling and reduced drag properties allowing the vessels to operate as intended. The applications for superrepellent surfaces of all variations are extensive, and so using fluorocarbons to provide the necessary properties has caused significant detriment to the environment. A shift toward sustainable materials and methods is underway, particularly for superhydrophobic and underwater oleophobic materials. Creating sustainable superoleophobic and superamphiphobic materials invokes a more complex challenge, and with few successes on the laboratory scale, producing these materials industrially for these applications requires much more research and development as will be discussed in this review."
} | 9,452 |
27093048 | PMC5113844 | pmc | 92 | {
"abstract": "Reef-building corals possess a range of acclimatisation and adaptation mechanisms to respond to seawater temperature increases. In some corals, thermal tolerance increases through community composition changes of their dinoflagellate endosymbionts ( Symbiodinium spp.), but this mechanism is believed to be limited to the Symbiodinium types already present in the coral tissue acquired during early life stages. Compelling evidence for symbiont switching, that is, the acquisition of novel Symbiodinium types from the environment, by adult coral colonies, is currently lacking. Using deep sequencing analysis of Symbiodinium rDNA internal transcribed spacer 2 (ITS2) PCR amplicons from two pocilloporid coral species, we show evidence consistent with de novo acquisition of Symbiodinium types from the environment by adult corals following two consecutive bleaching events. Most of these newly detected symbionts remained in the rare biosphere (background types occurring below 1% relative abundance), but one novel type reached a relative abundance of ~33%. Two de novo acquired Symbiodinium types belong to the thermally resistant clade D, suggesting that this switching may have been driven by consecutive thermal bleaching events. Our results are particularly important given the maternal mode of Symbiodinium transmission in the study species, which generally results in high symbiont specificity. These findings will cause a paradigm shift in our understanding of coral- Symbiodinium symbiosis flexibility and mechanisms of environmental acclimatisation in corals.",
"introduction": "Introduction The eukaryotic and prokaryotic microbial communities (that is, the microbiome) associated with animals and plants have essential roles in their health and functioning ( McFall-Ngai et al. , 2013 ). Reef-building corals form symbioses with a wide range of microbial symbionts, including phototrophic dinoflagellates in the genus Symbiodinium . As the coral host depends on photosynthate for nutrition, a prolonged breakdown of the symbiosis (referred to as coral bleaching) often leads to coral death ( Baker, 2003 ). Episodes of mass coral bleaching have increased in frequency and intensity due to climate change and have caused a substantial loss in coral cover in many coral reef regions over the last few decades ( Hoegh-Guldberg, 1999 ; Hoegh-Guldberg et al. , 2007 ; De'ath et al. , 2012 ). The role of Symbiodinium symbionts in acclimatisation of the coral holobiont to environmental changes has been extensively covered in the recent literature ( Blackall et al. , 2015 ). The genus Symbiodinium is classified into nine phylogenetic clades (A through I) based on DNA sequence analysis, with a range of different types (putative species) within each clade ( Pochon and Gates, 2010 ). Symbiodinium types can be transmitted directly from parent to offspring via eggs (vertical transmission) or aposymbiotic larvae/early recruits can acquire their symbionts from the environment (horizontal transmission) ( Harrison and Wallace, 1990 ; van Oppen, 2001 ; Padilla-Gamino et al. , 2012 ). Different Symbiodinium types have distinct physiological optima and stress tolerance levels, which confer different phenotypes to their coral hosts. For instance, corals dominated by Symbiodinium clade D are generally more thermally tolerant compared with those predominantly associating with types in other clades ( Berkelmans and van Oppen, 2006 ). More than one Symbiodinium type can exist simultaneously within a single coral host ( Mieog et al. , 2007 ; Correa et al. , 2009 ; Silverstein et al. , 2012 ); these can occur in high abundance as ‘dominant types' or in very low abundance known as ‘background types', that is, the ‘ Symbiodinium rare biosphere'. In other microbial ecosystems, the rare biosphere represents a low-abundance, high-diversity group (in terms of numbers of operational taxonomic units) representing <1% of relative abundance ( Sogin et al. , 2006 ; Reid and Buckley, 2009 ). Therefore, in the present study, all Symbiodinium types that occurred below this threshold were considered members of the ‘ Symbiodinium rare biosphere'. The capacity of reef-building corals to host different symbionts (symbiotic flexibility) suggests two potential adaptive mechanisms to environmental changes: symbiont ‘shuffling' and ‘switching' ( Buddemeier and Fautin, 1993 ; Fautin and Buddemeier, 2004 ). Some corals have been shown to resist and/or recover from thermal stress through changes in the relative abundance of Symbiodinium types that constitute the in hospite community, that is, symbiont shuffling ( Baker et al. , 2004 ; Rowan, 2004 ). This acclimatisation response is well documented ( Baker et al. , 2004 ; Chen et al. , 2005 ; Berkelmans and van Oppen, 2006 ; Jones et al. , 2008 ; Baskett et al. , 2009 ), but is believed to be limited to the Symbiodinium types acquired vertically or horizontally in early life stages. Symbiont ‘switching' refers to a change in the in hospite Symbiodinium community due to the uptake of new Symbiodinium types from the environment, potentially from the water column and sediments ( Buddemeier and Fautin, 1993 ; Fautin and Buddemeier, 2004 ). Preliminary studies have indicated that adult corals are unable to form stable symbioses with exogenous algal symbionts; therefore, this mechanism is believed to occur only during a relatively short period of the coral larval and early juvenile life stages ( Goulet and Coffroth, 2003 ; Little et al. , 2004 ; Coffroth et al. , 2010 ). Testing of this hypothesis has been hampered, however, by the use of genetic methods that lack sensitivity to detect Symbiodinium types that occur below 5–10% of total relative abundance. Here, we challenge this notion by exploring the Symbiodinium rare biosphere using next-generation sequencing, a cost-effective, high-throughput method that has been recently shown to accurately detect low-abundance Symbiodinium types ( Quigley et al. , 2014 ; Thomas et al. , 2014 ; Arif et al. , 2014 ; Green et al. , 2014 ; Edmunds et al. , 2014b ). We assess Symbiodinium communities in a time-series sample set to investigate (1) the cryptic diversity of the Symbiodinium rare biosphere within two common pocilloporid species; (2) possible changes within the Symbiodinium community over a period of time that spans two successive bleaching events; and (3) whether Symbiodinium shuffling and/or switching has occurred in pocilloporid corals from a subtropical reef at Lord Howe Island (LHI), eastern Australia.",
"discussion": "Discussion Extraordinary Symbiodinium diversity and symbiotic flexibility in LHI reef-building corals In both coral species, the deep sequencing analysis revealed an extraordinary diversity within the Symbiodinium community. In fact, the diversity reported here is almost five times greater than that reported in other recent next-generation sequencing studies on Symbiodinium diversity ( Quigley et al. , 2014 ; Thomas et al. , 2014 ; Arif et al. , 2014 ; Green et al. , 2014 ; Edmunds et al. , 2014b ). For example, a study on Acropora coral species ( Thomas et al. , 2014 ) at another high latitude reef (Abrolhos Island, Western Australia) found a Shannon diversity of 0.145 (vs 0.620 at LHI in September 2012). The high Symbiodinium diversity as well as the endemicity of LHI coral-algal symbioses (mostly composed of previously undescribed ITS2 Symbiodinium types) support previous studies showing that the LHI Symbiodinium community is genetically and physiologically distinct ( Wicks, 2009 ; Wicks et al. , 2010 ; Noreen et al. , 2015 ). Our results highlight a high level of symbiont diversity within LHI subtropical corals, with a mean of 11 symbiont types per coral host. Although only Symbiodinium belonging to clade C have been previously detected in LHI corals using a gel electrophoresis-based method ( Wicks et al. , 2010 ), here we detected Symbiodinium types from clades A, B, C, D, F and G. The association of Symbiodinium clade B with S. pistillata and clade G with both P. damicornis and S. pistillata found here have not been previously observed. Nevertheless, the majority of the symbionts detected here were members of Symbiodinium clade C, which explains the high level of specificity to clade C reported previously ( Wicks et al. , 2010 ). Further research is needed to investigate whether different Symbiodinium clade C types simultaneously hosted by a single colony can provide different physiological performance and potentially enable acclimatisation, as previously suggested for clade C types in Caribbean corals ( Sampayo et al. , 2008 ). Temperature anomalies may drive fine-scale changes within the Symbiodinium community During the two bleaching events, we did not observe any changes within dominant types; however, the Symbiodinium rare biosphere showed a dynamic pattern where both shuffling and switching events were quite common during thermal stress and recovery periods ( Supplementary Figures S1 and S2 ). For instance, we observed the new appearance of 104 and 80 Symbiodinium types in P. damicornis and S. pistillata , respectively, over all sampling periods. The substantial changes observed in the Symbiodinium community of both coral species following each of the two bleaching events suggest that environmental disturbance drives symbiont community changes in LHI corals ( Buddemeier and Fautin, 1993 ; Fautin and Buddemeier, 2004 ; Berkelmans and van Oppen, 2006 ; Jones et al. , 2008 ; Silverstein et al. , 2015 ) and that symbiotic associations in species that show maternal symbiont transmission are more flexible than previously thought. This concurs with a recent study showing that corals that are sensitive to environmental conditions display high intra- and inter-species flexibility ( Putnam et al. , 2012 ). Interestingly, 18 months after the two bleaching events, the recovered coral colonies harboured a completely different Symbiodinium assemblage with new dominant and background types. We hypothesise that the newly acquired dominant Symbiodinium type (LHI_C.28), and the type that was already present in the rare biosphere at the first sampling time point (C_I:53), may be better adapted to cope with temperature anomalies and the potentially altered environmental conditions following such events. Notably, we observed a switching event to Symbiodinium clade D and 90% of the Symbiodinium rare biosphere members were also newly acquired in 2012, which may provide more options to cope with future bleaching events. These findings overthrow the notion that the period for uptake of algal endosymbionts is narrow and only limited to early life stages in these reef-building corals. Role and importance of members of the ‘ Symbiodinium rare biosphere' There is increasing evidence to suggest that members of Symbiodinium clade D can confer enhanced thermal tolerance to the coral holobiont compared with other clades ( Stat and Gates, 2011 ). Repopulation of recovering bleached coral hosts with clade D types has been reported as a survival mechanism for elevated sea temperatures ( Chen et al. , 2005 ; Berkelmans and van Oppen, 2006 ; Mieog et al. , 2007 ; Jones et al. , 2008 ; Stat et al. , 2013 ; Silverstein et al. , 2015 ). This mechanism has, however, been primarily attributed to shuffling of Symbiodinium D pre-existing in the rare biosphere rather than de novo acquisition. Although the newly acquired D types in LHI corals occurred at low relative abundance in our results, studies on other microbial communities have demonstrated that rare species can be metabolically very active ( Campbell et al. , 2011 ; Logares et al. , 2014 ). It has also been shown that rare functionally important species can become dominant to maintain the integrity of functional processes when environmental conditions change ( Shade et al. , 2014 ). Moreover, a network theoretic modelling approach on coral- Symbiodinium associations under climate change ( Fabina et al. , 2013 ) predicts that both elevated symbiont diversity and types occurring at low abundance, which provide redundant or complementary symbiotic function, can significantly increase community stability in response to environmental change. Hence, following these predictions, our results indicate that the switch to clade D in the Symbiodinium rare biosphere and the increase in symbiont diversity documented here in LHI corals may enhance the ability of these corals to resist and/or recover from future bleaching events. The repopulation with previously undetectable clade D was also documented in an experimental study following two induced bleaching events ( Silverstein et al. , 2015 ). Even though the source of these newly dominant types could not be identified (from the rare biosphere or from the environment), the authors found an increase in the host thermotolerance and concluded that members of the Symbiodinium rare biosphere can be critical components of coral recovery ( Silverstein et al. , 2015 ). Similarly, the newly acquired Symbiodinium clade D documented here could increase their hosts' thermotolerance during future bleaching events. It is now well-established that the rare biosphere has significant ecological roles in ecosystems such as diazotrophic bacteria in seawater, bacterial and archaeal ammonia oxidisers in soils, methanogens in intestines ( Shade et al. , 2014 ), marine planktonic microeukaryotes in the ocean ( Logares et al. , 2014 ) and, our findings suggest the same is true for reef-building corals. Implications of symbiont switching for reef-building coral community structure Climate change is responsible for changes in species composition and population structure ( Ateweberhan et al. , 2013 ). In coral reef ecosystems, in particular, the general trend is the loss of stress-sensitive coral species and replacement by stress-tolerant species that survive the disturbance. For example, a study conducted over a 14-year period that included two thermal stress events (in 1998 and 2001) at the high latitude reef of Sesoko Island (Okinawa, Japan) reported a complete change in the coral community structure ( van Woesik et al. , 2011 ). The stress-sensitive branching pocilloporid corals were replaced by stress-tolerant massive corals such as poritids and brain corals. Our study suggests that symbiont switching to more thermally tolerant symbionts in the two pocilloporid coral species has the potential to assist the persistence of these environmentally-sensitive coral species over time. Given that the frequency of thermal stress events is predicted to increase (IPCC 2014), these findings have important implications for predicting coral assemblage recovery after mass bleaching events and will also help to refine evolutionary models that predict the future of coral reefs."
} | 3,745 |
21574576 | null | s2 | 93 | {
"abstract": "As a promising biomaterial with numerous potential applications, various types of synthetic spider silk fibers have been produced and studied in an effort to produce man-made fibers with mechanical and physical properties comparable to those of native spider silk. In this study, two recombinant proteins based on Nephila clavipes Major ampullate Spidroin 1 (MaSp1) consensus repeat sequence were expressed and spun into fibers. Mechanical test results showed that fiber spun from the higher molecular weight protein had better overall mechanical properties (70 KD versus 46 KD), whereas postspin stretch treatment in water helped increase fiber tensile strength significantly. Carbon-13 solid-state NMR studies of those fibers further revealed that the postspin stretch in water promoted protein molecule rearrangement and the formation of β-sheets in the polyalanine region of the silk. The rearrangement correlated with improved fiber mechanical properties and indicated that postspin stretch is key to helping the spider silk proteins in the fiber form correct secondary structures, leading to better quality fibers."
} | 280 |
29379177 | PMC5864192 | pmc | 94 | {
"abstract": "Coastal oceans are increasingly eutrophic, warm and acidic through the addition of anthropogenic nitrogen and carbon, respectively. Among the most sensitive taxa to these changes are scleractinian corals, which engineer the most biodiverse ecosystems on Earth. Corals’ sensitivity is a consequence of their evolutionary investment in symbiosis with the dinoflagellate alga, Symbiodinium . Together, the coral holobiont has dominated oligotrophic tropical marine habitats. However, warming destabilizes this association and reduces coral fitness. It has been theorized that, when reefs become warm and eutrophic, mutualistic Symbiodinium sequester more resources for their own growth, thus parasitizing their hosts of nutrition. Here, we tested the hypothesis that sub-bleaching temperature and excess nitrogen promotes symbiont parasitism by measuring respiration (costs) and the assimilation and translocation of both carbon (energy) and nitrogen (growth; both benefits) within Orbicella faveolata hosting one of two Symbiodinium phylotypes using a dual stable isotope tracer incubation at ambient (26 °C) and sub-bleaching (31 °C) temperatures under elevated nitrate. Warming to 31 °C reduced holobiont net primary productivity (NPP) by 60% due to increased respiration which decreased host %carbon by 15% with no apparent cost to the symbiont. Concurrently, Symbiodinium carbon and nitrogen assimilation increased by 14 and 32%, respectively while increasing their mitotic index by 15%, whereas hosts did not gain a proportional increase in translocated photosynthates. We conclude that the disparity in benefits and costs to both partners is evidence of symbiont parasitism in the coral symbiosis and has major implications for the resilience of coral reefs under threat of global change.",
"introduction": "Introduction Animal-microbe symbioses are common in the ocean, and are often associated with extreme environmental biogeochemistry. For example, deep-sea tubeworms host microbial symbionts that are capable of generating energy and biomass via chemoautotrophy of rich inorganic mineral substrates in the complete absence of sunlight. Conversely, sunlight is abundant in the shallow tropical oceans, while inorganic nutrients are limiting to ecosystem productivity. Therefore, partnerships between animals and autotrophic microbes like cyanobacteria and protists confers the host with the ability to harness light energy and transform myriad forms of nitrogen into the building blocks for important biomolecules like enzymes, proteins, and nucleic acids [ 1 ]. Through symbiosis, benthic animals such as sponges and corals have become some of the most successful competitors for space in the tropical seas. However, the conditions under which such symbioses evolved are rapidly changing in the Anthropocene [ 2 ]. The coral symbiosis, like many obligate symbiotic associations, is sensitive to environmental change. Rising sea surface temperatures have a direct impact on coral respiration and energy budgets. Coastal eutrophication is worsening which acts synergistically with thermal stress to exacerbate coral bleaching and alters the competitive advantages corals’ nutritional symbioses convey [ 3 ]. As a consequence, reefs are experiencing a perilous global decline which may be irreversible [ 2 , 4 ]. Under such conditions it has been increasingly suggested that the myriad species of coral-hosted Symbiodinium can be placed along a continuum from mutualist to parasite, with the latter implying that the symbionts parasitize their host of resources and reduce host fitness [ 5 , 6 ]. Indeed, several examples of this have been documented for stress-tolerant lineages of Symbiodinium , including phylotypes from the A and D ITS2 clades. However, evidence of parasitism and cheating in the coral nutritional symbiosis is sparse with the best examples revealing long-term reductions in host fitness associated with certain Symbiodinium types [ 5 – 8 ]. Nutritional exchanges are critical to understanding the performance of a symbiosis under stress. For corals, there are two simplified perspectives on the relative importance of mechanisms within the holobiont that underpin and precede the bleaching response. A “host-centered” perspective reveals that host species’ capacity for heterotrophy and lipid storage confers thermal tolerance, bleaching resistance, and post-bleaching resilience [ 9 ]. In sum, hosts that can store more energy and/or obtain nutrition from heterotrophic sources during thermal stress are more likely to survive bleaching events. A second perspective is “symbiont-centered“ and is based on thermal stress impairing photosynthetic performance of Symbiodinium by damaging the photosynthetic architecture including the thylakoid membrane [ 10 , 11 ]. When combined with Symbiodinium ’s propensity for photorespiration [ 12 ], damaged membranes produce reactive oxygen species [ 13 ] that induce host apoptosis, immune responses and ultimately, symbiont expulsion. However, some Symbiodinium tolerate temperatures exceeding 34 °C in hospite or even after expulsion (far higher than most corals’ bleaching threshold) and exhibit enhanced growth rates [ 14 , 15 ]. These exceptional observations challenge the theory that Symbiodinium are universally impaired at high temperatures and suggest that other factors in addition to photodamage are involved in the bleaching outcome. Wooldridge [ 16 ] presented an integrative model for the breakdown of the coral symbiosis, detailing how mutualistic symbionts might shift to parasitism in the warming period preceding bleaching [ 16 ]. The model is based on the need for the coral host to supply exogenous carbon for Symbiodinium photosynthesis. Although CO 2 can be sourced from host respiration of heterotrophic substrates, in many species the abundant ATP derived from photosynthesis is sufficient to sustain a positive feedback on the activity of carbonic anhydrase, which converts seawater bicarbonate into intracellular CO 2 . The diversity of active carbon concentrating mechanisms found in corals highlights the importance of this feedback [ 17 ]. Provided that symbiont growth is limited by lack of nutrients, cell division is minimized and hosts can accumulate translocated photosynthates as energy reserves to maintain symbiotic homeostasis. However, chronic metabolic imbalances caused by environmental perturbation, particularly eutrophication, but also high p CO 2 and irradiance favor the proliferation of Symbiodinium within the host [ 18 , 29 ]; Fagoonee 1999). As such, Symbiodinium are predicted to retain more carbon (as molecular skeletons for amino and nucleic acids) in order to sequester more nitrogen for cell division [ 19 ]. Thus, the flow of energy to the host is lessened, forcing increased respiration of lipid reserves to maintain metabolic functioning. At this point, host investments in supplying exogenous CO 2 for photosynthesis no longer yield returns as the flow of carbon ends in the Symbiodinium sink. Consequently, the likelihood of bleaching increases as a result of a weakened host unable to tolerate the burgeoning symbiont population which becomes primed for bleaching when temperature and irradiance are high [ 13 , 20 – 22 ]. As such, there is a presumption that the mutualistic symbiont has become more parasitic by “cheating” the host of resource subsidies while reducing host fitness. The transition from symbiont to parasite can be defined by increasing costs that exceed benefits gained by the host [ 23 , 24 ] and may be more evident in phylotypes that demonstrate “cheating” behaviors. Indeed, Lesser et al. [ 5 ] suggested that cheating in corals is evidenced by the observation of the temporary dominance of “infectious” and opportunistic Symbiodinium species such as S. trenchii . The competitive displacement and/or host-modulated winnowing of “inferior” symbionts further suggests that the host can influence this process, perhaps by optimizing subsidies for mutualists or implementing sanctions on parasites [ 49 ] (McIlroy & Coffroth 2017). These processes are thought to be mediated by symbiont competition, which is predicted to vary with environmental factors such as heat, light, and nutrient stress [ 25 , 26 ]. However, many of these theoretical processes lack empirical data for corals. The importance of filling this gap is urgent given that synergistic effects of temperature and nutrient stress reduce coral reef resilience. Wooldridge and Done (2009) showed that on Australia’s Great Barrier Reef, nearshore coral reefs exposed to chronic terrestrial nitrogen pollution bleached at temperatures 2 °C lower than offshore oligotrophic reefs. This finding implies that management policies aimed at reducing the future risk of coral bleaching need to consider both “local” nutrient pollution and “global” climate change minimization strategies [ 27 ]. In addition to this correlative field observation, a recently developed theoretical biogenergetic simulation model [ 28 ] was found to support the mechanistic bleaching predictions of the Wooldridge model. However, targeted experimental studies that directly test/validate the Wooldridge model are currently lacking. Only indirect observations are available for discrimination. For example, at the “front end” of the model, studies have shown that Symbiodinium growth is stimulated by exogenous nitrogen [ 29 , 30 ]. At the “back end” of the model, studies have confirmed that nitrogen has a direct effect on coral health [ 11 , 31 , 32 ], and that a large symbiont burden can make corals more vulnerable to bleaching and disease [ 3 , 18 ]. Together, these bodies of evidence suggest that symbiont benefits are not synonymous with host benefits and that parity is lost under times of stress; an indication of the unsustainable cost of symbiosis. Yet, none of these studies have examined the capacity for symbionts to cheat or parasitize their hosts which requires a focus on metabolism and nutrient assimilation and translocation during periods of stress. Here we ask the question, do Symbiodinium become parasites at elevated temperatures?",
"discussion": "Discussion We exposed O. faveolata sourced from the shallow and deep limits of their distribution to an ecologically relevant pre-bleaching warming experiment (+5 °C over 10 days [ 33 ] followed by a snapshot assessment of biogeochemical cycling including determinations of holobiont productivity (NPP), Symbiodinium carbon and nitrogen assimilation, and their subsequent translocation to the host. Previous studies focused on carbon metabolism in scleractinian corals under varied environmental conditions have reported a negative impact of temperature on Symbiodinium photophysiology [ 10 , 38 ]. In contrast to those works, we observed that Symbiodinium metabolism was increased at sub-bleaching temperatures (31 °C) despite the fact that the O. faveolata holobiont had a 60% reduction in NPP relative to controls (Table 1 ). At the same time, photosynthesis (as measured by 13 C fixation) was constant among deep/C7 corals and increased 14% among shallow/A3 corals (Fig. 1a ). Thus, warming did not impair Symbiodinium photosynthesis but increased holobiont respiration which collectively caused a significant drop in holobiont NPP. Therefore, we conclude that warming induces greater energetic demands on the holobiont before Symbiodinium performance is compromised [ 16 , 39 , 40 ]. Moreover, this increase in respiration was associated with a significant (15%) drop in host carbon content, but not in Symbiodinium , suggesting that hosts were bearing the costs of elevated respiration, perhaps by catabolizing lipid reserves. Although the host was still receiving fixed carbon from their symbionts, the overall loss of carbon in host tissues is evidence of passing a tipping point, where the benefits of symbiosis no longer exceeded the costs; a sign of a parasitic interaction [ 24 ]. This confirms that during thermal anomalies coral hosts are rapidly taxed by unsustainable respiration that is not offset by photosynthetic benefits. A logical outcome is that a weakened host would be less capable to mitigate oxidative damage induced by a large and healthy symbiont burden [ 22 ]. By sampling 2 holobionts adapted to conditions at the extreme ends of the host species’ depth distribution we were able to assess the relative importance of light adaptation, Symbiodinium phylotype and temperature on productivity. Relative to deep/C7, shallow/A3 O. faveolata had twice the Symbiodinium density, but similar chla concentrations per cell (Table 1 ). This suggests that photoprotective mechanisms (e.g., self-shading and host pigments for non-photochemical quenching) were upregulated at shallow depths. In comparison, deep/C7 corals were duller in appearance, and had fewer Symbiodinium with no reduction in chlorophyll per cell, which is indicative of light-limiting conditions [ 41 ]. As such, deep/C7 corals brought to surface light conditions had 34% more carbon assimilation than shallow/A3 corals owing to more efficient use of light, and showed no response to warming (Fig. 1a ). In contrast, warming caused a 14% increase in carbon assimilation in shallow/A3 corals illustrating that heat can increase photosynthesis even when photoprotective mechanisms are present. Thus, host regulation of photoprotective mechanisms appears to have a stronger influence on symbiont photosynthesis than warming alone. The host’s ability to regulate symbiont photosynthesis via protection and inhibition is a plausible negative feedback mechanism for bleaching resistance. Such mechanisms may help to explain the different bleaching tolerances observed across the diversity of coral holobionts. However, warming had no proportional benefit to the host carbon subsidy, indicating that Symbiodinium either retained their additional carbon or it was respired by the holobiont (Fig. 1b ). In contrast to carbon, Symbiodinium nitrogen assimilation was universally enhanced with warming with no proportional increase in nitrogen obtained by the coral host (Fig. 1c, d ). Deep/C7 corals assimilated >37% more nitrogen than their shallow conspecifics at both temperatures. This can be attributed to either (1) Symbiodinium identity, which is not mutually exclusive from, or (2) greater photosynthetic efficiency, or (3) different nitrogen exposure histories or (4) a higher nitrogen demand for PSII repair. More importantly, warming and depth/clade were equal in their effect on Symbiodinium nitrogen assimilation. Both deep/C7 and shallow/A3 Symbiodinium assimilated >31% more nitrogen when warmed to 31 °C. Although the coral host still benefitted from symbiont nitrogen translocation, there was no proportional increase in translocation due to either warming or depth/clade. Given that nitrate assimilation is an energy intensive process, Symbiodinium effectively lowered their cost by maintaining a fixed subsidy to the host and retaining the surplus assimilated nitrogen [ 23 , 42 ]. This finding has several major implications. It suggests that when nitrogen is not limiting to primary productivity that warming (1) favors the assimilation of nitrogen and thus, the proliferation of Symbiodinium [ 29 , 31 ]; (2) nitrate increasingly “evades” host mechanisms (via passive transport, perhaps) to modulate Symbiodinium access to nitrogen, as proposed for organic nitrogen and ammonium [ 42 , 43 ], and (3) Symbiodinium either translocate a fixed nitrogen subsidy to their host, or hosts are physically or biochemically limited in the amount of nitrogen they can receive. Taken together, these data validate models that predict a synergistic effect of warming and nitrogen on selfish symbiont behavior and coral bleaching [ 16 ]. Indeed, the most significant pattern we observed was the increased disparity between host and symbiont carbon and nitrogen assimilation at sub-bleaching temperatures. The difference in resource acquisition between the partners (ΔAP 15 N symbiont-host and ΔAP 13 C symbiont-host ) showed that the nitrogen gap increased by >67% for all holobionts (Fig. 2c, d ), and the carbon gap increased by 39% in shallow/A3 corals at 31 °C (Fig. 2a ). That deep/C7 corals had similar ΔAP 13 C symbiont-host at both temperatures could be interpreted as evidence of a stable mutualistic interaction, but we caution that such comparisons between symbiont types are confounded by the physiological adaptations to the deep and shallow forereef. Regardless, these differences were evident after just one day of incubation, and we predict that they would compound over time as Symbiodinium cell division increases. We simulated such an increase based on the initial cell densities and final mitotic index of our incubated clones, which revealed an interesting pattern (Fig. 3 ). With just a 1% higher MI stimulated by heat and abundant resources, cell densities can theoretically increase in shallow corals by 25% in just 20 days, exceeding 5 million cells cm −2 . In deep corals with lower initial symbiont abundance, the time to reach a similar cell density may exceed 45 days. Such densities have been observed in O. faveolata prior to bleaching events [ 38 ] and may result from the synergistic effects of elevated nutrients and temperature. Indeed, Symbiodinium quickly assimilated nitrate and proportionally fewer organic derivatives were translocated to the host at elevated temperatures. Therefore, during weeks of heat accumulation preceding thermal bleaching events Symbiodinium can sequester significantly more nitrogen and energy than their hosts, which they can subsequently utilize for cell division [ 29 ]. A logical next step is to link our observations with long-term changes in symbiont population densities ( sensu [ 22 ]). Together, the observation of increased holobiont respiration and reduced carbon content (i.e., increased cost to the host) and unchanged translocation (i.e., no net change in cost to the symbiont) and increased resource storage and mitotic index (increased symbiont benefit) is indicative of the mutualism gradually shifting to parasitism [ 5 ]. Moreover, these observations validate the assumption that symbiont resource hoarding precedes the proliferation of an unsustainable symbiont burden on the host as predicted by the Wooldridge model. Fig. 3 Hypothetical growth curves of Symbiodinium density based on initial cell densities and final mitotic index (MI) values following 10 days of incubation at elevated temperature and nutrients, for corals adapted to the deep and shallow forereef environment. The ‘high symbiont burden’ threshold is derived from Kemp et al. [ 38 ]. For model equations and assumptions, please refer to Supplemental Materials . Although this study is limited to a single coral host ( O. faveolata ), we were able to evaluate the metabolic costs and nutritional benefits of hosting Symbiodinium partners that span the continuum from a sensitive mutualist to stress-tolerant, opportunistic parasite. S. fitti (ITS2 clade A3) are known for their tolerance of high light and increasing temperature, respectively [ 6 ], while being a less beneficial partner to their hosts than clade C Symbiodinium [ 5 , 44 ]). Indeed, Symbiodinium within the ITS2 clade C are generally considered to be more mutualistic as they confer more fitness benefits to their hosts, including carbon [ 45 ] and growth limiting nitrogen [ 26 ]. C7 is one such mutualist, which is specific to and dominates Orbicella in low-light habitats though is less tolerant of thermal stress [ 38 , 46 ]. Yet, we observed enhanced symbiont resource uptake (i.e., revenue) with unchanged host translocation (costs), which led to a higher net “profit” in the form of accumulated resources that led to a concomitant increase in cell division (benefits) by both of these symbionts in hospite under thermal stress. At the same time, their hosts sustained a large increase in respiration, which was associated with a reduction in carbon content. Taken together, these observations suggest that warming can create conditions for symbiont parasitism. Indeed, the functional significance of Symbiodinium diversity in the coral symbiosis is a critically important question to resolve if we are to understand the tolerance, adaptive capacity, and resilience of corals to a warming world [ 44 ]. We predict that a truly mutualistic symbiont would confer more resources to their host at a wider range of temperatures, whereas a commensal or parasite would retain more resources, particularly when conditions optimize their fitness [ 5 , 47 ]. Here we have observed the latter, where both an opportunistic and stress-tolerant symbiont and a purported low-light adapted mutualist have both become selfish as increasing temperatures favored the retention of resources while simultaneously increasing the respiratory demands of the holobiont. Future work should focus on how these patterns vary across the diversity of host–symbiont interactions, and hosts which vary in their trophic plasticity [ 9 ]. In conclusion, we have experimentally shown that warming reduces the benefits of symbiosis when nutrients are non-limiting, which bolsters calls for nutrient pollution to be included with greenhouse gas emissions in climate change mitigation policies that aim to preserve coral reefs [ 3 , 27 , 48 ]."
} | 5,370 |
22012252 | null | s2 | 95 | {
"abstract": "The two Flag/MaSp 2 silk proteins produced recombinantly were based on the basic consensus repeat of the dragline silk spidroin 2 protein (MaSp 2) from the Nephila clavipes orb weaving spider. However, the proline-containing pentapeptides juxtaposed to the polyalanine segments resembled those found in the flagelliform silk protein (Flag) composing the web spiral: (GPGGX(1) GPGGX(2))(2) with X(1) /X(2) = A/A or Y/S. Fibers were formed from protein films in aqueous solutions or extruded from resolubilized protein dopes in organic conditions when the Flag motif was (GPGGX(1) GPGGX(2))(2) with X(1) /X(2) = Y/S or A/A, respectively. Post-fiber processing involved similar drawing ratios (2-2.5×) before or after water-treatment. Structural (ssNMR and XRD) and morphological (SEM) changes in the fibers were compared to the mechanical properties of the fibers at each step. Nuclear magnetic resonance indicated that the fraction of β-sheet nanocrystals in the polyalanine regions formed upon extrusion, increased during stretching, and was maximized after water-treatment. X-ray diffraction showed that nanocrystallite orientation parallel to the fiber axis increased the ultimate strength and initial stiffness of the fibers. Water furthered nanocrystal orientation and three-dimensional growth while plasticizing the amorphous regions, thus producing tougher fibers due to increased extensibility. These fibers were highly hygroscopic and had similar internal network organization, thus similar range of mechanical properties that depended on their diameters. The overall structure of the consensus repeat of the silk-like protein dictated the mechanical properties of the fibers while protein molecular weight limited these same properties. Subtle structural motif re-design impacted protein self-assembly mechanisms and requirements for fiber formation."
} | 464 |
27064312 | null | s2 | 96 | {
"abstract": "Spider silks have unique mechanical properties but current efforts to duplicate those properties with recombinant proteins have been unsuccessful. This study was designed to develop a single process to spin fibers with excellent and consistent mechanical properties. As-spun fibers produced were brittle, but by stretching the fibers the mechanical properties were greatly improved. A water-dip or water-stretch further increased the strength and elongation of the synthetic spider silk fibers. Given the promising results of the water stretch, a mechanical double-stretch system was developed. Both a methanol/water mixture and an isopropanol/water mixture were independently used to stretch the fibers with this system. It was found that the methanol mixture produced fibers with high tensile strength while the isopropanol mixture produced fibers with high elongation."
} | 217 |
33311595 | PMC7733507 | pmc | 97 | {
"abstract": "Reservoir computing (RC) is a machine learning algorithm that can learn complex time series from data very rapidly based on the use of high-dimensional dynamical systems, such as random networks of neurons, called “reservoirs.” To implement RC in edge computing, it is highly important to reduce the amount of computational resources that RC requires. In this study, we propose methods that reduce the size of the reservoir by inputting the past or drifting states of the reservoir to the output layer at the current time step. To elucidate the mechanism of model-size reduction, the proposed methods are analyzed based on information processing capacity proposed by Dambre et al. (Sci Rep 2:514, 2012). In addition, we evaluate the effectiveness of the proposed methods on time-series prediction tasks: the generalized Hénon-map and NARMA. On these tasks, we found that the proposed methods were able to reduce the size of the reservoir up to one tenth without a substantial increase in regression error.",
"introduction": "Introduction Efficiently processing time-series data is important for various tasks, such as time-series forecasting, anomaly detection, natural language processing, and system control. Recently, machine-learning approaches for these tasks have attracted much attention of researchers and engineers because they not only require little domain knowledge but also often perform better than traditional approaches. In particular, machine-learning models that employ recurrent neural networks, such as long short-term memory, have achieved great success in natural language processing and speech recognition 1 , and their fields of applications continue to expand. However, the standard learning algorithms for recurrent neural networks, which include backpropagation through time 2 and its variants 3 , require large computational resources. These computational burdens often hinder real-world applications, especially when computing is performed near end users or data sources instead of data centers. Such computing has been attracting considerable interest because the amount of data often exceeds the network bandwidth capacity, which leads to network congestion and makes it difficult to efficiently send data to data centers. In addition, transferring personal data across networks is often avoided due to privacy issues. This new computing paradigm is called “edge computing,” which is characterized by limited computational power and limited battery capacity 4 , 5 . Reservoir computing (RC) is a machine-learning algorithm that aims to reduce the computational resources required for predicting time series without reducing accuracy. As shown in Fig. 1 , a typical RC consists of three parts: an input layer, a “ reservoir ” layer where neurons are randomly connected, and an output layer 6 , 7 . Because only the weights between the reservoir layer and the output layer are trained while the other weights remain fixed, the learning process of RC is much faster than that of backpropagation through time 8 – 10 . Therefore, RC is expected to be a lightweight machine-learning algorithm that enables machine learning in edge computing 11 . Figure 1 Typical RC architecture. The reservoir layer consists of randomly connected neurons. The connections between the input and reservoir layer \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$W^\\text {in}$$\\end{document} W in and connections within the reservoir layer \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$W^\\text {res}$$\\end{document} W res are fixed (solid arrows), whereas the output weights \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$W^\\text {out}$$\\end{document} W out are trained (dashed arrows). The RC training process is fast and accurate. In addition, RC has shown high performance on various time-series forecasting tasks, including chaotic time-series 12 – 14 , weather 15 , wind-power generation 16 , and finance 17 . Moreover, the range of applications of RC has extended into control engineering 18 , 19 and video processing 20 – 22 . To develop the applications for RC in edge computing, its hardware implementation must be improved to enhance its computational speed and energy efficiency. For realizing such efficient hardware implementation, variants of RC models, some of which employ delay-feedback systems 23 , simple network topologies such as ring-topology and delay lines 24 – 26 , and billiard systems 27 , have been proposed. Efficient hardware based on these variants have been implemented using field programmable integrated gate arrays (FPGAs) 28 – 31 . Moreover, numerous types of implementation employing physical systems, such as photonics 32 – 34 , spintronics 35 , mechanical oscillators 36 , and analog integrated electronic circuits 37 , 38 , have been demonstrated 39 . Although these implementations have exhibited the superiority of RC in computational speed and energy efficiency, the maximum size of the reservoir and, in turn, the forecasting accuracy, is limited by the physical size of the hardware. In this study, we propose three methods that reduce the size of the reservoir without any performance impairment. The three methods share the concept that the number of the effective dimension of the reservoir is increased by allowing additional connections from the reservoir layer at multiple time steps to the output layer at the current time step. We analyze the mechanism of the proposed methods based on the information processing capacity (IPC) proposed by Dambre et al. 40 . We also demonstrate how the proposed methods reduce the size of the reservoir in the generalized Hénon-map and NARMA tasks.",
"discussion": "Discussion In this study, we proposed three methods to reduce the size of an RC reservoir without impairing performance. To elucidate the mechanism of the proposed methods, we analyzed the IPC. We found that the value of the total IPCs almost reaches \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$N^\\text {res} (P+1)$$\\end{document} N res ( P + 1 ) using the proposed methods, whereas the importance of their components (the first-, third-, and fifth-order IPCs) changes drastically. We also found that the delay structures of the IPCs depend on the values of Q and P . To investigate the applicability of the proposed methods on complex data, we presented the experimental results on generalized Hénon-map and NARMA tasks. We found that when the target task has a relatively simple temporal structure, as demonstrated with the Hénon-map tasks, selecting an appropriate value of Q enhances the performance substantially. In contrast, when the target task contains complex temporal structure, as demonstrated in the NARMA tasks, adjusting the value of Q does not enhance the performance. However, in those cases, we found that increasing the value of P can reduce the size of the reservoir without impairing performance. We have demonstrated that the number of neurons in the reservoir can be reduced by up to one tenth in the NARMA10 task. Here, we note the relationship between our work and the relevant previous works 23 , 49 . In Ref. 23 , the authors proposed the time-delay reservoir that consists of a single node with delayed feedback. The states of the single node at multiple time steps correspond to virtual nodes, which may look similar to the delay-state concatenation. However, these two methods are different because the virtual nodes have output connections only to the current outputs, whereas the delay-state concatenation adds output connections from the nodes in the reservoir not only to the current outputs but also to future outputs. We also note that the delay-state concatenation can be also applied to the time-delay reservoir. In Ref. 49 , the authors proposed a method that is similar to delay-state concatenation. Their proposed model corresponds to the case when the number of additional connected past states is one (i.e., \\documentclass[12pt]{minimal}\n\t\t\t\t\\usepackage{amsmath}\n\t\t\t\t\\usepackage{wasysym} \n\t\t\t\t\\usepackage{amsfonts} \n\t\t\t\t\\usepackage{amssymb} \n\t\t\t\t\\usepackage{amsbsy}\n\t\t\t\t\\usepackage{mathrsfs}\n\t\t\t\t\\usepackage{upgreek}\n\t\t\t\t\\setlength{\\oddsidemargin}{-69pt}\n\t\t\t\t\\begin{document}$$P=1$$\\end{document} P = 1 ). They observed that performance enhancement depends on the value of Q , as we showed in this paper. However, to the best of our knowledge, the dependence of the performance on the value of P has not been reported. In addition, the other two proposed methods, drift-state concatenation and delay-state concatenation with transient states, are introduced for the first time in this paper. Moreover, the authors of Ref. 49 explained the mechanism of their proposed method in terms of the delayed embedding theorem 50 . In contrast, we have provided a more intuitive explanation based on the IPC 40 . Because the proposed methods do not assume a specific topology for the reservoir, they can readily be implemented in FPGAs and physical reservoir systems, such as photonic reservoirs 39 . Therefore, the proposed methods could be an important set of techniques that facilitates the introduction of RC in edge computing."
} | 2,499 |
28580425 | PMC5451198 | pmc | 98 | {
"abstract": "Stretchable, transparent nanogenerator enabled by ionic hydrogel converts motion energy into electricity and senses touch pressure.",
"introduction": "INTRODUCTION The past decade has witnessed the rapid growth of flexible/stretchable electronics, with the advent of various revolutionary multifunctional devices ranging from flexible transistors ( 1 , 2 ) and integrated circuits ( 3 , 4 ), stretchable luminescence devices ( 5 , 6 ), and roll-up displays ( 7 ) to smart sensor-integrated electronic skins ( 8 – 10 ). These advancements impose the challenge on corresponding power devices that they should have comparable flexibility/stretchability. For example, stretchable and transparent actuator ( 11 ) and touch panel ( 12 ) have been recently demonstrated, but no reported energy device can simultaneously achieve the high transparence and stretchability. Meanwhile, the growing wearable consumer electronics, either integrated in clothes/wears or attached on or implanted in curved human body, rely on power devices that are stretchable, shape-adaptive, and biocompatible. Because of the intrinsic energy conversion mechanism, it is hard for some energy harvesters to achieve high stretchability, for example, the strong magnetic field required for conventional electromagnetic generator; on the contrary, the recently developed triboelectric nanogenerator (TENG) is naturally flexible and has potential for high stretchability ( 13 – 16 ). The TENG, converting mechanical energy into electricity based on the coupling effect of contact electrification and electrostatic induction, has been demonstrated to be versatile in harvesting different types of energies and has the advantages of simple structure, vast material choice, and low cost ( 17 – 22 ). Several stretchable TENGs have been recently reported ( 23 – 28 ) with similar strategy as most reported stretchable devices, which are enabled by anchoring percolated networks of conductive materials (carbon nanotubes, graphene, carbon paste, silver nanowires, etc.) on prestrained elastomer substrates. However, the stretchability or ultimate strain (ε ult ) for this strategy is limited, typically below the ultimate strain of the elastomer (for example, 400 to 700% for silicones), due to the markedly increased sheet resistance when being stretched and the loss of percolation at large strain. Another recently reported stretchable TENG achieved ~300% tensile strain by using water or ionic liquid as the electrode, but its application is limited by the use of liquid materials ( 29 ). Hydrogels, composed of hydrophilic polymer networks swollen with water or ionic aqueous solution, are in solid form, soft, stretchable (ε ult , ~2000%), and biocompatible ( 30 ). Some hydrogels are transparent in full visible spectrum ( 11 ). Meanwhile, the increments of their resistivity with strain are orders of magnitude lower than those of percolated conductive networks ( 11 ). Many stretchable functional devices have been demonstrated with ionic hydrogels as the electrode conductors, such as strain sensors ( 31 ), loudspeakers ( 11 ), and electroluminescent devices ( 32 ). Here, we report a soft skin-like TENG (STENG) that enables both biomechanical energy harvesting and touch sensing by using elastomer and ionic hydrogel as the electrification layer and electrode, respectively. Different from previously reported TENGs using electrical conductors as the electrode, this soft STENG uses ionic conductors as the electrode. Polyacrylamide (PAAm) hydrogel containing lithium chloride (LiCl) is used as the ionic hydrogel (PAAm-LiCl hydrogel), and two commonly used elastomers, that is, polydimethylsiloxane (PDMS) Sylgard 184 and 3M VHB 9469, are adopted. For the first time, ultrahigh stretchability (ultimate stretch λ of up to 12.6 or strain ε of 1160%) and high transparency (up to 96.2%) are achieved simultaneously for an energy device because all engaged materials are stretchable and transparent. These soft skin-like generators are capable of outputting an open-circuit voltage of up to 145 V and an instantaneous areal power density of 35 mW m −2 . Meanwhile, the STENG-based electronic skin can sense pressure of as low as 1.3 kPa. The current study presents an energy harvester that is superstretchable, biocompatible, and transparent for the first time, allowing energy generation and touch sensing without interference of optical information’s delivering and could thus have potential in smart artificial skins, soft robots, functional displays, wearable electronics, etc.",
"discussion": "DISCUSSION We reported here a soft STENG composed of a sandwich structure of ionic hydrogels and elastomer films for energy conversion and tactile sensing. These material combinations and structural designs allow the following advantages: (1) High stretchablity (up to λ = 12.6 or strain ε = 1160%) and transparency (up to 96.2% average transmittance to full spectrum of visible light) are achieved, which are both much higher than those of previously reported stretchable and/or energy conversion devices using percolated conductive networks or ITO as the electrode materials ( 24 , 29 , 33 , 34 , 36 ). No performance degradation is observed at stretched states as well. (2) The mechanism of STENG is slightly different from that of previously reported TENGs using electrical conductors as the electrode. Ionic conductors (that is, ionic hydrogels) are used, which is demonstrated to be stable and will not be electrolyzed at the interface. (3) The unique capability of energy harvesting and tactile perception of the STENG may lead to the soft/stretchable self-powered sensors or self-charging power systems in the future ( 49 , 50 ). For example, self-powered soft robot or electronic skin might be possible in the future by the combination of soft energy-harvesting and energy storage devices and soft sensors and actuators. (4) Both elastomers and hydrogels are low cost, lightweight, and biocompatible. It is also possible to design the STENG into arbitrary, complicated shapes as long as a thin elastomer film wraps or seals the hydrogels. Meanwhile, the fabrication is facile for scale-up synthesis. Considering the biocompatibility, the STENG has potential for power devices attached on the skin or implanted inside the human body ( 51 ). Furthermore, both the elastomer and hydrogel have large categories of different materials with various unique properties. The combination of the hydrogel and elastomer has recently led to many multifunctional devices, as mentioned in the Introduction. The STENG with the hydrogel/elastomer combination reported here opens up opportunities for energy generation devices with new functional properties (stretchability, transparency, biocompatibility, etc.) and many potential applications ranging from soft robots, foldable touch screens, and artificial skins, to wearable electronics. In summary, our approach of soft/stretchable energy harvesting allows the energy conversion from human motions to electricity. The capability of scavenging biomechanical energies of the STENG was demonstrated when applying it as an artificial skin. Capacitors or batteries can be charged by the artificial skin to power wearable electronics. Finally, the sensitivity of the STENG to the input pressure was investigated, which enabled it as an electronic skin for pressure or tactile perception. The applicable temperature of the STENG is optimal at 0° to 60°C; otherwise, the freezing or boiling of the ionic solution in the hydrogel may cause the malfunction of the device. For future research, more multifunctional devices can be explored by developing STENGs with other functional hydrogels/elastomers; the interface bonding between the hydrogel and elastomer should be enhanced to further improve the mechanical performances of the STENG; output performances should be improved by maximizing the surface electrostatic charge density through surface treatments/modifications or materials optimization."
} | 1,993 |
36563153 | PMC9788778 | pmc | 99 | {
"abstract": "With increasing computing demands, serial processing in von Neumann architectures built with zeroth-order complexity digital circuits is saturating in computational capacity and power, entailing research into alternative paradigms. Brain-inspired systems built with memristors are attractive owing to their large parallelism, low energy consumption, and high error tolerance. However, most demonstrations have thus far only mimicked primitive lower-order biological complexities using devices with first-order dynamics. Memristors with higher-order complexities are predicted to solve problems that would otherwise require increasingly elaborate circuits, but no generic design rules exist. Here, we present second-order dynamics in halide perovskite memristive diodes (memdiodes) that enable Bienenstock-Cooper-Munro learning rules capturing both timing- and rate-based plasticity. A triplet spike timing–dependent plasticity scheme exploiting ion migration, back diffusion, and modulable Schottky barriers establishes general design rules for realizing higher-order memristors. This higher order enables complex binocular orientation selectivity in neural networks exploiting the intrinsic physics of the devices, without the need for complicated circuitry.",
"introduction": "INTRODUCTION Digital systems based on von Neumann architectures and built with zeroth-order complexity circuits have carried the workload of computing till date. However, with the exponential growth of computing needs, serial processing in such architectures is quickly saturating in terms of both computational capacity and power, entailing research into alternate paradigms ( 1 ). Because of their large parallelism, low energy consumption, and high error tolerance, brain-inspired neuromorphic systems are attracting considerable interest, especially for tasks such as classifying billions of images and powering speech recognition services ( 2 ). At the hardware level of the computing stack, the discovery of memristors has fueled approaches based on intrinsic device dynamics to replace complicated digital circuits, paving way for more efficient and simpler in-memory computing architectures ( 3 , 4 ). However, most demonstrations have thus far centered only around mimicking primitive lower-order biological complexities using devices with first-order dynamics ( 5 , 6 ). Although theoretical predictions of the benefits of higher-order devices exist, experimental demonstration of memristors with higher-order complexity is far and few ( 7 – 9 ). Memristors with higher-order complexities are predicted to solve problems that would otherwise require increasingly elaborate circuits ( 10 ), but no generic design rules exist. One of the intriguing features of biological neural networks (NNs) is their plasticity, which helps them to learn through experiential change in configuration. The human brain constantly evolves over time, creating new synaptic associations dependent on lifelong learning experiences and knowledge. Reproducing this ability of plasticity to perform in-memory computations in hardware is at the very core of neuromorphic engineering ( 11 ). Bearing functional resemblance to biological synapses, memristors are at the heart of such in-memory computing technology, and hence, biorealistic realization of synaptic plasticity in memristors is considered a crucial step toward realizing NNs with high accuracy and unsupervised learning capabilities. Need for complex learning rules In this context, selection of a plasticity model plays a vital role in designing neuromorphic systems. The first generation of neuromorphic systems typically implements some form of the pair or doublet spike timing–dependent plasticity (DSTDP) model—a local event-based weight update scheme that maps synaptic weight changes as a function of the timing between the pre- and postsynaptic spikes ( 12 – 15 ). This simple timing-based model is highly convenient because it allows for low-power operations within a specifically defined domain. However, the positive-feedback process this paired timing–based model adopts, in which strong synapses are further strengthened and weak synapses are further weakened, does not explain several key aspects of biological plasticity ( 16 ). It destabilizes the useful dynamic range of synaptic weights and fails to address time-variant problems such as online modeling of dynamic processes in visual surveillance. Hence, we need to look beyond simple DSTDP rules to model the next generation of NNs. Information processing in the brain involves a high connectivity—each neuron is estimated to be connected to up to 10 4 other neurons via synaptic junctions. Thus, synaptic plasticity can be intuitively considered to be a multifactor phenomenon. In biology, several factors are hypothesized to contribute to the learning process such as the timing between spikes ( 17 ), rate of pre- and postsynaptic firing ( 18 ), historical pattern of activity at the synapse ( 19 ), and global parameters like electrochemical environment, ionic concentration, and temperature ( 20 ). Despite the impressive progress already demonstrated with memristor-based computing architectures, many of the abovementioned factors are hitherto unaddressed, entailing innovative hardware approaches to emulate the plasticity and connectivity of biological NNs. This calls for a second generation of neuromorphic materials and devices, whose switching physics are capable of adhering to biorealistic plasticity models that capture both timing- and rate-based correlations, and encompass history-dependent activation and global regulatory controls. In this work, we show second-order dynamics in halide perovskite semiconductors, an archetypal ionic-electronic material. With a compositional space of >10 6 formulations that can be explored via solution-based simple processing, halide perovskites, as a material technology platform, offer a wide range of design options for memristive and neuromorphic devices. These materials are relevant for a wide range of neuromorphic architectures because they support a rich variety of switching physics, such as electrochemical metallization reactions with reactive electrodes, valence change mechanisms via halide ion migration, spin-dependent charge transport, and multiferroicity ( 21 – 23 ). Their mixed ionic-electronic conductivity enables comprehensive demonstration of Bienenstock-Cooper-Munro (BCM) learning rules, capturing both timing- and rate-based plasticity effects in a memdiode configuration. Ion migration and back diffusion result in modulable Schottky barriers at the halide perovskite–transport layer interfaces that are exploited by a triplet spike timing–dependent plasticity (TSTDP) scheme. This protocol establishes general design rules for realizing higher-order memristors with similar ionic-electronic materials. Going beyond the conventional Hebbian learning rule, the BCM rule is a biorealistic pattern-based plasticity law that captures the effect of both the timing between paired spikes (as in the case of common DSTDP) and the spike train rate, also known as spike rate–dependent plasticity (SRDP), and describes history-dependent synaptic modification ( Fig. 1A ). In contrast to previous investigations that use SRDP and DSTDP schemes ( 24 – 26 ), we exploit the TSTDP plasticity model ( 27 ) to map BCM rules in our memristive diodes, also known as memdiodes. Using a spike train stimulation protocol, we faithfully emulate the high connectivity of biological neurons and demonstrate advanced plasticity features, going beyond simple synaptic learning functions previously shown using single and paired spikes, e.g., excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), and DSTDP. The migration and back diffusion of ions in halide perovskites introduce an internal timing factor akin to Ca 2+ dynamics in biology that, together with a last spike–dominating rule and state-dependent forgetting effects, captures both temporal and rate-based correlations. We successfully demonstrate two main characteristics of the BCM rule, frequency dependence and sliding threshold ( 28 ), and establish a negative feedback process to regulate synaptic weight updates within a useful dynamic range, thus improving the stability of the NN. Inspired by the BCM rules that explain orientation selectivity in the mammalian visual cortex, we develop simulations of binocular orientation–selective NNs where the mechanism of plasticity involves temporal competition between input patterns instead of spatial competition between synapses as in Hebbian learning. We demonstrate all the features predicted by BCM learning with memristive devices. Fig. 1. Design of higher-order ionic-electronic memristors. ( A ) Conventional first-order electronic devices are capable of capturing only simple timing-based plasticity rules such as DSTDP (highlighted in the blue box on the left). On the other hand, higher-order memristors can follow a multifactor BCM learning rule (highlighted in the blue box on the right), where both timing and rate of firing are captured for a more robust learning. High firing rates induce LTP because they evoke strong postsynaptic depolarization and calcium signals, low to moderate firing rates induce LTD because they evoke moderate depolarization and calcium signals, and very low firing rates do not induce plasticity. Plasticity depends on the pre-post spike timing for different ranges of firing rate, illustrated by the colored boxes and arrows ( 63 ). Thus, the net plasticity reflects an interaction between the pre-post spike timing and firing rate. Here, second-order dynamics are observed in halide perovskite memdiodes with the structure ITO/SnO 2 + PCBA/MAPI/P3HT/MoO 3 /Ag. ( B ) Scanning electron microscopy cross-sectional image of the sample. The built-in potential due to band alignment and the Schottky barrier introduced at the MAPI-P3HT interface allows tunable temporal dynamics, a critical design feature of the second-order halide perovskite memdiode. ( C ) The intrinsic ion/ion vacancy migration in halide perovskites locally dopes the perovskite–transport layer interfaces, enabling finely modulable conductance/weight changes. The back diffusion of ions introduces an additional rate dependency, which we exploit to capture the BCM learning rules.",
"discussion": "DISCUSSION Performing computing based on the intrinsic device dynamics, where each device replaces complicated digital circuits in a functional sense, is a potential strategy to enable adaptive complex computing ( 53 , 54 ). Second-order memristors such as the ones presented in this work enable us to capture both timing- and rate-based learning rules using the devices’ intrinsic physics ( 7 ). In comparison to digital circuit implementations of higher-order synapses and first-order memristors ( 55 ), these devices portray advantages in area and circuit complexity. The need for second-order memristor comes from the complexity of implementing synaptic learning rules with first-order memristors. In the latter devices, the implementation of plasticity rules, such as spike time dependent plasticity (STDP), requires to encode the timing information in the shape of programming pulse. The memristor is used as a simple programmable memory in which the overlapping of spikes results in the right shape of amplitude and duration to encode the timing between presynaptic and postsynaptic neurons. These mechanisms are necessary because there are no other ways to encoding timing information in a first-order memristor. Instead, in second-order memristors, because of a second internal state variable, the activity of synapse controls the plasticity rather than the amplitude or pulse duration. The history of activity of the memristor is stored in the device itself and influences the future behavior ( 7 ). In our work, the halide perovskite memdiodes show second-order characteristics encoding timing and rate of spikes, because of their mixed ionic-electronic conduction. The possibility to encode this information in the activity of the synapses instead of a particular shape and/or duration of the pulses permits the use of these devices as second-order elements instead of a simple memory element for high-complexity neuromorphic computing. These devices act as a new building block to implement algorithms and systems without the need for complicated timing circuitry and unaffordable system complexity that first-order elements and digital implementations require. High complexity in this context refers to all neuromorphic computing systems in which a simple first-order memory element is not sufficient to implement the desired learning rule or algorithm, such as the BCM learning rule. Because of the specific physical properties of our devices, we successfully demonstrate the two main characteristics of the BCM rule, namely, the frequency dependence and the sliding threshold. The weight update trace reveals multiplicative correlations between presynaptic and postsynaptic activities and a nonmonotonic behavior in the depression region (EDE)—features that previous investigations ( 24 – 26 ) with SRDP and DSTDP schemes fail to address. In comparison to filamentary memristors, these devices have a larger dynamic range due to the rate-dependent negative-feedback process and the EDE region. The richer dynamics can be attributed to the back diffusion of ionic vacancies that introduce an additional modulatory mechanism along with the inbuilt electronic Schottky barrier (due to band alignment) and stimulation history. As mentioned before, P3HT is chosen specifically to introduce a significant Schottky barrier with MAPI at the hole extraction side ( Fig. 1B ), and thus, we focus on this part of the device. For analysis, we compare the initial states of a (i) low-experienced conductance state G 0 = 2.1 μS, (ii) medium-experienced conductance state G 0 = 3.9 μS, and (iii) high-experienced conductance state G 0 = 7.1 μS. The two extreme states are shown in Fig. 6 . Here, the Schottky barriers arising from the ionic vacancy accumulation are schematically represented for qualitative understanding. As shown, the high-experienced conductance state G 0 = 7.1 μS has a smaller Schottky barrier when compared to the low-experienced ( G 0 = 2.1 μS) and medium-experienced ( G 0 = 3.9 μS) conductance state due to accumulation of large number of negatively charged V Pb ′ and V MA ′ during the initialization process. Consequently, upon bias removal, more metastable ionic vacancies exist at the MAPI-P3HT interface for back diffusion in the case of G 0 = 7.1 μS, resulting in larger relative changes in the Schottky barrier, and enhanced forgetting and depression effects. Fig. 6. Mechanistic illustration of frequency dependence and sliding threshold of BCM learning rules with halide perovskite memdiodes. Schematic diagram of the memristive mechanism showing the accumulation of ion vacancies, dynamic change in energy band alignment of the MAPI-P3HT Schottky interface, and ion vacancy relaxation at different memristive states. In comparison to the recent demonstration with second-order oxide memristor ( 42 ) and two-dimensional (2D) heterostructure memtransistor ( 56 ), the mixed ionic-electronic conduction of halide perovskites offers a simpler processing route, device architecture, and higher yield approach to implement homeostatic regulatory mechanisms at the individual device level, thus establishing a universal design strategy. While other devices require preprogramming to a high conductance state to enable EDE, our device design allows EDE control via band structure and interface engineering and requires no preprogramming step, resulting in power saving. The above observations are expected to provide inspiration for similar ionic-electronic materials systems, such as lithium-intercalated battery-like synapses ( 57 ) and proton-doped organic electrochemical transistor–based synapses ( 58 , 59 ). These device properties enable the implementation of new learning mechanisms exploiting temporal competition between inputs in contrast to classical Hebbian learning where spatial competition between synapses is captured. Further studies, however, are necessary to investigate different materials with second-order dynamics that can add different physical time constants to cover a large spectrum of temporal processing capability. A large set of second-order devices is required to cover different applications with different specifications of operative frequencies and timing. Moreover, the presence of two state variables in these devices requires a deeper understanding of the underlying physics and suitable models to achieve proper optimization. On this point, it is important to note that the classical BCM model is parabolic, while experiments show a more complex functional shape. However, the perfect fitting of experimental curve is detrimental, increasing the complexity of the model without adding any critical features. The crucial point of this rule is to follow the dynamics of the variation of the weight rather than the absolute value of the variation of the weight ( 19 , 28 , 50 ). The latter would just result in a small change in the convergence speed, while the dynamics that we properly reproduce determines the properties and stability of the system. To conclude, simulations of binocular orientation–selective networks ( 60 ) mimicking visual cortex cells demonstrate an example of the relevance of halide perovskite memdiodes in the context of high-complexity computing: The timing/frequency processing properties of these devices enabled the development of a totally unsupervised system that implements a temporal-competition processing between input patterns, which can also be useful in many other general applications ( 25 , 61 ). This concept will enable a new generation of NNs with higher-order spatiotemporal functions that are useful to capture time-variance features in dynamic environments ( 62 ). Natural candidates that could benefit from that are video and audio processing systems, which, with these properties, become more similar to the biological learning mechanisms seen in mammalian brains. Furthermore, self-supervised learning for edge computing and efficient spatiotemporal recognition systems will also benefit from these devices, thus introducing an important new building block, significantly advancing beyond state-of-the-art demonstrations."
} | 4,630 |
39145295 | PMC11322636 | pmc | 100 | {
"abstract": "Spiking neural networks (SNNs) offer a promising energy-efficient alternative to artificial neural networks (ANNs), in virtue of their high biological plausibility, rich spatial-temporal dynamics, and event-driven computation. The direct training algorithms based on the surrogate gradient method provide sufficient flexibility to design novel SNN architectures and explore the spatial-temporal dynamics of SNNs. According to previous studies, the performance of models is highly dependent on their sizes. Recently, direct training deep SNNs have achieved great progress on both neuromorphic datasets and large-scale static datasets. Notably, transformer-based SNNs show comparable performance with their ANN counterparts. In this paper, we provide a new perspective to summarize the theories and methods for training deep SNNs with high performance in a systematic and comprehensive way, including theory fundamentals, spiking neuron models, advanced SNN models and residual architectures, software frameworks and neuromorphic hardware, applications, and future trends.",
"conclusion": "7 Future trends and conclusions This article provides an overview of the current developments in various theories and methods of deep SNNs, including relevant fundamentals, various spiking neuron models, advanced models, and architectures, booming software tools and hardware platforms, as well as applications in various fields. However, there are still many limitations and challenges. (1) Currently, only a few aspects of the intelligent brains have been applied to instruct the construction and training of SNNs, lacking enough biological plausibility. Therefore, to improve SNNs' capability, it is necessary to introduce more types of spiking neurons, rich connection structures, multiscale local-global-cooperative learning rules, system homeostasis, etc., into SNNs to more accurately mimic the cognitive and intelligent characteristics emerging in the brains. For example, it deserves more efforts to train SNNs with self-supervised or unsupervised learning (Zhou Z. et al., 2024 ), as children mainly receive unlabeled data during growth. Besides, the brain is actually a complex network, thus it is worthy of more effort to study graph SNNs, although some attempts already exist (Li et al., 2024 ; Yin et al., 2024 ). (2) Recent neuroscience studies have found that astrocytes can naturally realize Transformer operations (Kozachkov et al., 2023 ), which provides a new direction for the improvement of SNNs. In addition, astrocytes have the function of regulating neuronal firing activity and synaptic pruning (Lee et al., 2021 ; Liu et al., 2022 ), which provides ideas for the performance improvement and lightweight of SNNs in the future. (3) Information encoding methods and training algorithms for SNNs are mostly based on average firing rates, lacking the ability to represent temporal dynamics adequately. There should be more exploration of time-dependent information encoding strategies and corresponding training algorithms, to further enhance the spatiotemporal dynamic characteristics of SNNs and strengthen their temporal processing capability. (4) The training of SNNs mainly employs time-dependent methods, like BPTT, which greatly increases the training cost, compared to conventional DNNs. Thus, there is a need to develop brain-like SNNs that can be trained in parallel, and dedicated software and hardware that support their computation, reducing training time and power consumption. (5) As there are obstacles to conversion and interaction between different neuromorphic platforms, it is needed to establish a common standard to improve interoperability. Further, more brain-inspired principles or technologies should be incorporated into the design of neuromorphic systems, to enhance the computational performance of the chips, in terms of processing speed and energy efficiency. (6) Large-scale SNNs are mainly applied to classification tasks. Their potential in handling tasks that need to process continuous input streams, such as videos, languages, events from neuromorphic vision sensors, etc., has not been fully explored. Moreover, the introduction of various neuromorphic sensors and neuromorphic chips into autonomous robotics, cooperating with conventional sensors and processing chips, might be an efficient and effective way to achieve embodied intelligence. Further studies are needed to fully leverage the features and advantages of SNNs. In summary, studies and applications of SNNs are growing rapidly, but there is still great potential to improve the effectiveness and efficiency of SNNs. Efforts should be made in multiple directions, including model architectures, training algorithms, software frameworks, and hardware platforms, to promote the coordinated progress of models, software, and hardware.",
"introduction": "1 Introduction Regarded as the third generation of neural network (Maass, 1997 ), the brain-inspired spiking neural networks (SNNs) are potential competitors to traditional artificial neural networks (ANNs) in virtue of their high biological plausibility, and low power consumption when implemented on neuromorphic hardware (Roy et al., 2019 ). In particular, the utilization of binary spikes allows SNNs to adopt low-power accumulation (AC) instead of the traditional high-power multiply-accumulation (MAC), leading to significantly enhanced energy efficiency and making SNNs increasingly popular (Chen et al., 2023 ). There are two mainstream pathways to obtain deep SNNs: ANN-to-SNN conversion and direct training through the surrogate gradient method. Firstly, in ANN-to-SNN conversion (Cao et al., 2015 ; Hunsberger and Eliasmith, 2015 ; Rueckauer et al., 2017 ; Bu et al., 2022 ; Meng et al., 2022 ; Wang Y. et al., 2022 ), a pre-trained ANN is converted to an SNN by replacing the ReLU activation layers with spiking neurons and adding scaling operations like weight normalization and threshold balancing. This conversion process suffers from long converting time steps, which causes high computational consumption in practice. In addition, the converted SNNs obtained in this way are constrained by the original ANNs' architecture and are hard to adapt to dynamic signal (DVS, DAVIS, ATIS data) processing. Thus, the direct exploration of the virtues of SNNs is limited in ANN-to-SNN conversion. Secondly, in the field of direct training, SNNs are unfolded over simulation time steps and trained with backpropagation through time (Lee et al., 2016 ; Shrestha and Orchard, 2018 ). Due to the non-differentiability of spiking neurons, the surrogate gradient method is employed for backpropagation (Neftci et al., 2019 ; Lee et al., 2020b ; Fang et al., 2021a , b ; Zhou Z. et al., 2023 ). On one hand, this direct training method can handle temporal data and also achieve decent performance on large-scale static datasets, with only a few time steps. On the other hand, it can provide sufficient flexibility for designing novel architectures specifically for SNNs and exploring the properties of SNNs directly. Therefore, the direct training method has received more attention recently. Given the significant benefits and rapid advancement of directly trained deep SNNs, particularly the emergence of high-performance transformer-based SNNs, this review systematically and comprehensively summarizes the theories and methods for directly trained deep SNNs. Combining theory fundamentals, spiking neuron models, advanced SNN models and residual architectures, software frameworks and neuromorphic hardware, applications, and future trends, this article offers fresh perspectives into the field of SNNs. This review is structured as follows: Section 2 presents the evolution and recent advancements in spiking neuron models. Section 3 introduces the fundamental principles of spiking neural networks. Section 4 focuses on the most recent advanced SNN models and architectures, especially transformer-based SNNs. Section 5 concludes the software frameworks for training SNNs and the development of neuromorphic hardware. Section 6 summarizes the applications of deep SNNs. Finally, Section 7 points out future research trends and concludes this review.",
"discussion": "6.3 Discussion on SNN applications Deep SNNs have achieved great success in many fields in recent years, but there still exist some limits that need to be addressed. Firstly, although many studies demonstrated that deep SNNs achieved comparable accuracy to their ANN counterparts on many tasks, they still lag behind conventional ANN SOTA, especially for large datasets like ImageNet, which asks for more endeavors. Secondly, many studies claimed that the proposed SNNs consumed much less energy compared to ANN counterparts, through calculating the number of addition operations, without considering the cost of other operations like data movement. Therefore, it is meaningful to deploy well-performed SNNs on neuromorphic chips or corresponding simulators to fully exploit the event-driven nature and measure the actual energy cost. Thirdly, as for applications requiring high processing speed and low power consumption, like robotics, it is promising to adopt neuromorphic vision/audio sensors and neuromorphic processing chips due to their event-driven nature, besides network pruning and weight quantization. Meanwhile, to fully exploit the advantages of events, it deserves more efforts to explore how to directly process neuromorphic sensing events using SNNs, without converting events into frames as current studies usually do. Fourthly, as for transformer-based SNNs used in language or video processing, how to choose the input clip for one simulation step, to reconcile the temporal resolution of the input sequence and the simulation step of SNNs, is worth studying. Last but not least, as SNNs have an additional temporal dimension, how to achieve the speed-accuracy trade-off as humans is a problem worth of study. In other words, how to assign a suitable simulation duration or how to decide when to make a choice, are important questions to realize the balance between computation cost and prediction accuracy."
} | 2,530 |
35541967 | PMC9082732 | pmc | 101 | {
"abstract": "In this paper, we demonstrate a newly designed hybridized triboelectric nanogenerator (TENG) fabric incorporating multiple working modes, which can effectively harvest ambient mechanical energy for conversion into electric power by working in a hybridization of a contact–separation mode, a sliding mode and a freestanding triboelectric layer mode. The power generation of each mode of the TENG fabric was systematically investigated and compared along different directions, under different frequencies and at different locations. Owing to the advanced structural design, the as-fabricated TENG fabric could be switched between multiple working modes according to its real working situation. High output voltage and current of about 140 V and 0.6 μA, respectively, were obtained from a larger size of TENG fabric, which could be used to light up 120 LEDs in series. Compared to the previously reported TENGs, such a hybridized TENG fabric based on hybridized modes has much better adaptability for harvesting energy (such as human walking, running, and other human motion) in different directions. This work presents the promising potential of hybridized TENG fabric for power generation and self-powered wearable devices.",
"conclusion": "4. Conclusion In this paper, the concept of a hybridized energy harvester incorporating different modes (including a contact mode, a sliding mode and a freestanding triboelectric mode) was proposed, so that each mode could be used effectively and in a complementary manner in a TENG fabric. The working principle of each mode of the TENG fabric was analyzed in order to fully understand the power generation process. The power generation of each mode was systematically investigated and compared along different directions, under different frequencies, and in different locations. A typical TENG fabric with a larger size can deliver a maximum output voltage of 140 V, a maximum output current of 0.6 μA, which can be used to light up 120 LEDs. Compared to the previously reported TENGs, such a hybridized TENG fabric, based on hybridized modes, has much better adaptability for harvesting energy (such as human walking, running, and other human motion) in different directions.",
"introduction": "1. Introduction Wearable electronic devices have received great attention due to their promising applications in a vast range of fields such as health monitoring, wearable smart phones, artificial skin sensors and motion tracking. 1–9 Typically, any of these electronic devices need an external power source to operate. However, traditional rigid batteries as power sources remain a bottleneck hindering the practical and sustainable uses of wearable electronics, due to their heavy weight, bulky volume, and limited capacity and lifetime. In order to improve the flexibility of the battery, the development of new technologies that can harvest energy from the environment as sustainable self-powered sources has become a newly emerging field. 10–12 Triboelectric nanogenerators (TENGs) have been invented as an effective way to harvest energy from our living environment (such as human motion, flowing water and airflow) based on the coupling of contact-electrification and electrostatic induction. 13–16 Since 2012, TENGs have been successfully demonstrated as promising energy-harvesting technology for self-powered devices, such as motion tracking systems, velocity sensors, biosensors and so on, 9,17–21 due to their cost-effectiveness, high efficiency, low-cost, environmental friendliness, and universal availability. 22–25 Very recently, attempts have been made to fabricate fabric-based TENGs for promising applications in self-powered wearable electronics. 26–28 For example, Zhou et al. 29 reported a woven structured TENG based on a freestanding triboelectric-layer mode for scavenging energy from human motion, which showed a stable output voltage and current of about 10–30 V and 0.5–1 μA, respectively. Pu et al. 28 have demonstrated a wearable fabric TENG by using a grating-structure which could convert low-frequency human motion energy into high-frequency electrical outputs. However, these structural designs of TENGs can only work in 1D or 2D directions, and are unable to harvest mechanical energy from arbitrary directions, which largely limits their versatility and applicability in wearable electronics, because of the variability of human movement. Further research is still urgently required to design fabric-based TENGs that have 3D motion energy scavenging abilities for maximizing energy harvesting. Meanwhile, integration of fabric-based TENGs that incorporate different modes (including a contact mode, a sliding mode and a freestanding triboelectric mode) has seldom been found in literature, though several works have been reported of TENGs integrated with other energy harvesters (such as solar cells). It is desirable to incorporate two or more modes that can be integrated together, giving the device the ability to collect either sliding movement or separation movement according to the actual conditions, which has much better adaptability in wearable electronics. Herein, we demonstrate a hybridized TENG fabric incorporating different working modes (including a contact–separation mode, a sliding mode and a freestanding triboelectric-layer mode), each mode of which can be used effectively and in a complementary manner. The working principles were analyzed in order to fully understand the power generating process of each TENG fabric. The output performance of each mode was systematically investigated under different directions, frequencies, and locations. Additionally, the TENG fabric was also employed in applications for lighting up commercial LED bulbs.",
"discussion": "3. Result and discussion 3.1 Structure design of the hybridized TENG fabric The hybridized TENG fabric was structurally composed of two parts: one part was the woven-structured fabric (W-fabric); and the other part was the nylon fabric with a grid-patterned back electrode (G-fabric), as shown in Fig. 1c . In general, the hybridized TENG fabric had two basic structures: one was a free-standing TENG (named FS-TENG), in which the structured grid-patterned electrodes A and B were chosen as the terminals, and the top W-fabric acted as the freestanding triboelectric layer. The other structure was a contact–separation type TENG (named CS-TENG) where electrodes A and B were connected together and acted as one of the terminals, and electrode C acted as another terminal. The photograph and corresponding scanning electron microscopy (SEM) image of the grid-patterned Au electrode on nylon are shown in Fig. 1d and e , demonstrating its flexibility and electrode continuity. The proposed hybridized TENG could be easily integrated to human clothes, because of its flexibility, wearability, and air permeability. 3.2 Working mechanisms and output performances of the TENGs Because of the complexity of human motion, both FS-TENG and CS-TENG have at least two modes of movement, one being pressing and the other sliding. In order to harvest mechanical energy in different directions, our TENGs were designed to work in both a contact–separation mode and a sliding mode simultaneously. To fully understand the power generating process of the TENG fabrics, each mode of the TENG fabrics was carefully analyzed. \n Fig. 2a shows the scheme of the FS-TENG fabric in a contact–separation mode motion. For the sliding mode, as the top W-fabric contacts with the G-fabric, because PTFE and nylon fabrics have different abilities in attracting electrons, there are positive triboelectric charges on the nylon surface and negative ones on the PTFE surface ( Fig. 2a stage I). When the W-fabric starts to slide toward the right-hand side, the induced negative charges on the top triboelectric layer (PTFE) result in an instantaneous electron flow from electrode B to electrode A, finally reaching equilibrium when the W-fabric stops on the right side of the nylon ( Fig. 2a stage III). Once it moves back from the right side to the left side, the electrons flow back from electrode A to electrode B, until another equilibrium is achieved 30,31 ( Fig. 2a stage IV). For the contact–separation mode, when they are separated, an electric potential difference is produced, driving electrons to flow from electrode B to electrode A through an external circuit in order to balance the generated triboelectric potential ( Fig. 2b stage II). As the separation between the W-fabric and the G-fabric is maximized, the flow of electrons is stopped because an electrostatic equilibrium is reached ( Fig. 2b stage II); once the W-fabric is driven to contact with the nylon layer again, electrons flow from electrode A back to electrode B, until another equilibrium is achieved ( Fig. 2b stage II). Fig. 2 The working mechanisms of the FS-TENG fabrics: (a) the sliding mode; (b) the contact–separation mode; the output performance of the FS-TENG fabrics working in the sliding mode (c) and the contact–separation mode (d). Subsequently, the output performances of FS-TENG in both working modes were also evaluated. For the sliding mode, the open-circuit voltage ( V OC ), and the short-circuit current ( I sc ) are shown in Fig. 2c and d , with peak values of around 15 V and 1 μA, respectively. Meanwhile, the output performance of FS-TENG in the contact–separation mode was also characterized in the sliding mode. As exhibited in Fig. 2 , the values of the open-circuit voltage ( V OC ), and the short-circuit current ( I sc ) were 2 V and 0.3 μA, and were smaller than those in the sliding mode. We can see that by having a freestanding triboelectric layer (W-fabric), FS-TENG could move freely without any constraint. Moreover, FS-TENG in the sliding mode also had the capability of converting low-frequency motion into high-frequency signals, i.e. each sliding motion generated multiple peaks rather than one single peak, as was observed for the previous reported TENG, which will largely expand the applications of TENGs for versatile mechanical energy harvesting, especially for wearable devices. Similarly, the energy-harvesting of the CS-TENG fabric is also schematically illustrated in Fig. 3 . Because of the large differences in the ability to attract electrons, the triboelectrification left the nylon surface with positive charges and the PTFE with net negative charges with equal density. For the sliding mode, once the top W-fabric slides outward ( Fig. 3a stage II), the induced potential difference drives a current to flow from electrode A(B) to electrode C through an external circuit in order to balance the generated triboelectric potential. Subsequently, when the W-fabric was slid backward ( Fig. 3a stage IV), the separated charges began to get in contact again, and the redundant charge transferred electrons flowed back through an external circuit, resulting in a current flow from electrode C to electrode A(B). 13,32 For the contact–separation mode of CS-TENG, when the W-fabric is separated from the nylon surface, the induced potential drives a current to flow from electrode A(B) to electrode C until the potential difference is fully offset by the transferred charges ( Fig. 3b stage II); when the W-fabric is reverted to contact the nylon surface again, electrons flow back from electrode C to electrode A(B), until another equilibrium is achieved ( Fig. 3b stage II). 33 Fig. 3 The working mechanism of the CS-TENG fabrics: (a) the sliding mode; (b) the contact–separation mode; the output performances of the CS-TENG fabrics working in the sliding mode (c) and the contact–separation mode (d). The output performances of CS-TENG both in the sliding and contact–separation working modes were measured as shown in Fig. 3c and d . For the sliding mode, the V OC was measured to be about 120 V OC and the I sc was measured to be about 0.6 μA. It is noteworthy that in this mode, electric signals are only generated when the W-fabric slides out of the T-fabric; i.e. there is little output when the W-fabric slides only inside the G-fabric (as shown in Fig. S1 † ), which is because there is no polarization generated in this case. 33 For the contact–separation mode of CS-TENG, the measured V OC and I sc were about 120 V and 0.8 μA. It can seen be that both modes of CS-TENG have higher output performances than those of FS-TENG. 3.3 Influencing factors on the TENG fabric Since the mechanical energy from the environment is always irregular and varies in frequency, it is necessary to study the dependence of the TENGs’ outputs on motion-direction. For the sliding mode, both output performances of FS-TENG and CS-TENG working along three typical directions (0°, 45° and 90°) were investigated for contrast. For FS-TENG, when the W-fabric slid along 0° and 90°, the output voltage and current remained unchanged (15 V and 1 μA) (as shown in Fig. 4 ); while when the sliding angle was 45°, the output voltage and current were decreased slightly from 15 V and 1 μA to 8 V and 0.5 μA, respectively. Fig. 4d–e show the corresponding output signals of CS-TENG sliding along 0°, 45° and 90°. From Fig. 4 , we can see that CS-TENG can harvest sliding energy from all directions with the output performance remaining almost unchanged. In this case, the grid-patterned Au electrodes were perceived as a whole because A and B were connected together. Therefore, the sliding path does not have a big impact on the output performance. For the contact–separation mode, the contrasting output performances of FS-TENG and CS-TENG under different contact–separation positions were measured as shown in Fig. 5 . We can see that when the W-fabric and the G-fabric changed contact–separation points from the matched style I (where the W-fabric matched exactly with the grid-patterned electrodes below) to the unmatched style II (where the W-fabric was unmatched with the grid-patterned electrodes below), both the output voltage and the current of FS-TENG decreased greatly from 4 V and 0.5 μA to 0.5 V and 0.1 μA, respectively. While, for CS-TENG, the output performance was almost constant when changing the contact–separation points, as shown in Fig. 5e and f . Fig. 4 Output performance of both the FS-TENG and the CS-TENG fabrics sliding along different motion-directions. (a) Schematic illustrations of three typical sliding paths (sliding angles: 0°, 45° and 90°). (b and c) Output voltage and current of the FS-TENG fabric sliding at different angles. (d and e) Output voltage and current of the CS-TENG fabric sliding at different angles. Fig. 5 Output performances of both FS-TENG and CS-TENG in the contact–separation mode with different contact–separation points. (a and b)Schematic illustrations of two typical contact–separation points (style I and II). (c and d) Output voltage and current of the FS-TENG fabric working in different contact–separation points from style I and style II. (e and f) Output voltage and current of CS-TENG working in different contact–separation points from style I and style II. From the results above, we can easily see that CS-TENG shows more excellent advantages, not only because of its higher output performance, but also because it is basically not affected by the influence of movement direction, including the sliding path or the contact–separation point, which might mean it has much better adaptability for harvesting energy. However, FS-TENG also has its advantages in unique working situations. For example, FS-TENG could have the ability of harvesting small amplitude sliding mechanical movements; while CS-TENG can only collect large amplitude motions, because no signal is generated when the W-fabric slides only inside the G-fabric. Therefore, the as-fabricated hybridized TENG fabric could be switched between both structures, according to the real working situation, by changing the connection. Besides the motion-direction, another factor that could influence the output of the TENGs is the frequency. As a supplement, the relationship between the electrical outputs of the TENGs and the frequency was systematically investigated as shown in Fig. S2. † As for the V OC , the peak value remained almost unchanged under different frequencies (f1 < f2 < f3). According to the previous reports, the V OC is independent to the sliding velocity, and is only determined by the displacement. 33 As for the I sc , the peak heights for the four working modes present an obvious increasing trend with increased frequency (f1 < f2 < f3), because higher frequency will not only result in more transferred charges as discussed before, but also more importantly contributes to a higher charge transfer rate. High output performances can be achieved by the TENG fabric with a large segment size. Here, a larger laser-scribed mask, with a single square window with size of about 20 × 20 mm 2 , was also utilized to obtain a larger piece of TENG fabric with dimensions of 5 × 5 mm 2 . The performances of both the FS-TENG and the CS-TENG fabrics were tested under both the sliding mode for FS-TENG and the contact–separation mode for CS-TENG. Fig. 6a and b display typical output signals of FS-TENG in the sliding mode, which delivers a maximum output voltage of 140 V, and a maximum output current of 0.6 μA, respectively. Through the use of a full-wave bridge rectifier, the energy harvesting capability of the FS-TENG fabric was also evaluated under the sliding mode for charging three commercial capacitors (2.2 μF, 3.3 μF and 10 μF, respectively) in a direct charge cycle. With a similar frequency of ∼0.5 Hz, three commercial capacitors can be charged from 0 to 850 mV (black line), 600 mV and 400 mV (red line), respectively, within 25 s as shown in Fig. 6c . Furthermore, a total of 120 commercial LED bulbs were assembled in series on a piece of electrical circuit board ( Fig. 6d ), and were connected to the TENG fabric. When manually triggered, both the FS-TENG and the CS-TENG fabrics could directly and simultaneously light up all of these 120 LED bulbs. (ESI Videos 1 and 2 † ). This demonstration suggested that the TENG fabrics have potential as sustainable power sources for driving wearable smart electronics. Fig. 6 (a and b) Typical output performances of the FS-TENG fabric with a larger size working in the sliding mode. (c) Charging curves of three capacitors (2.2 μF, 3.3 μF and 10 μF) charged by the FS-TENG fabric under the sliding mode. The inset is the equivalent electrical circuit of the self-charging power system. (d) A snapshot of the 120 commercial LED bulbs in serial-connection directly driven by the FS-TENG fabric under the sliding mode. (An equivalent snapshot of the CS-TENG fabric is shown in the ESI † )."
} | 4,684 |
35479675 | PMC9032853 | pmc | 102 | {
"abstract": "The next-generation multifunctional soft electronic devices require the development of energy devices possessing comparable functions. In this work, an ultra-stretchable and healable hydrogel-based triboelectric nanogenerator (TENG) is prepared for mechanical energy harvesting and self-powered sensing. An ionic conductive hydrogel was developed with graphene oxide and Laponite. as the physical cross-linking points, exhibiting high stretchability (∼1356%) and healable capability. When using the hydrogel as the electrode, the TENG can operate normally at 900% tensile strain, while the electrical output of the TENG can fully recover to the initial value after healing the damage. This hydrogel-based TENG is demonstrated to power wearable electronics, and is used as a self-powered sensor for human motion monitoring and pressure sensing. Our work shows opportunities for multifunctional power sources and potential applications in wearable electronics.",
"conclusion": "4. Conclusion In summary, we fabricated an ultra-stretchable and healable TENG, where VHB elastomer and ionic hydrogel were used as the electrification layer and electrode, respectively. The hydrogel exhibited high stretchability (∼1356%) and high healable capability (∼99% healing efficiency). The electricity generation capability of the TENG was demonstrated, which could be used as a sustainable self-powered power source to power commercial LED lights and electronic watches. The hydrogel-based TENG could also operate normally at 900% tensile strain and its electrical performances could be fully recovered after damages. In addition, this soft TNEG was demonstrated to be self-powered sensors for pressure sensing and human motion detection.",
"introduction": "1. Introduction Wearable electronic devices have attracted intense attention in the last decade with multiple functions being realized, such as transparency, stretchability, biocompatibility, healable capability, and biodegradability. 1–6 A variety of electronic devices have been designed and developed, but most of them are powered by traditional power sources, such as heavy batteries and capacitors which have limited life span and are not environmentally friendly. It is still a bottleneck challenge to provide sustainable power to these electronic devices without sacrificing their advantages in stretchability or other functions. Therefore, it has become an alternative technology approach to develop energy harvesters with compatible multifunctions, so as to scavenge the distributed renewable energies in the working environment of these devices for the power sources. Triboelectric nanogenerators (TENGs) have intensively been studied in recent for energy harvesting and self-powered sensors. 7–11 The TENG generates electricity from various kinetic energies based on the effects of triboelectric electrification and electrostatic induction. It has the advantages of low cost, high flexibility, high efficiency, and environmental friendliness. Efforts have been also made to develop multifunctional TENGs, such as stretchable, transparent, healable, or biodegradable TENGs. 12–21 The key to achieve stretchable TENGs is the development of stretchable electrodes. Stretchable TENGs have been reported using electrodes of conductive filler-percolated composites, 22–24 liquid metals, 25–28 ionic solutions, 29 or ionic gels. 30,31 The ion-conducting gel electrodes, including hydrogels, ionogels and organogels, have the advantages of low modulus and high stretchability. Even though their conductivity is generally much lower than that of electronic conductors, it overcomes the low stretchability of electronic conductors, since the ion conduction is based on the ion transfers along the trapped solution inside the molecular networks. 32 Furthermore, it is easier for gel electrodes to achieve other multifunctions through the design of proper molecular chain networks, such as the healable capability, biodegradability or transparency. In particular, the hydrogels have the highest conductivity and are more environmentally friendly or biocompatible than the other two counterparts. Therefore, efforts are still being made to optimize the hydrogel electrode-based multifunctional TENGs. Here, we present an ultra-stretchable and healable hydrogel-based TENG for mechanical energy harvesting. Graphene oxide (GO) and Laponite were used as collaborative physical crosslinking points in the hydrogel, leading to its high stretchability and high healable efficiency. The hydrogel with 1 wt% GO achieved a tensile strength of 106 kPa and an ultimate tensile strain of 1356%, and the ionic conductivity is 38.8 mS cm −1 at 25 °C. All the aforementioned characteristics make this hydrogel useful as an electrode for stretchable and healable TENG. The obtained TENG can not only maintain high stretchability, but can also fully recover its electricity generation capability after healing the damages autonomously. Moreover, these TENG could also function as sensors for detecting human motions and contacting pressures.",
"discussion": "3. Results and discussions 3.1 Fabrication and characterization of the hydrogel The hydrogel polymer chains are prepared by co-polymerization of three monomers, i.e. 2-acrylamido-2-methylpropane sulfonic acid (AMPS), acrylic acid (AA) and N -[3-(dimethylamino)propyl]methacrylamide (DMAPMA), as shown in Fig. 1a . The GO and Laponite were added as physical crosslinking points ( Fig. 1b ). Laponite and GO, which both have similar laminated structures, can enhance the mechanical properties of hydrogels, achieving high strength and high stretchability. 35,36 Meantime, they can endow the systems healable property due to the formation of abundant hydrogen bondings. The Laponite–GO-crosslinked poly(AMPS- co -AA- co -DMAPMA) hydrogels with 2 M LiCl was noted as LPG–LiCl hydrogel thereafter. The scanning electron microscopy (SEM) images of the freeze-dried hydrogel samples showed that the polymer matrix had a typical three-dimensional porous morphology, which is beneficial for the uptake of aqueous solution ( Fig. 1c ). From the FTIR spectra ( Fig. 1d ), the reflection peak of stretching vibrations of Si–O–Si in Laponite at 657 cm −1 shifted to a lower wavenumber of 640 cm −1 in the hydrogel, indicating that crosslinking by hydrogen bonds was formed between Laponite and polymer chains. As for the GO, reflection peak at 3430 cm −1 was attributed to the O–H stretching; reflection peaks at 1730 cm −1 and 1379 cm −1 were attributed to the C \n \n\n<svg xmlns=\"http://www.w3.org/2000/svg\" version=\"1.0\" width=\"13.200000pt\" height=\"16.000000pt\" viewBox=\"0 0 13.200000 16.000000\" preserveAspectRatio=\"xMidYMid meet\"><metadata>\nCreated by potrace 1.16, written by Peter Selinger 2001-2019\n</metadata><g transform=\"translate(1.000000,15.000000) scale(0.017500,-0.017500)\" fill=\"currentColor\" stroke=\"none\"><path d=\"M0 440 l0 -40 320 0 320 0 0 40 0 40 -320 0 -320 0 0 -40z M0 280 l0 -40 320 0 320 0 0 40 0 40 -320 0 -320 0 0 -40z\"/></g></svg>\n\n O and C–OH in the COOH groups, respectively; reflection peak at 1046 cm −1 and C–O–C stretching vibrations. However, the C O peak of GO in the hydrogel shifted to a lower wavenumber of 1725 cm −1 , indicating the existence of hydrogen bonds between GO and polymer chains. Fig. 1 Preparation of the nanocomposite LPG–LiCl hydrogels. (a) Schematic of copolymerization of poly(AMPS- co -AA- co -DMAPMA) with potassium persulfate (KPS) and N , N , N ′, N ′-tetramethylethylenediamine (TEMED) as initiator and catalyst, respectively. (b) Schematic network structure of LPG–LiCl hydrogels with Laponite and GO as cross-linking points. (c) SEM images of a freeze-dried LPG–LiCl hydrogel. (d) FTIR spectra of Laponite, GO, and LPG–LiCl hydrogels. \n Fig. 2a showed the variation of the tensile strength and strain at break of the LPG hydrogel with different GO contents. The tensile strength and strain at break for the hydrogel without GO is 43 kPa and 2948%, respectively. When the GO content was 1 wt%, the tensile strength was 106 kPa, more than 2 times higher than that without GO, but the strain decrease to 1356%. When the content of GO was further increased, the tensile strength of the hydrogel would decrease. Therefore, adding appropriate amount of GO as crosslinker can strengthen the hydrogel, but the strengthening effect was weakened when the GO content was high. Meanwhile, the conductivity of the hydrogel with different GO contents was tested. As shown in Fig. 2b , the content of GO has no significant effect on the electrical conductivity (38.8–41.6 mS cm −1 ) of the hydrogel. In the following study, the hydrogel with a 1 wt% GO content was used, unless otherwise stated. Fig. 2 Mechanical performance and healable property of the nanocomposite hydrogel. Tensile stress–strain curves (a) and ionic conductivity (b) of hydrogels with different contents of GO. Cyclic tensile behaviors of the hydrogel under different strains (c) and under a fixed strain without resting times (d). (e) Healable property of hydrogels healed at 80 °C for different time. (f) Time evolution of the healable process for the hydrogel by the real-time resistance measurements. (g) Schematic diagram of healable mechanism of hydrogels. (h) A circuit comprising hydrogel in series with a blue LED indicator: (i) original, (ii) completely bifurcated, (iii) healed. Inset: optical microscope image for the hydrogel at each state. Scale bar: 1 cm (i, ii, iii) and 200 μm (inset). In order to evaluate the rebounding resilience properties of hydrogels, a series of tensile loading and unloading cycles were applied to the hydrogels. As shown in Fig. 2c , the closed area of the stress–strain curve formed during the loading-unloading process increased with increasing the tensile strain, indicating that the dissipated energy also increased. This was due to the fact that the increase of strain leaded to the slip of more polymer chain segments and the dissipation of more energy. 37,38 Subsequently, we investigated the anti-fatigue properties of the hydrogel. As shown in Fig. 2d , in 8 consecutive cyclic tensile tests, the cycle curve of the hydrogel almost coincided with the initial state of the hydrogel, indicating that the hydrogel had good resilience and fatigue resistance. The thermal properties of hydrogels were studied by differential scanning calorimetry (DSC) experiments. The results show that dried hydrogels have a relatively high glass transition temperature ( T g ) of 50.6–53.7 °C (Fig. S1 † ). The healable properties of hydrogels were evaluated through the mechanical properties after cutting the hydrogel completely into two pieces and healing for different times. The healing efficiency calculated based on the strain to fracture is 50%, 60% and 99% after 12 h, 24 h and 36 h healing at 80 °C, respectively. Nevertheless, the healing efficiency is only 48% after 36 h healing at 25 °C (Fig. S2 † ). Besides, the healable properties of hydrogels were evaluated through the electrical properties. The time evolution of the resistance change is shown in Fig. 2f . When the hydrogel was completely cut off, the resistance increased sharply from the steady state to the open circuit state. Once the two broken samples were spliced together, the resistance value would quickly return to the initial value. \n Fig. 2g showed a schematic diagram of the healable mechanism of the hydrogel. There were lots of oxygen-containing groups (–COOH, –OH, etc .) at the surface of GO. 39–41 There are also abundant –OH groups on the surface of Laponite plates; the surface of Laponite is negatively charged and the edge of Laponite is positively charged. 42 Therefore, both Laponite and GO can form hydrogen bondings with the –CONH 2 groups of the polymer chains and serve as the physical crosslinking points. These dynamic hydrogen bondings can then lead to the healable capability of the system, since they can be reversibly re-formed after the damages. Furthermore, the reversible electrostatic interaction between the GO and the Laponite can also contribute to healable property of the hydrogel. For demonstration, the hydrogel was connected to the power source through the hydrogel and a commercial light-emitting diode (LED) bulb ( Fig. 2h(i) ). Hydrogel was cut into two pieces and attached back ( Fig. 2h(ii) and 2h(iii) ). The 67 μm wide cleft had almost disappeared, showing good repair ability (the inset photos in Fig. 2h ). Moreover, there was no difference before and after the broken/healed process, thus verifying the healable of the electrical performance of the hydrogel. As for polymer ionic conductors, the ion conduction pathways are fully recovered as long as the molecular chains are healed; while for percolated polymer composites, the electron conduction pathways may not be thoroughly recovered after the healing process. 3.2 The output of TENG The hydrogels and VHB elastomer are used as the electrodes and dielectric layer for TENGs. The TENG works in the single-electrode mode with a sandwich structure, in which hydrogel is used as an electrode, sealed between two layers of 3 M VHB membranes and connected to an external load with a copper wire. A commercial polytetrafluoroethylene (PTFE) film is used as the other electrification layer. Fig. 3a shows the working mechanism of TENG. When the PTFE surface is in contact with the VHB layer of the TENG, static charges of the same quantity and opposite polarity are generated on the PTFE and VHB surfaces, respectively ( Fig. 3a–i ). There is no potential difference between the two surfaces and no current is generated in the circuit. When the dielectric material PTFE is separated from the triboelectric layer VHB of the TENG, the un-screened positive static charges in VHB will induce the negative ions flowing to the gel/VHB interfaces. Thereby an electric double layer will be formed at the metal wire/gel interfaces, leading to the electrons flowing from the ground to the metal/gel interfaces through the external loading ( Fig. 3a-ii ). The electron stop flowing until all the static charges in the VHB are screened and the charge quantity of the electric double layer reaches maximum ( Fig. 3b-iii ). When PTFE is approaching by external force back to contact the TENG, the process is reversed, so that electrons will flow from the copper wire to the external circuit ( Fig. 3b-iv ). Alternating current will be generated when PTFE and TENG are engaged in continuous contact/separation movements. Fig. 3 The working principles and the output of the TENG with a single-electrode mode. (a) Scheme of the working mechanism of the TENG. The electric output performances of the TENG: (b) V oc , (c) I sc , and (d) Variation of the output current density and power density with the external loading resistance. (e) V oc and (f) I sc of the TENG under different driven frequency from 1 to 5 Hz. The reciprocal contact-separation motion between a PTFE and a TENG (area, 2 × 2 cm 2 ) was realized by a linear motor. The motion frequency was firstly fixed to 2 Hz to measure the electrical output. As shown in Fig. 3b, c , and Fig. S3, † the open-circuit voltage ( V oc ), short-circuit current ( I sc ), and short-circuit charge ( Q sc ) of TENGs are 75 V, 0.6 μA, and 25 nC, respectively. In addition, the output power density of the TENG was measured by changing the external resistance. When the external resistance is about 800 MOhm, the output power density of TENG reaches the maximum value of 260 mW m −2 . The TENG has obvious advantages over traditional electromagnetic generators in generating high-voltage electricity from irregular and low-frequency mechanical energies (<5 Hz). 43 By increasing the frequency, the Q sc (Fig. S4 † ) and V oc ( Fig. 3e ) showed very slight increase, as the generated static charge quantities are not sensitive to the motion frequency. Nevertheless, as the frequency increased from 1 Hz to 5 Hz, the I sc increased from 0.3 μA to 1.5 μA ( Fig. 3f ). This is due to the fact that the time for charge transfer in each cycle is shortened at high frequency though the transferred charge quantity is about the same. The actual application of TENG requires external mechanical stimulation or internal mechanical friction, which will inevitably damage TENG and directly affect the performance of TENG. 12,44 Therefore, it is necessary to develop TENG with healable property to recover its performances after damages. This hydrogel electrode has the healable capabilities, and the adhesive VHB film is also recoverable (Fig. S5 † ). It can be confirmed that the tensile strain of the VHB film can be recovered to be 55% after attaching the cut two pieces back into together in 1 min. We compared the output of TENG before and after healing process ( Fig. 4 and Fig. S6 † ). There were no significantly changes in all three electrical output parameters ( I sc , V oc and Q sc ), indicating that the hydrogel-based TENG could fully recover its output performances after damage. Fig. 4 The comparison of output for TENG before and after the healing process. (a) Photographs of the TENG before and healed after being cut in half. The comparison of the TENG output performances before and after healing. (b) V oc and (c) I sc . (d) V oc and (e) I sc of the TENG stretched to different strains. (f) The durability of the hydrogel-based TENG. The stretchability of the hydrogel-based TENG was evaluated at the stretched states ( Fig. 4d–e and Fig. S7 † ). The output performances enhanced with the increase of strains. Compared with the initial state, the V oc , I sc and Q sc reached ∼164 V, ∼1.6 μA and ∼54 nC when increasing the strain to 500%. In the measurement process, the shape and size of PTFE film were larger than that of TENG at all stretched states. The increase in outputs is mainly caused by the change of contact surface area and dielectric layer thickness. The surface area of the TENG would increase under deformation conditions compared with the initial state, resulting in more static charges at a larger contact area. Meantime, the thickness of the VHB would decrease, which was also beneficial for the higher output. Similar results have also been reported. 30,31 Therefore, this TENG demonstrated exceptional ultra-stretchable performances. In addition, the durability of the device was also tested. As shown in Fig. 4f , TENG still showed stable performance after more than 3000 cycles under 2 Hz, indicating that the device had excellent long-term reliability. 3.3 TENG as power source for commercial electronics TENG can usually be used as a sustainable nanoscale power source, harvesting tiny mechanical energy of human motion to power small electronics such as LED lights. The test was carried out by a TENG with an area of 2 × 2 cm 2 , and the human body can be regarded as a reference electrode or ground. As shown in Fig. 5a–c , 14 LED lights were successfully lit by tapping the TENG with a hand. Moreover, the TENG can be used to charge capacitors. Since TENG generates an AC signal, it cannot be used to power an electronic watch directly. A rectifier bridge is usually needed to manage the entire circuit. As shown in Fig. 5d and e , by connecting a rectifier, the generated electric energy can charge a 2.2 μF capacitor to 2 V within 40 s, and then power an electronic watch. The corresponding charging circuit diagram was shown in Fig. 5e . Fig. 5 Demonstration of the TENG to power commercial electronics. (a–c) Photograph of green LEDs connected in series driven by tapping the TENG with a hand. (d) An electric watch was powered by TENG. (e) Voltage profile of a 2.2 μF capacitor being charged by the TENG and powering the electronic watch. (f) The equivalent circuit of a self-charging system for the hydrogel based-TENG to power electronics. 3.4 Human motion detecting and pressure sensing by the TENG The TENG could be utilized in various applications such as human motion monitoring and pressure sensing. As shown in Fig. 6a , the TENG (30 mm × 5 mm) was designed and fixed on the finger to achieve contact separation movement when the finger was bent and straightened. When the bending angle of knuckles are periodically changed from 0° to 30–60–90°, the output V oc of the TENG sensor increases from 0.7 V to 2.6 V correspondingly. When the bending angle increases, the contact area between the finger and the dielectric layer also increase, resulting in a corresponding enhance in the output of the TENG sensor. Several periodic repeatable outputs at each fixed angle demonstrate the dependability and repeatability of this self-powered wearable sensor. Moreover, TENG can also be used as self-powered sensors for force recognition. The output of TENG sensors were measured under a series of applied loading forces ( Fig. 6b and c ). Under higher pressure, higher output voltage and current can be observed. This is due to the fact that the micro-scale elastic deformation under high contact pressure leads to closer contact at the interface. So the effective contact area will increase, generating a higher electrical output. Fig. 6d shows the peak values of voltage output as a function of applied pressure. When the applied pressure is lower than 65 kPa, the peak value of V oc increases at a rate of 0.46 V kPa −1 . When the pressure is higher than 65 kPa, the voltage values reach saturation. Fig. 6 TENG sensor for human motion detection and pressure sensing. (a) Open-circuit voltage responses when the TENG sensor has various bending angles on the finger joints. The V oc (b) and I sc (c) of the TENG as pressure sensor at different contact forces. (d) Summarized variation of peak amplitudes of the voltage with the contact pressure."
} | 5,436 |
39431068 | PMC11483375 | pmc | 104 | {
"abstract": "Spiders can produce up to seven different types of silk,\neach with\nunique mechanical properties that stem from variations in the repetitive\nregions of spider silk proteins (spidroins). Artificial spider silk\ncan be made from mini-spidroins in an all-aqueous-based spinning process,\nbut the strongest fibers seldom reach more than 25% of the strength\nof native silk fibers. With the aim to improve the mechanical properties\nof silk fibers made from mini-spidroins and to understand the relationship\nbetween the protein design and the mechanical properties of the fibers,\nwe designed 16 new spidroins, ranging from 31.7 to 59.5 kDa, that\nfeature the globular spidroin N- and C-terminal domains, but harbor\ndifferent repetitive sequences. We found that more than 50% of these\nconstructs could be spun by extruding them into low-pH aqueous buffer\nand that the best fibers were produced from proteins whose repeat\nregions were derived from major ampullate spidroin 4 (MaSp4) and elastin.\nThe mechanical properties differed between fiber types but did not\ncorrelate with the expected properties based on the origin of the\nrepeats, suggesting that additional factors beyond protein design\nimpact the properties of the fibers.",
"conclusion": "Conclusions We have shown that the basic NT-Rep-CT\nconstruct can host a broad\nrange of different amino acid sequences in the Rep region and still\nmaintain the ability to be spun into continuous fibers in our all-aqueous\nspinning system. Interestingly, we show that the primary structure\nof the Rep has a rather moderate influence on the resulting fibers’\nmechanical properties, at least for repeat regions ranging from 32\nto 43 kDa. We also found that spidroins with a molecular weight >45\nkDa were difficult to spin in the spinning setup used in this work.\nMoving forward, optimizing protein design, spinning methodologies,\nand postprocessing techniques will be crucial for realizing the full\npotential of artificial spider silk in various applications.",
"introduction": "Introduction Spider major ampullate silk is among the\nmost impressive fibers\nknown in terms of mechanical properties, which stem from a unique\ncombination of high tensile strength and extensibility. 1 − 4 In addition, spiders make up to seven distinct types of silk, 5 , 6 each endowed with unique properties tailored to each type’s\nspecific function. 7 , 8 For instance, major ampullate\nsilk possesses the highest strength and is used as a lifeline, 9 aciniform silk is more extensible and used for\nprey wrapping, 10 while flagelliform silk\ncoated with aggregate glue is extremely extensible and hence used\nfor making the capture spiral. 11 All silk types are mainly composed of spider silk proteins (spidroins),\nwhich in general are very large with a molecular weight of up to 300\nkDa. 12 , 13 The spidroins share a common architecture,\nfeaturing a repetitive (Rep) region, which is flanked by globularly\nfolded N-terminal (NT) 14 and C-terminal\n(CT) 15 domains. NT and CT are responsible\nfor the high solubility of the spidroins during storage and for polymerization\nof the spidroins when exposed to shear and a decreased pH. 15 − 19 The divergent mechanical properties of the silk fiber are likely\nto stem from disparities in the spidroins’ Rep region. 20 − 22 For instance, the major ampullate spidroins (MaSps) carry characteristic\npoly-Ala blocks that alternate with Gly-rich regions, 23 and according to the current model of the fiber structure–function\nrelationship, the poly-Ala form nanosized β-sheet crystals in\nthe mature fiber conferring the fiber strength, 24 while the Gly-rich segments make up an amorphous and flexible\nmatrix in which these crystals are embedded. 25 , 26 Another example comes from the flagelliform silk which is composed\nof spidroins that have Pro-rich repeat regions, which are believed\nto contribute to the extreme extensibility of this silk type. 27 Nevertheless, recent results suggest that this\ncorrelation is not straightforward, but that a combination of spidroins\nand other proteins is necessary to obtain the strength and extensibility\nof major ampullate silk. 28 In the\nsame way that spiders use their different silk types for\nspecific purposes, 29 fibers with different\nproperties are needed for specific industrial applications. For instance,\nan important selection criterion of fibers for clothing is comfort, 30 , 31 while sports goods could benefit from lightweight, elastic, and\ntough fibers. 32 The diverse and unique\nproperties of spider silks make the material attractive for many applications,\nfor instance, biomedical 33 − 36 and aerospace 37 applications,\nas durable components in robotics, 38 as\nsustainable wearables, 39 in sports goods, 40 and in other high-end textiles. 8 , 41 Most artificial silk fibers are made from recombinant spidroin-like\nproteins that often consist of the Rep unit alone (no terminal domains).\nFibers produced from large Rep proteins (300 kDa 42 − 44 ) can feature\nGPa strength. 43 , 44 However, this production method\ncomes with the disadvantage that solvents such as hexafluoropropanol\nand methanol are needed during the spinning process and that the protein\nyield is rather low. Another strategy for producing artificial spider\nsilk is to express natively folded mini-spidroins that consist of\nthe NT, a short Rep, and a CT. 45 − 50 The advantage of this approach is that the proteins can be purified\nusing native conditions, and the artificial silk fibers are formed\nby decreasing the pH and/or using shear forces. 45 − 50 This strategy, on the other hand, comes with the disadvantage that\nfibers are weaker, reaching a maximum strength of 250 MPa. 48 , 51 One explanation for the reduced strength compared to the non-native\nspinning methodology could be the comparatively small size of the\nmini-spidroins of less than 100 kDa. 46 − 49 , 51 , 52 With the aim to improve and specifically\ntailor the mechanical\nproperties of artificial silk fibers made from mini-spidroins and\nto understand the relationship between the protein design and the\nmechanical properties of the fibers, we systematically modified the\nmini-spidroin NT2RepCT. NT2RepCT has a Rep from MaSp1 featuring two\npoly-Ala blocks located between NT and CT, expresses at high very\nyields of up to 21 g/L with Escherichia coli , 53 and is extremely soluble in aqueous\nbuffers (500 mg/mL). 54 To modify NT2RepCT,\nwe explored four different strategies: (1) to vary the poly-Ala segment\nlength, (2) to increase the size of the MaSp1 region, (3) to insert\nRep segments from different spidroin types, and (4) to insert Rep\nsegments from other naturally occurring fibrous proteins. Thus, a\nbattery of 16 mini-spidroins was designed ( Figure 1 ), and fibers spun from these proteins were\ncharacterized by determining their mechanical properties. Figure 1 Using the basic\nmini-spidroin concept described by Andersson et\nal., 45 the Rep region in NT2RepCT was replaced\nwith other natural and engineered repeats. Altogether 16 proteins\nwere designed (listed in Table 1 ) with Rep regions originating from MaSp1 (three different\nconstructs with an increased size of the Rep region and two with shortened\npoly-Ala blocks), MaSp4 (two constructs), elastin (two constructs),\nFlSp (two constructs), TuSp (two constructs), resilin (one construct),\nand MiSp (one construct). Examples of Rep regions are shown in the\nlower panel, the complete sequences of NT, CT, and all Rep regions\nare listed in Table S1 , and a short description\nof each insert is provided in Table S2 .\nThe figure was made with Biorender.com using structures\nfrom Protein Data Bank (PDB) 2MFZ and 4FBS .",
"discussion": "Results and Discussion Expression and Purification In total, we designed 16\nconstructs that carry different Rep regions inserted between the terminal\ndomains (complete sequences are listed in Tables S1 and S2 ). Three of these carried Rep regions from the MaSp1\nfrom Euprosthenops australis , three\nfrom a MaSp4 from Caerostris darwini , 55 two from elastin from Homo sapiens , two from a tubuliform spidroin (TuSp)\nfrom Trichonephila clavipes , 56 two from a flagelliform spidroin (FlSp) from T. clavipes , 56 one from\nresilin from Drosophila simulans , 57 and one from minor ampullate spidroin (MiSp)\nfrom Araneus ventricosus . 58 These proteins were named after the protein\nfrom which the repetitive segment was derived and the number of amino\nacid residues in the repeat region. Two additional constructs were\nbased on the MaSp1 Rep region, in which the natural 15 and 14-residue-long\npoly-Ala stretches were replaced by four or eight consecutive Ala\nresidues, which were named MaSp1_A 4 and MaSp1_A 8 , respectively. As reference, already reported values for solubility\nand yield of the previously characterized NT2RepCT 51 , 53 , 54 (according to the nomenclature described\nherein, NT2RepCT would be called MaSp1_77) and (A 3 I) 3 -A 14 52 proteins were\nused. Most constructs expressed well and were soluble after\ncell lysis in a Tris-HCl buffer, reaching yields after purification\nof 26 and 250 mg/L in shake flask cultivations ( Table 1 and Figure S1 ). Overall, lower\nyields after purification were obtained for mini-spidroins that carried\nlonger Rep regions compared to shorter variants ( Figure S2 ). This is in line with previous publications 46 − 48 and could be due to the inability of the prokaryotic translational\nmachinery to cope with the high demands of specific amino acid residues, 42 increased aggregation propensity of long repetitive\nproteins, toxicity of the expressed proteins, and/or unfavorable secondary\nstructure formation of the repetitive mRNAs. 59 Table 1 List of Mini-Spidroins Studied in\nThis Work a strategy construct length of rep (# aa) M w (kDa) solubility after cell lysis yield after\npurification (mg/L)* spinnability control NT2RepCT 77 33.2 +++ 250 52 ++ (A 3 I) 3 -A 14 77 33.3 +++ 207 52 – poly-Ala length MaSp1_A 4 56 31.7 +++ 209 + MaSp1_A 8 64 32.2 +++ 170 – MaSp1 length MaSp1_110 110 35.7 +++ 183 ++ MaSp1_173 173 40.7 n.a 0 n.a MaSp1_237 237 45.8 n.a 0 n.a spidroin\ntypes MaSp4_125 125 38.2 +++ 231 ++ MaSp4_175 175 42.8 +++ 80 ++ MaSp4_252 252 49.7 ++ 98 – MiSp_206 206 42.1 +++ 114 ++ FlSp_132 132 36.8 ++ 157 ++ FlSp_232 232 44.7 +++ 61 – TuSp1_174 174 43.2 +++ 114 ++ TuSp2_345 345 59.5 +++ 35 – natural repeats Resilin_142 142 39.5 ++ 48 ++ Elastin_116 116 36.5 +++ 88 ++ Elastin_221 221 45.5 ++ 26 – a A compilation of 18 mini-spidroins\nincluding two previously characterized proteins (NT2RepCT 51 , 53 , 54 and (A 3 I) 3 -A 14 52 ). Expression values\nfor NT2RepCT and (A 3 I) 3 -A 14 were\nreported earlier by Arndt et al. 52 *Average\nof a 10 × 1L cultures. Solubility after cell lysis was rated\naccording to the following thresholds: + ++ almost all; + + more than\n50%; + less than 50%; – not soluble; n.a. since no protein\nwas expressed. The spinnability of the different constructs was rated\nas – not spinnable in that fibers were weak and fractured either\nwhen the fiber exited the capillary or while picking it up from the\nbath and guiding it to the collection wheel. + Spinnable, but discontinuous.\nTo collect enough fibers for tensile testing, the fiber had to be\npicked up several times and guided to the collection wheel. + + Continuous\nspinning. The fiber was picked up and continuously collected for 1–2\nmin yielding up to 60 m of fiber, before the process was aborted to\nensure evenly spaced fibers on the collection wheel. Spinning of Artificial Spider Silk The mini-spidroins\nthat could be isolated in a soluble form were concentrated and subjected\nto spinning using a protocol developed for NT2RepCT. 51 This protocol is purely aqueous-based but differs slightly\ncompared to previous studies published by our group 52 , 54 in terms of buffer in the coagulation bath, reeling speed, and the\ntip diameter of the capillary used for extrusion of the spinning dope\n(see the Materials and Methods section for\ndetails). Using these spinning conditions, the different mini-spidroins\nwere characterized as being continuously spinnable, discontinuous,\nor not spinnable. Spinnable fibers could be guided through the bath\nand collected onto a wheel rotating at 36 m/min. Some mini-spidroins\nformed fibers that were too fragile to be collected onto the rotating\nwheel and were then categorized as nonspinnable. Of the 18 mini-spidroin\nvariants (16 novel + 2 controls), 10 were possible to spin into continuous\nfibers, seven were nonspinnable, and one type could be collected onto\nthe wheel, but the spinning was not continuous ( Table 1 ). Mini-Spidroins with Different Lengths of the Poly-Ala Blocks The poly-Ala blocks found in the MaSp Rep region form crystals\nin the fiber, which are believed to confer the strength of the fibers.\nMolecular simulations have suggested that restricting the size of\nthe crystals is imperative for optimizing the strength since large\ncrystals will be more likely to bend and fracture than smaller ones. 24 , 60 At the same time, poly-Ala stretches composed of less than seven\nresidues are not expected to form β-sheet crystals. 61 To build on these results, Hu and co-workers\ninvestigated recombinant spidroins containing 5, 8, or 12 Ala residues\nin the poly-Ala motif. 62 They found that\nthe construct containing 12 Ala residues produced artificial silk\nfibers that were weaker compared with the other constructs. The eight\nAla spidroin produced the strongest (623 MPa) and toughest (107 MJ\nm –3 ) artificial silk fibers. Accordingly, we designed\ntwo variants of NT2RepCT in which the poly-Ala blocks were shortened\nto eight or four residues, respectively (MaSp1_A 8 and MaSp1_A 4 ). Surprisingly, MaSp1_A 8 was not spinnable, as\nfibers broke when guided through the spinning bath. Likewise, it was\ndifficult to spin and collect MaSp1_A 4 fibers, which is\nreflected in the mechanical properties of these fibers which were\namong those with the lowest strength and extensibility of all fibers\ninvestigated herein ( Table 2 ). Hence, reducing the number of Ala residues\nin the poly-Ala motif of NT2RepCT did not lead to improvement of the\nfibers’ mechanical properties. Table 2 Mechanical Properties of Spinnable\nConstructs in Order according to the Strength a construct diameter (μm) strain\nat break (%) eng. strength (MPa) Young’s modulus (GPa) toughness modulus (MJ m –3 ) MiSp_206 11.8 ± 0.6 93% ± 43% 74 ± 8 2.5 ± 0.4 59 ± 27 MaSp1_A 4 17.7 ± 6.3 6% ± 2% 75 ± 28 2.0 ± 1.0 2 ± 1 TuSp1_174 19.8 ± 3.4 6% ± 2% 76 ± 19 2.3 ± 0.6 3 ± 1 Resilin_142 13.5 ± 3.9 21% ± 26% 97 ± 18 2.7 ± 0.8 17 ± 24 FlSp_132 b 9.9 ± 3.0 93% ± 43% 98 ± 38 2.7 ± 0.8 74 ± 49 MaSp1_110 9.7 ± 1.7 106% ± 32% 105 ± 25 2.6 ± 0.9 75 ± 23 NT2RepCT a 8.8 ± 2.7 113% ± 28% 110 ± 30 2.2 ± 0.6 90 ± 30 MaSp4_125 b 9.0 ± 2.9 89% ± 26% 134 ± 38 2.4 ± 0.6 84 ± 35 MaSp4_175 9.8 ± 2.5 83% ± 31% 143 ± 78 3.3 ± 1.6 96 ± 61 Elastin_116 b 9.8 ± 2.9 105% ± 32% 143 ± 32 2.9 ± 1.1 109 ± 47 a For each construct, at least 10\nfibers were tensile tested. Outliers were not removed. The values\nreported after ± represent one standard deviation. Representative\nstress–strain curves are shown in Figure S2 . a Added for comparison purposes where the values\nrepresent the average from 89 fibers tensile tested that originate\nfrom 8 different spinning occasions, as reported in Schmuck et al. 51 b The values are averages of fibers\nmade from two or more spinning occasions and tensile testing of more\nthan 20 fibers. A graphical representation and statistical significance\ncompared to NT2RepCT is shown in Figure S4 . Representative stress–strain curves for all fiber types,\nexcept for NT2RepCT, are shown in Figure S5 . Representative stress–strain curves for NT2RepCT are shown\nin Schmuck et al. 51 MaSp1 Mini-Spidroins with Different Lengths of the Rep Region To investigate the impact of longer repeat regions, we extended\nthe repeat region of NT2RepCT so that the resulting mini-spidroins\nencompassed three poly-Ala blocks (MaSp1_110), five poly-Ala blocks\n(MaSp1_173), or seven poly-Ala blocks (MaSp1_237). We first assessed\nthe expression levels and benchmarked against NT2RepCT that can be\nexpressed and purified at high yields reaching 250 mg/L in shake flask\ncultivations. 52 , 54 From shake flask cultures, MaSp1_110\ncan be obtained at 183 mg/L, but surprisingly, we could not identify\nthe longer MaSp1_173 and MaSp1_237 on an SDS-PAGE of the cells after\nexpression ( Figure S3 ), which indicates\na very low or nonexisting expression. Interestingly, others such as\nHeidebrecht and co-workers succeeded in expressing constructs having\n24 poly-Ala blocks with a standard E. coli strain, showing that it is in principle possible to recombinantly\nexpress spidroins with longer Rep. 63 In\nthis particular example, each poly-Ala block hosted only 5 Ala residues,\nwhich is much fewer compared to the up to 15 consecutive Ala residues\nfound in our MaSp1 Rep segment. Nevertheless, we believe that the\nlonger poly-Ala block is probably not the only cause for the failed\nexpression of MaSp1_173 and MaSp1_237, since a recent study by Hu\net al. succeeded to express spidroins having 13 repeats where each\nrepeat hosted a block of 12 Ala residues with an expression level\nof 2.5 g/L using a bioreactor. 62 As far\nas we could judge, the authors in the two examples did not use any\nadditional vectors during expression with E. coli to elevate the alanyl- or glycyl-tRNA pool. 42 An additional factor that could explain the loss of expression would\nbe if the longer mini-spidroin variants are toxic to the bacteria,\nbut the exact reason for the abrupt drop in protein expression is\nnot known. The expression system presented here could potentially\nbe improved by screening different bacterial strains and promotors,\nusing a different codon optimization strategy, and possibly also by\ndesigning different mini-spidroin variants. Next, we attempted\nto spin fibers from MaSp1_110, which was successful. The mechanical\ncharacterization of these fibers showed no significant difference\ncompared to that of NT2RepCT fibers ( Table 2 and Figure S4 ). This may be a result of that the MaSp1_110 Rep region is not long\nenough to mediate a significant impact on fiber mechanical properties. In theory, longer repeat regions should result in an increased\nnumber of intermolecular interactions and, thereby, stronger fibers.\nIn line with this, several reports in the scientific literature investigated\nthis aspect, applying all-aqueous and native conditions for expression,\npurification, and spinning without the use of denaturing agents in\nany step of the process. Zhou et al. 48 found\na strong dependency of strength and size when inserting aciniform\nspidroin repeat segments into a mini-spidroin. When such a small mini-spidroin\nwas spun (45.8 kDa), the resulting fibers had a strength of 52 MPa,\nbut when the repetitive segment was extended, which resulted in a\n104 kDa construct, the resulting fibers had a strength of 245 MPa.\nEven though these results are quite impressive, the expression was\noverall low, and the larger 104 kDa variant was not soluble in aqueous\nbuffers. In a follow-up study, 44 and 96 kDa recombinant spidroins\nwith repeats from FlSp were spun into fibers. Interestingly, despite\nthe relatively large difference in molecular weight, the fiber strength\nwas moderately increased from 182 to 253 MPa using larger spidroins.\nAt the same time, these fibers were rather brittle with a strain at\nbreak of 12%, which is surprising considering the impressive extensibility\nof the flagelliform silk. 46 In the literature,\nthere are also conflicting reports regarding the relationship between M w and the mechanical properties of the fibers.\nIn a study where spidroins containing different lengths of the MaSp1\nrepeat region were spun into fibers, the largest constructs (60.8\nkDa) gave inferior fibers compared to a 42.5 kDa variant, which could\nbe spun into a fiber with a strength of 149 MPa. 47 The emerging inconclusive picture of the effect of longer\nRep segments on the mechanical properties could be a result of numerous\nparameters, which are discussed further below. Since we were\nunsuccessful in expressing the mini-spidroins based\non MaSp1 from Euprosthenops australis with repeat regions exceeding 110 residues, we attempted to produce\nlonger variants with repeats originating from other spidroins. Mini-Spidroins with Rep Regions from Different Spidroin Types The mechanical properties of different spider silks vary significantly,\nand this is related to differences in the primary structure of the\nRep region. 22 , 64 As mentioned before, the major\nampullate silk, made up primarily of MaSps, 65 , 70 is the strongest of the different silks. 66 The MaSp4 has been identified in the spider species Caerostris darwini ( 55 , 68 ) and Araneus ventricous ( 67 ) and\nhas been associated with an increased extensibility of the fiber. 55 To examine the influence of MaSp4 Rep on our\nartificial silk is especially interesting because this spidroin type\nhas been scarcely studied. MiSps are the main constituents of the\nminor ampullate silk, which have a lower tensile strength than major\nampullate silk, but increased extensibility. Flagelliform silk stands\nout as the most extensible fiber (>200%) of all of the different\nsilk\ntypes which relates to the Pro-rich repeat region of the FlSps, while\ntubuliform silk is intermediate in terms of both strength and strain. 66 Thus, we expected that the insertion of different\nRep between NT and CT would make the artificial silk fibers substantially\ndifferent compared to reference NT2RepCT fibers, even more so as we\nhave previously seen a drastic effect by introducing only very few\nmutations in the Rep of NT2RepCT. 52 The new mini-spidroins containing Rep from different spidroin types\ncould all be successfully spun into fibers ( Table 2 ), except for constructs having more than\n220 residues in Rep (FlSp_232, MaSp4_252, and TuSp2_345). The longest\nconstructs were classified as not spinnable because the fibers were\ntoo fragile to be collected or because of aggregate formation in the\nspinning dope, as in the case of TuSp2_345, for which extrusion was\ninhibited due to clogging. Among the eight mini-spidroin variants\nwith Rep regions from different spidroin types, only MaSp4_125 and\nMaSp4_175 were significantly stronger. These fibers had an average\nstrength of 134 and 143 MPa, respectively, compared to 110 MPa for\nNT2RepCT ( Table 2 and Figure S4 ). Since NT2RepCT, MaSp4_125, and MaSp4_175\nhave similar M w and identical terminal\ndomains, the increase in strength could be related to the nature of\nthe MaSp4 repeat. The results are surprising since the MaSp4 is Pro-rich\nand void of poly-Ala blocks, which should give a more extensible fiber\naccording to the current structure–function models. Fibers\nobtained from the mini-spidroin constructs MiSp_206, and TuSp_174\nwere significantly weaker. The strain at break for all constructs\nwith Rep from different spidroins was above 83%, which is comparable\nto NT2RepCT fibers. An exception is the construct TuSp1_174 which\nresulted in brittle fibers. In summary, we did not achieve mechanical\nproperties that correlated to the behavior of the native silk fibers\nthat the respective repeats were derived from, and moreover, the mechanical\nproperties of the fibers made from the new mini-spidroins were very\nsimilar to the mechanical properties of NT2RepCT fibers. Mini-Spidroins with Rep Regions from Other Fibrous Proteins In nature, the elastomers tropoelastin and resilin stand out as\nexamples of proteins that are found in tissues with impressive mechanical\nproperties. Tropoelastin is found in connective tissues and is cross-linked\nvia Lys residues into elastin. 71 Silk-elastin\nhybrid materials are frequently described in the literature in the\ncontext of making biomaterials, for instance, to assemble tissue scaffolds\nby electrospinning, 72 to tune the mechanical\nproperties of hydrogels, 73 and for modulating\nthe self-assembly kinetics of silk/elastin hybrids to make nanomaterials. 74 Herein, we could efficiently produce and spin\nfibers from a 36.5 kDa silk-elastin hybrid protein (Elastin_116),\nbut the longer variant of this protein (Elastin_221) could not be\nspun into continuous fibers, as was also the case for the longer constructs\nwith Rep from other spidroins. Notably, the fibers spun from Elastin_116\nproteins were the toughest among all fibers investigated herein, reaching\n143 MPa in strength, a strain at break of 105%, a Young’s modulus\nof 2.9 GPa, and a toughness modulus of 109 MJ/m 3 . Resilin is found in elastic tissues of insects and arthropods and\nthe proteins are cross-linked via Tyr residues. 75 , 76 Resilin and resilin/silk hybrids have been mainly investigated for\ntheir use as biomaterials, 77 but there\nare also attempts to increase the mechanical properties of silkworm\nsilk by inserting the resilin gene into the silkworm genome. 78 In our hands, fibers made from a mini-spidroin\ncarrying Rep from Resilin (Resilin_142) could be continuously spun.\nThe strength of Resilin_142 fibers was within the experimental error\ncompared to NT2RepCT fibers, but the strain at break was significantly\nreduced ( Figure S4 and Table 2 ). To our knowledge, there\nare no previous examples of silk-elastin\nor silk-resilin hybrid proteins that have been continuously wet-spun\ninto macroscopic fibers. Silk-resilin/elastin hybrid proteins could\nin theory harness the high deformability and resilience of the respective\nnatural materials, but to achieve this, we likely have to find ways\nto allow the repeat regions to fold properly and to achieve the intermolecular\ncross-links that are present in the native materials. Control Proteins As already mentioned before, the spinning\ncondition used in this study were optimized for NT2RepCT 51 and have not been tested on the engineered variant\n(A 3 I) 3 -A 14 , which has been shown\nto form fibers that are as tough as native spider silk in a recent\nstudy. 52 Interestingly, the (A 3 I) 3 -A 14 was in this study categorized as nonspinnable\nand is therefore not listed in Table 2 , because the fiber broke as soon as it was placed\non the collection wheel. This is likely due to that the spinning conditions\nused herein were different from those in Arndt et al. 52 Specifically, the glass capillaries used in the current\nstudy had a larger opening, the spinning bath did not contain NaCl,\nand the speed of collection was much faster. These results attest\nto previously published notions, that in order to achieve the best\npossible mechanical properties of a fiber spun from a specific mini-spidroin,\nthe spinning conditions need to be screened and optimized systematically. 51 However, performing such large screens for the\nrelatively large number of proteins investigated herein would require\nseveral years’ work. General Discussion The main intention of the present\nstudy was to investigate if and how different primary structures of\nthe Rep region in mini-spidroins affect the mechanical properties\nof artificially spun spider silk when using all-aqueous spinning conditions.\nTo our surprise, the strength of all as-spun fibers using identical\nspinning conditions was found to be within a rather narrow range of\n74–143 MPa. Interestingly, the only constructs that were better\nin terms of strength compared to the control mini-spidroin NT2RepCT\n(MaSp1_77) are constructs based on MaSp4 (up 143 MPa) 55 and elastin (143 MPa) ( Table 2 ). None of the fibers developed herein had\na higher strain at break than the control NT2RepCT (113%), but the\nmajority of the constructs (73%) resulted in fibers with a strain\nat break above 50%. Only the MaSp1_A 4 and TuSp1_174 fibers\nfailed at the yielding point. The fiber with the highest strength\nthat was significantly different compared to NT2RepCT was made from\nElastin_116. The strain at break values of NT2RepCT and Elastin_116\nfibers were comparable, as was the toughness modulus ( Figure S4 ). We therefore conclude that the mechanical\nproperties of fibers spun from mini-spidroins with diverse primary\nstructures in their repeat regions are surprisingly similar. The relatively\nsmall improvements we could achieve in this cohort of proteins are\nin the range that can be achieved when changing the spinning conditions. 51 The question remains why most mini-spidroins\ninvestigated in this study did not give rise to fibers with more diverse\nmechanical properties and failed to reflect the properties of the\nrespective native silk fibers. One explanation for this observation\ncould be that the spinning conditions were not optimized for the respective\nconstruct, which could have substantially different optima concerning\nthe buffer type, buffer concentration, pH, extrusion rate, and dope\nconcentration. The biology of all glands, except the major ampullate\ngland, remains largely unexplored, which means that there may be unknown\nconditions or components that are important for the correct polymerization\nof a specific silk type. Specifically, several studies have shown\nthat there are several nonclassical spider silk proteins in the major\nampullate silk, all of which have unknown functions. 65 , 70 , 69 These proteins and the combination\nof different proteins could potentially be important for the mechanical\nproperties. However, even though the presence of other spidroins and\nnonspidroins should not be neglected, 65 , 70 the bulk of\neach fiber type is made from the corresponding spidroin type, 70 and thus they should be a main contributor to\nthe mechanical properties. Another factor to take into account could\nbe that the spidroin terminal domains may have evolved to function\noptimally with a specific repeat, which could also influence the results\nas we used the same NT and CT in all protein constructs. Yet another\ncontributing factor could be that the repeat regions affect the ability\nof the mini-spidroins to form liquid–liquid phase-separated\n(LLPS) droplets before polymerization, which has been suggested to\nbe an important step in the spinning process and deserves further\ninvestigation. 50 , 79 − 82 Moreover, shear forces are important\nfor alignment and to induce the formation of β-sheets, 83 but the extrusion of the spidroins through the\nglass capillary used herein may not be enough to align the larger\nspidroins, thus causing a suboptimal arrangement and as a consequence\na weaker fiber is obtained. The significantly smaller size of the\nmini-spidroins compared to the native spidroins could of course also\naffect the properties of the fiber if they are too short to form correct\nintermolecular interactions ( Figure 2 ). When using denaturing conditions for the preparation\nof the spinning dope, the literature suggests that spidroins with\nan M w comparable to native spidroins (up\nto 300 kDa) are required for making fibers with a strength of 1 GPa. 42 , 43 , 63 , 84 Considering this, it may not be too surprising that we could not\nimprove the strength more by spinning longer proteins, as the largest\nspinnable spidroin in this study had an M w of only 43 kDa. Although our results suggest that there is no correlation\nbetween strength and molecular weight ( Figure 2 ), it should be noted that molecular weight\ndifference between the smallest and largest spidroin is only 37%,\nwhich may be too little to detect effects of protein molecular weight. Figure 2 Mechanical\nproperties for artificial silk fibers plotted against\nthe molecular weight of the mini-spidroins used in this study. The\ndifferent colors of the circles are used to represent the different\nmini-spidroins where white represents MaSp1_A 4 , black NT2RepCT,\ngray MaSp1_110, dark blue FlSp_132, blue Elastin_116, light blue MaSp4_125,\ndark green Resilin_142, green MiSp_206, light green MaSp4_175, and\npurple TuSp1_174. R 2 was obtained by least-squares regression.\n\nGeneral Discussion The main intention of the present\nstudy was to investigate if and how different primary structures of\nthe Rep region in mini-spidroins affect the mechanical properties\nof artificially spun spider silk when using all-aqueous spinning conditions.\nTo our surprise, the strength of all as-spun fibers using identical\nspinning conditions was found to be within a rather narrow range of\n74–143 MPa. Interestingly, the only constructs that were better\nin terms of strength compared to the control mini-spidroin NT2RepCT\n(MaSp1_77) are constructs based on MaSp4 (up 143 MPa) 55 and elastin (143 MPa) ( Table 2 ). None of the fibers developed herein had\na higher strain at break than the control NT2RepCT (113%), but the\nmajority of the constructs (73%) resulted in fibers with a strain\nat break above 50%. Only the MaSp1_A 4 and TuSp1_174 fibers\nfailed at the yielding point. The fiber with the highest strength\nthat was significantly different compared to NT2RepCT was made from\nElastin_116. The strain at break values of NT2RepCT and Elastin_116\nfibers were comparable, as was the toughness modulus ( Figure S4 ). We therefore conclude that the mechanical\nproperties of fibers spun from mini-spidroins with diverse primary\nstructures in their repeat regions are surprisingly similar. The relatively\nsmall improvements we could achieve in this cohort of proteins are\nin the range that can be achieved when changing the spinning conditions. 51 The question remains why most mini-spidroins\ninvestigated in this study did not give rise to fibers with more diverse\nmechanical properties and failed to reflect the properties of the\nrespective native silk fibers. One explanation for this observation\ncould be that the spinning conditions were not optimized for the respective\nconstruct, which could have substantially different optima concerning\nthe buffer type, buffer concentration, pH, extrusion rate, and dope\nconcentration. The biology of all glands, except the major ampullate\ngland, remains largely unexplored, which means that there may be unknown\nconditions or components that are important for the correct polymerization\nof a specific silk type. Specifically, several studies have shown\nthat there are several nonclassical spider silk proteins in the major\nampullate silk, all of which have unknown functions. 65 , 70 , 69 These proteins and the combination\nof different proteins could potentially be important for the mechanical\nproperties. However, even though the presence of other spidroins and\nnonspidroins should not be neglected, 65 , 70 the bulk of\neach fiber type is made from the corresponding spidroin type, 70 and thus they should be a main contributor to\nthe mechanical properties. Another factor to take into account could\nbe that the spidroin terminal domains may have evolved to function\noptimally with a specific repeat, which could also influence the results\nas we used the same NT and CT in all protein constructs. Yet another\ncontributing factor could be that the repeat regions affect the ability\nof the mini-spidroins to form liquid–liquid phase-separated\n(LLPS) droplets before polymerization, which has been suggested to\nbe an important step in the spinning process and deserves further\ninvestigation. 50 , 79 − 82 Moreover, shear forces are important\nfor alignment and to induce the formation of β-sheets, 83 but the extrusion of the spidroins through the\nglass capillary used herein may not be enough to align the larger\nspidroins, thus causing a suboptimal arrangement and as a consequence\na weaker fiber is obtained. The significantly smaller size of the\nmini-spidroins compared to the native spidroins could of course also\naffect the properties of the fiber if they are too short to form correct\nintermolecular interactions ( Figure 2 ). When using denaturing conditions for the preparation\nof the spinning dope, the literature suggests that spidroins with\nan M w comparable to native spidroins (up\nto 300 kDa) are required for making fibers with a strength of 1 GPa. 42 , 43 , 63 , 84 Considering this, it may not be too surprising that we could not\nimprove the strength more by spinning longer proteins, as the largest\nspinnable spidroin in this study had an M w of only 43 kDa. Although our results suggest that there is no correlation\nbetween strength and molecular weight ( Figure 2 ), it should be noted that molecular weight\ndifference between the smallest and largest spidroin is only 37%,\nwhich may be too little to detect effects of protein molecular weight. Figure 2 Mechanical\nproperties for artificial silk fibers plotted against\nthe molecular weight of the mini-spidroins used in this study. The\ndifferent colors of the circles are used to represent the different\nmini-spidroins where white represents MaSp1_A 4 , black NT2RepCT,\ngray MaSp1_110, dark blue FlSp_132, blue Elastin_116, light blue MaSp4_125,\ndark green Resilin_142, green MiSp_206, light green MaSp4_175, and\npurple TuSp1_174. R 2 was obtained by least-squares regression."
} | 9,215 |
39868331 | PMC11760428 | pmc | 105 | {
"abstract": "Quorum sensing (QS) is a mechanism of intercellular communication that enables microbes to alter gene expression and adapt to the environment. This cell-cell signaling is necessary for intra- and interspecies behaviors such as virulence and biofilm formation. While QS has been extensively studied in bacteria, little is known about cell-cell communication in archaea. Here we established an archaeal model system to study QS. We showed that for Haloferax volcanii, the transition from motile rods to non-motile disks is dependent on a possibly novel, secreted small molecule present in cell-free conditioned medium (CM). Moreover, we determined that this putative QS molecule fails to induce the morphology transition in mutants lacking the regulatory factors, DdfA and CirA. Using quantitative proteomics of wild-type cells, we detected significant differential abundances of 236 proteins in the presence of CM. Conversely, in the Δ ddfA mutant, addition of CM resulted in only 110 proteins of significant differential abundances. These results confirm that DdfA is involved in CM-dependent regulation. CirA, along with other proteins involved in morphology and swimming motility transitions, is among the proteins regulated by DdfA. These discoveries significantly advance our understanding of microbial communication within archaeal species.",
"introduction": "Introduction Microorganisms employ many different forms of intercellular communication critical for survival. The most well-studied mechanism of microbial cell-cell communication is quorum sensing (QS), in which microorganisms synthesize and secrete small molecules or peptides called autoinducers. Recognition of the autoinducers at varying extracellular concentrations enable the microbe to regulate gene expression as a result of changes in their population density. Only when a density threshold is reached, or quorum of molecules is recognized, do the single cells express genes to orchestrate collective behaviors such as bioluminescence, biofilm formation, competence, motility and sporulation 1 - 4 . Studies about bacterial intercellular communication began with competence factor ComX in Streptococcus pneumoniae 5 , 6 and the acyl-homoserine lactone (AHL) regulating bioluminescence in Vibrio species 7 - 9 . In subsequent years, many other QS molecules were identified and characterized, such as autoinducer-2 (AI-2) 10 , 11 , Pseudomonas quinolone signal (PQS) 12 , autoinducing peptide (AIP) of gram-positive bacteria 13 , and diffusible signal factor (DSF) 14 . These QS signaling molecules are synthesized and recognized by different mechanisms. One of the best-defined QS pathways is the LuxI/LuxR system. LuxI catalyzes the synthesis of AHL molecules that can diffuse across the membrane. At a critical extracellular concentration of AHL molecules, signifying that a quorum of bacterial kin is present, AHL can bind to and stabilize the LuxR dimer, which then promotes transcription of QS activated genes 15 . As another example of a QS pathway, in Gram-positive bacteria, QS behaviors can be mediated by two-component systems. Autoinducer molecules bind histidine kinase receptors in the membrane leading to phosphorylation of cognate response regulator proteins. The majority of response regulators in Gram-positive QS pathways function through transcriptional regulation of target genes 16 . In this way, characterizing both the molecules and mechanisms driving microbial intercellular communication enables us to eavesdrop on the microbial chatter, deepening our understanding of single-cellular life as well as providing targets for therapeutics or disinfectants that can interrupt the crosstalk 17 . In recent years, QS mechanisms have been discovered in non-bacterial microorganisms as well. For the fungus Candida albicans, farnesol is an important signal for morphology changes 18 , for the methanogenic archaeon Methanosaeta harundinacea 6Ac, a carboxylated AHL regulates cell assembly and carbon metabolic flux 19 , and even for the virus, phi3T phage, the SAIRGA peptide mediates the phage lysis versus lysogeny decision in the host 20 . These discoveries emphasize the evolutionary significance of density-dependent signaling and its fundamental role in microbial life. Archaea are ubiquitous in the environment 21 , play critical roles in geochemical cycles 22 , and have demonstrated applications in biotechnology 23 , 24 , yet only a handful of studies have explored the potential for intercellular signaling in archaea. In two bioinformatic studies, putative LuxR solos (LuxR homologs without a cognate LuxI homolog) were identified in many archaeal species 25 , and a strong correlation was observed between metabolic regulation and candidate QS genes in ammonia-oxidizing archaea 26 . Interestingly, multiple studies have revealed that crude supernatant extract from different archaeal species can stimulate AHL-dependent QS phenotypes of bacterial bioreporters, including M. harundinacea supernatant containing the carboxy-AHL and Haloterrigena hispanica supernatant containing a class of diketopiperazine (DKP) compounds as the active molecules 19 , 27 - 30 . These findings suggest the potential for inter-domain crosstalk. However, the identity of the active molecules in most of the archaeal strains tested in the various bioreporter assays remain unknown 27 , 29 , 30 , and it is also unknown whether the DKP molecules from Htr. hispanica act as true QS signals that induce changes in the archaeon’s behavior 28 . Only three examples of true archaeal QS, or population density-dependent phenotypic changes in archaea mediated by secreted molecules, have been described: M. harundinacea morphology changes mediated by carboxy-AHL 19 , growth-phase dependent biosynthesis of Natrialba magadii Nep protease (unknown signal molecule) 31 , and increased biofilm biomass of Halorubrum lacusprofundi upon the addition of culture supernatant (unknown signal molecule) 29 , 32 . Furthermore, M. harundinacea filI gene is the only gene discovered in any archaeon that has been empirically determined to be involved in a QS pathway 19 . The dearth of information about archaeal QS may be a result of the potential challenges associated with culturing archaea, such as the specific metabolic and environmental needs of extremophilic archaea, as well as the general lack of archaeal genome annotations 33 . Because observing QS-dependent phenotypes in the lab requires microbial cultures, there remains a need for the development of an easily culturable and genetically tractable archaeal model system to study archaeal cell-cell signaling. The halophilic archaeon Haloferax volcanii is easy to culture in the lab and has a wide array of established genetic and biochemical assays 34 . This model archaeon also displays distinct morphology differences at varying population densities, indicating the potential for QS-related behaviors. In liquid culture, cells initially present as motile rods at early log growth phase. As the culture propagates, cells begin to transition to pleomorphic disks, presumably increasing the surface area to volume ratio. By the time the cell population levels off at its highest density, cells are exclusively disk-shaped 35 - 37 . This morphological change can also be observed in swimming motility halos on soft agar plates, with non-motile disks at the dense center of the halo, and motile rods at the less dense, leading edge 38 . In this report, we demonstrate that Hfx. volcanii is an invaluable model for deciphering QS in archaea. Our interdisciplinary studies strongly suggest that a small molecule, one that has not been previously identified as involved in density-dependent signaling, is secreted by Hfx. volcanii to mediate the morphology transition. We identified two components, DdfA and CirA, involved in the QS response and employed the deletion strain Δ ddfA in quantitative proteomics to identify additional potential QS components. Thus, by developing methods to analyze a previously uncharacterized QS system in the model archaeon Hfx. volcanii, we pave the road for future studies into archaeal cell-cell communication.",
"discussion": "Discussion In this interdisciplinary study, we developed a system to learn about archaeal QS using the model halophile Hfx. volcanii, combining microbiology with proteomics and analytical chemistry. Our results demonstrate that Hfx. volcanii secretes a QS signal (DFS) into the extracellular environment that regulates the rod-to-disk transition ( Fig. 1 ). We show that DFS is a population-density dependent signal because the addition of increasing amounts of CM to fresh cultures as well as higher OD 600 of culture from which CM is prepared leads to a higher proportion of disk-shaped cells at early log ( Fig. 1b , 1c ). In addition to the QS-mediated morphology transition, we show that the addition of CM to motility agar can arrest Hfx. volcanii motility ( Fig. 2 ). A larger volume of CM is necessary for visualizing the motility cessation phenotype than for the shape transition phenotype ( Fig. S2 ), possibly due to agar sequestration of DFS or shaking during liquid culture incubation. Thus, in this study we establish two distinct, easily observable and reproducible QS-mediated phenotypes in Hfx. volcanii. Furthermore, we demonstrate that DFS is a small molecule that can be enriched via a centrifugal filter ( Fig. 5 ) and can be extracted by ethyl acetate ( Fig S5 ), similar to other QS signals 9 . Prior to our study, no proteins involved in the archaeal signaling pathways in response to QS signals had been identified. Using our model system, we were able to identify two existing mutants, Δ ddfA and Δ cirA, that retained the rod shape and their ability to swim in the presence of CM ( Fig. 3 ). In fact, the previously shown hypermotility phenotypes of Δ ddfA and Δ cirA 39 , 40 may be caused by their lack of DFS signal recognition and response. While little is known about the function of DdfA in disk formation 39 , our previous work has shown that CirA plays a role in transcriptional regulation of the genes encoding archaella biosynthesis: without CirA, downregulation of the arl genes arlA1, arlI, and arlJ does not occur 40 . Since we discovered Δ ddfA and Δ cirA as, to the best of our knowledge, the first mutants that do not display the Hfx. volcanii QS response, we can now expand our understanding of the roles of DdfA and CirA in QS. Subsequent quantitative proteomics of wild-type Hfx. volcanii grown with and without CM revealed significant differential abundance of 236 proteins, including significant abundance decreases of proteins important for motile rods and increases in proteins important for disks in the presence of CM ( Fig. 4b , Table S2 ). We also observed an increase in CirA and proteins involved in N -glycosylation. Previous data suggest glycosylation may play a role in motility and shape regulation, and colonies of Δ agl15 look similar to Δ cirA : smaller and darker than wild-type colonies 40 , 45 . Similar to previous studies 39 , 41 , DdfA was not identified in our MS analyses, however, further insight into its role in CM-dependent regulation was obtained from quantitative proteomics of the Δ ddfA strain in the presence and absence of CM. Although the morphology and motility phenotype of Δ ddfA suggests that this disk-defective mutant lacks certain responses to DFS, the identification of 110 proteins with significant differential abundance in Δ ddfA when grown in 1% CM suggests DdfA-independent CM responses ( Fig. 4d , Table S4 ). Because 77 proteins overlap in the “wild type, no CM to CM” and “Δ ddfA, no CM to CM” comparisons ( Fig. 4c ), it is likely that those proteins are involved in the DFS response pathway independent of the role of DdfA or are responding to other components of CM in a DdfA-independent way. A subset of proteins that increases or decreases in this manner is depicted in Fig. 7 . Similarly, the set of 159 proteins that change in abundance in the “wild type, no CM to CM” comparison but not in the “Δ ddfA, no CM to CM” comparison provide us with a set of proteins that are regulated in a DdfA-dependent manner in response to CM. This set includes CirA, CetZ1, ArlG, chemotaxis proteins, certain glycosylation pathway proteins, and potential proteins in two-component signaling. At the moment it is not clear whether DdfA directly responds to DFS or whether DdfA is indirectly affected by DFS. However, these quantitative proteomics results show the importance of QS in archaeal regulation, provide an invaluable dataset of proteins involved in QS responses, and give us first insights into the QS-mediated regulatory network in Hfx. volcanii ( Fig. 7 ). Finally, our data reveal not only the importance of QS in archaea but also provide a strong foundation for further studies into archaeal QS using the model archaeon Hfx. volcanii. We showed that DFS is a stable small molecule ( Fig. 5 ), but it also appears that DFS is highly potent at low concentrations ( Fig. 1b ), rendering chemical isolation and purification difficult. We also cannot exclude that more than one molecule is responsible for the QS-mediated phenotypes. Because there is a possibility that DFS is synthesized via an autoregulatory pathway, similar to AHL in the LuxI/LuxR system 15 , the proteins involved in DFS biosynthesis may be found in our proteomics datasets and could shed light on the chemical structure of this QS signal. Therefore, future studies using multi-pronged approaches that combine chemical purification strategies and screens for mutants unable to produce DFS, along with the proteomic results obtained from this study, will aid in delineating the Hfx. volcanii DFS pathway. These interdisciplinary approaches will not only advance our understanding of how Hfx. volcanii communicates with itself, but also help us to learn more about interspecies and inter-domain interaction."
} | 3,499 |
36903366 | PMC10003856 | pmc | 108 | {
"abstract": "Silk from silkworms and spiders is an exceptionally important natural material, inspiring a range of new products and applications due to its high strength, elasticity, and toughness at low density, as well as its unique conductive and optical properties. Transgenic and recombinant technologies offer great promise for the scaled-up production of new silkworm- and spider-silk-inspired fibres. However, despite considerable effort, producing an artificial silk that recaptures the physico-chemical properties of naturally spun silk has thus far proven elusive. The mechanical, biochemical, and other properties of pre-and post-development fibres accordingly should be determined across scales and structural hierarchies whenever feasible. We have herein reviewed and made recommendations on some of those practices for measuring the bulk fibre properties; skin-core structures; and the primary, secondary, and tertiary structures of silk proteins and the properties of dopes and their proteins. We thereupon examine emerging methodologies and make assessments on how they might be utilized to realize the goal of developing high quality bio-inspired fibres.",
"conclusion": "9. Conclusions We have herein outlined the experimental processes and procedures for examining silk molecular and structural organization; mechanical, vibratory, optical, thermal and conductive properties; gross structural integrity; and protein amino acid composition and molecular weight, among others, using a range of standard and improvised biological, chemical, and physical techniques. It is clear that any analysis hoping to tie together the molecular, structural, and bulk fibre functional properties of any kind of silk should perform a suite of experiments across a range of scales, as described herein. There have been many studies that have successfully utilized a broad suite of analyses, and it is through these studies that we have gained exceptional insights into how silk functions across scales. Moving forward, we recommend future studies utilize similar, or even wider, scopes of procedures. This is particularly important in light of the growing interest in investing a lot of time and resources into transgenic and recombinant technologies to build new high performance fibres at a scale to solve some of the most pressing challenges currently facing textiles and other industries.",
"introduction": "1. Introduction Silk is a natural proteinaceous secretion with a high molecular weight and repetitive peptide sequences that is produced by many invertebrates [ 1 , 2 , 3 , 4 ]. Silks are most commonly secreted as a fibre, but may be secreted as a glue or other non-fibrous material [ 4 ]. Of the impressive diversity of silks produced, the fibrous silks of the silkworm moths (the silk used in textiles) and spiders are by far the most well researched, so they are the primary focus of this review. Spider and silkworm silk fibres generally have a density of around 1.3 g cm −3 [ 5 ], which is lower than most other natural fibres (e.g., wool or cotton, at ~1.5 g cm −3 [ 6 ]). Despite having low density, many silks exhibit immense strength, elasticity, and toughness [ 5 , 6 ]. It is these properties, along with its biocompatibility, electro-thermal and vibratory conductivity, optical properties, and environmentally benign methods of synthesis, that render the fibres of immense interest to researchers and engineers seeking to develop bio-inspired technologies [ 7 , 8 ]. For thousands of years, sericulture has enabled the utilization of strong, tough, smooth, and shiny silkworm silk fibres as fabrics and textiles. More recently, silk fibres have been reconstituted for use as biomedical and other functional devices [ 4 , 9 , 10 , 11 ]. Spiders produce up to seven unique silks [ 12 ], with each silk type comprised of multiple proteins [ 13 , 14 ] produced in seven distinct glands, with unique transcriptional regulation across species yielding fibres with variable structural and mechanical properties. Of the spider silks, it is dragline or major ampullate (MA) silk that has the greatest toughness [ 15 ], and so is the most well characterized. The extremely tough spider dragline silks cannot be commercially harvested, as spiders must be forcibly silked for dragline collection in a labour-intensive process and are not able to be domesticated due to their territorial and cannibalistic nature [ 16 ]. Accordingly, attempts are being made to create recombinant and/or transgenic lines of silk using bacteria, yeast, plant, or animal hosts. While recent production of intact fibres from recombinant and transgenic silk production lines is showing promise [ 9 , 17 , 18 ], it is still notoriously difficult for artificial silks to recapture the mechanical and other properties of their natural counterparts. In spite of many physiological differences among silkworm and spider silk glands, they all secrete a liquid spinning dope of high ionic strength and concentrated with silk protein monomers, which are spun into fibres at spinnerets. Within the silk gland and duct, physiological and biochemical mechanisms regulate the transformation of the dope to a fibre. For instance, a fall in pH within the duct induces an increase in protein concentration, thus an enhanced viscosity of the dope. This produces a fluid shear stress on the dope as it flows from the sac into the spinning duct of the gland. The folds within the gland induce additional shear stresses on the dope as it flows through it, before solidifying prior to secretion at the spinneret. A skin consisting of glycoprotein, lipid, and/or various kinds of coating proteins is secreted along the surface of the fibril core at extrusion [ 4 , 19 , 20 , 21 , 22 ]. Figure 1 shows the hierarchical structure of spider and silkworm silks. The spun fibres of both have a unique hierarchical structure, consisting of a protein core of tightly packed nanofibrils within a protein matrix that can be separated by solubilizing in high salt-reducing solutions, surrounded by a thin glycoprotein-rich skin [ 2 ]. Silkworm silks are further surrounded by a layer of sticky protein called sericine. The fibril core is composed of proteins called fibroins or, in the case of spider silk, spidroins (a derivation of spid er fib roins ) [ 2 , 23 , 24 ]. These consist of highly conserved amino (N-) and carboxyl (C-) terminal domains, between which are the repetitive regions that make up the bulk (~95%) of the protein [ 25 , 26 ]. Most of the repetitive region is made up of short repetitive amino acid motifs of predominantly alanine and glycine with, in some instances, other amino acids, such as proline, tyrosine, and serine [ 27 , 28 ]. Within the silk fibre, the proteins are orientated along the fibre axis and formed primarily of β-sheets, nanocrystals that stack up to become hard and stiff, thus imparting strength to the fibre. The nanocrystals are suspended within a rubbery matrix of random coils and α-helices, the so-called “amorphous region” (see Figure 1 ), which is responsible for the silk’s extensibility [ 29 , 30 , 31 ]. Multiple experiments and simulations show how arrangements within the skin and core influence the properties of silk [ 2 , 32 ], but a grasp of the mechanisms by which the structure induces function is not clearly discernible since most of the studies have been primarily descriptive and only examined features at a single hierarchical scale. It is through meticulous testing of native silk fibres, isolated silk proteins, and the soluble protein monomers found within the precursor spinning dopes that we can begin to understand how the properties of native silks are realized. The accumulated evidence suggests silk has a hierarchical structure, as depicted in Figure 2 , whereby its sub-molecular primary structural order specifically affects its bulk fibre conformational properties via secondary structure conformations. In this review we thus comprehensively overview various procedures that can, and should, be effectively used together to examine the nano-scale interactions, production processes, and resultant multi-scaled properties of silkworm and/or spider silks."
} | 2,037 |
31683537 | PMC6862673 | pmc | 111 | {
"abstract": "One of the main obstacles for memristors to become commonly used in electrical engineering and in the field of artificial intelligence is the unreliability of physical implementations. A non-uniform range of resistance, low mass-production yield and high fault probability during operation are disadvantages of the current memristor technologies. In this article, the authors offer a solution for these problems with a circuit design, which consists of many memristors with a high operational variance that can form a more robust single memristor. The proposition is confirmed by physical device measurements, by gaining similar results as in previous simulations. These results can lead to more stable devices, which are a necessity for neuromorphic computation, artificial intelligence and neural network applications.",
"conclusion": "5. Conclusions Two new types of memristor networks have been introduced, which are able to emulate more reliable memristors. Measurements have been successfully carried out for both the previously presented networks and the new networks. The measurements provided new information about the macro-characteristics of memristor networks compared to the previous simulations. The increased switching speed of memristor networks should be further investigated. This solution can be used with existing devices to support the implementation of neuromorphic applications.",
"introduction": "1. Introduction Since the theoretical [ 1 ] and practical [ 2 ] discovery of memristors, they have been extensively studied [ 3 , 4 , 5 ] as elementary building blocks for artificial intelligence and neuromorphic computing applications. The expected properties of memristors for such applications are wide and analog resistance range, low variance of device parameters and high device stability during long-term operation. Research has been done [ 6 ] to find optimal materials that satisfy these expectations, but even then there are other possibilities to further increase the capabilities of memristors. In binary memory applications, three important properties should be considered. The first one is having two clearly distinguishable states and these state declarations should apply to every element in a memory array. The second one is having a fast switching speed between the states. To reach the performance of the current complementary metal–oxide–semiconductor (CMOS) technology’s RAM the switching speed should be less than 10 ns. The third one is cycle endurance, which is the number of write–erase cycles without permanent device failure. In crossbar-network applications, a certain amount of uniformity of the memristors is necessary. The programming voltage and current levels are the same for every element and thus one expects that they will behave similarly for the same input signals. In the case of ANN applications, more deviance could be tolerated, but many state devices are needed, so the memristors developed for binary or multi-state memory purposes will not be sufficient. The mass production of devices, which can reliably fulfill these requirements, is not trivial. If the production yield of single devices is less than 100 percent (as they are not functioning as memristors or they are outside of the accepted range of parameters), then they can also affect the access circuit and the encompassing parts of the neuromorphic system. If the production yield of single devices is less than 100 percent (as they are not functioning as memristors or they are outside of the accepted range of parameters), then they can also affect the access circuit and the encompassing parts of the neuromorphic system. In very large scale integration (VLSI) device manufacturing, it is often easier and tends to cause fewer faults to make the same device many times, and use it as a building block to emulate other devices, instead of creating fewer, but different devices [ 7 ]. The same approach can be applied to memristors, but one should take into consideration their special nonlinear behavior in the voltage–current domain. This idea is further supported by the fact that memristors as two-terminals, could be manufactured more easily on many layers on microchips [ 8 ] than transistors. However, with every extra layer, the probability of device defects could also increase. In order to maintain or even improve the virtual yield of the production, interconnected structures of the memristor network are proposed. These circuits and the presented measurement results provide a response to the above mentioned challenges. Our proposed circuit constructions can be efficiently implemented on microchips, stacking the memristors of the circuit on top of each other. If a decent multilayer production technology arises with memristors, the disadvantage of the usage of several layers for the implementation of a single layer of memristor would be neglectable. This paper is organized as follows: after the above problem proposal, the measurement environment is introduced and explanatory discussion is given about our circuitry. The third section contains the proposed circuits and the measurement results that are more detrimental to the yield. This circuitry effectively addresses the proposed task. In the fourth section, the results are summarized and analyzed. The article is closed with a brief summary of the results in the conclusion section.",
"discussion": "4. Discussion The checkerboard type of network was practically unable to switch its state significantly compared to other solutions. This was probably due to the limited number of parallel connections in the network, which produced less possible routes to open. Higher control voltage could change its state, but the risk of device damage increases with the increased after-switch current. Longer pulses could also help, but it makes the writing process slower. The results for the H-fractal type of network are very similar to the newly introduced network regarding the writing of the state into low resistance position. However, this type of network has problems with erasing the state into the OFF state and could get stuck at an in-between state. The ON and OFF resistance values of the network with twelve memristors are lower than the other networks due to the reduced number of serial layers. However, when one compares it to a single memristor, it has lower ON resistance value and higher OFF resistance, meaning the network is more sensitive to control signals than only one memristor. In other words, a pulse with the same voltage level could make a clearer distinction between the initial and after states. The previous simulation results suggested that the switching speed could decrease using memristor grids. Surprisingly, the switching speed did not decrease, but increased instead. The networks are approximately three times faster than a single memristor. This is fairly unexpected, as the control voltage stayed constant in both measurements, which means that the voltage on any single memristor in a network measurement had to be strictly lower than in the case of a single device measurement at any given time during measuring. One explanation of this phenomenon could be the following: under the threshold voltage, the device behaves as a very small capacitor. As the metal flows into the dielectric matter to build up the filament, the partially charged capacitor discharges, causing a short-time high-energy electric current burst. The other devices are sensitive to fast current changes and the filament forming is starting in them as well. It can be seen as a “domino effect” with the consecutive memristors. If any of the OFF state memristors in a series switches to the ON state, the rest will automatically switch as well immediately after. If any of the memristors which closes the source in the series, opens, the rest will automatically open immediately after. Based on the above presented measurements the following parameter values were acquired, presented in Table 1 . The resistance values are the average ON/OFF ratio values of the 50 cycle long measurement sequence. Another important feature of th networks to note is the stronger nanobattery effect [ 16 ]. This causes the visible shift of the zero current level after the erasing pulse. The nanobattery effect is undesired in most applications, but can be dealt with by an appropriate control voltage and timing. It can also be taken advantage of, in some scenarios."
} | 2,100 |
34469744 | null | s2 | 112 | {
"abstract": "Metabolic cross-feeding frequently underlies mutualistic relationships in natural microbial communities and is often exploited to assemble synthetic microbial consortia. We systematically identified all single-gene knockouts suitable for imposing cross-feeding in Escherichia coli and used this information to assemble syntrophic communities. Most strains benefiting from shared goods were dysfunctional in biosynthesis of amino acids, nucleotides, and vitamins or mutants in central carbon metabolism. We tested cross-feeding potency in 1,444 strain pairs and mapped the interaction network between all functional groups of mutants. This network revealed that auxotrophs for vitamins are optimal cooperators. Lastly, we monitored how assemblies composed of dozens of auxotrophs change over time and observed that they rapidly and repeatedly coalesced to seven strain consortia composed primarily from vitamin auxotrophs. The composition of emerging consortia suggests that they were stabilized by multiple cross-feeding interactions. We conclude that vitamins are ideal shared goods since they optimize consortium growth while still imposing member co-dependence."
} | 291 |
37283898 | PMC10239617 | pmc | 113 | {
"abstract": "Ocean warming and marine heatwaves induced by climate change are impacting coral reefs globally, leading to coral bleaching and mortality. Yet, coral resistance and resilience to warming are not uniform across reef sites and corals can show inter- and intraspecific variability. To understand changes in coral health and to elucidate mechanisms of coral thermal tolerance, baseline data on the dynamics of coral holobiont performance under non-stressed conditions are needed. We monitored the seasonal dynamics of algal symbionts (family Symbiodiniaceae) hosted by corals from a chronically warmed and thermally variable reef compared to a thermally stable reef in southern Taiwan over 15 months. We assessed the genera and photochemical efficiency of Symbiodiniaceae in three coral species: Acropora nana, Pocillopora acuta, and Porites lutea . Both Durusdinium and Cladocopium were present in all coral species at both reef sites across all seasons, but general trends in their detection (based on qPCR cycle) varied between sites and among species. Photochemical efficiency ( i.e. , maximum quantum yield; F v /F m ) was relatively similar between reef sites but differed consistently among species; no clear evidence of seasonal trends in F v /F m was found. Quantifying natural Symbiodiniaceae dynamics can help facilitate a more comprehensive interpretation of thermal tolerance response as well as plasticity potential of the coral holobiont.",
"conclusion": "Conclusions We did not detect clear evidence of seasonal trends in dominant Symbiodiniaceae genera or photochemical efficiency in our study species, rather differences were more apparent among sites with contrasting thermal regimes and among coral species—highlighting the importance of species-specific studies. Reef site patterns that we observed in Symbiodiniaceae genera detection, using a coarse presence/absence qPCR approach, merit more comprehensive investigation ( i.e., genera or species-level quantification) to better assess the influence of thermal regime on Symbiodiniaceae associations among coral hosts. Baseline seasonal data under non-stressed conditions are pertinent to improve our understanding of energy provision sources relevant to coral thermal tolerance. Identifying typical ranges of normal variability in coral and Symbiodiniaceae physiology, coupled with an appreciation of the role that species traits and reef characteristics play, will allow for a better understanding of coral holobiont resistance and resilience in a warming ocean.",
"introduction": "Introduction The health and persistence of corals in the Anthropocene are most highly threatened by climate change ( Hoegh-Guldberg et al., 2017 ; Hughes et al., 2017 ), in particular ocean warming and marine heatwaves ( Spalding & Brown, 2015 ; Leggat et al., 2019 ). Elevated seawater temperatures can lead to oxidative stress in corals ( Oakley & Davy, 2018 ), which prompts the expulsion of the coral’s symbiotic dinoflagellate algae (family Symbiodiniaceae), from which scleractinian ( i.e., reef-building) corals typically obtain most of their energy ( Yellowlees, Rees & Leggat, 2008 ; Van Oppen & Lough, 2018 ). This physiological response leads to coral bleaching and may result in mortality, which can have detrimental effects on reef ecosystems and coral-dependent communities ( e.g. , phase shifts to algal-dominated states ( Ostrander et al., 2000 ; Vaughan et al., 2021 ); homogenization of fish populations Richardson, Graham & Hoey, 2020 ). The effect of elevated temperature on corals, however, is not uniform among species and sites. Corals with ‘weedy’ life-history traits ( e.g. , Pocillopora spp . ), which typically have a brooding reproductive strategy and are recruitment pioneers, appear to fare better under elevated temperatures than ‘competitive’ corals ( e.g. , Acropora spp . ), which are fast-growing broadcast spawners, but neither performs as well as ‘stress-tolerant’ corals ( e.g. , massive Porites spp . ), which are slow-growing broadcast spawners ( Darling, McClanahan & Côté, 2013 ; Kubicek et al., 2019 ). In addition, reef site characteristics can also affect coral thermal tolerance ( Camp et al., 2018 ; Burt et al., 2020 ). Reef sites that have shown higher resistance to bleaching include some that experience chronic disturbance ( e.g. , Guest et al., 2016 ), a wide seasonal range of temperatures –including chronic high maximum temperatures ( e.g. , Riegl & Purkis, 2012 ) and/or highly variable thermal regimes ( e.g. , Wyatt et al., 2020 ). The primary mechanisms that allow some corals to persist and prosper are likely underpinned on a fundamental level by the capacity for effective and consistent energy acquisition. Here we focus on coral performance in relation to the genera and photochemical efficiency of the coral’s symbiotic algae. The symbiotic relationship between scleractinian corals and Symbiodiniaceae has allowed coral reefs to flourish in nutrient-poor marine environments due to the efficient transfer of photosynthetic products from the symbiont to the coral host ( Muscatine, 1990 ; Roth, 2014 ). Corals can associate with a range of Symbiodiniaceae genera, often more than one at a time ( Baker, 2003 ; Silverstein, Correa & Baker, 2012 ), and the genus of Symbiodiniaceae can influence the host’s bleaching resistance ( Berkelmans & van Oppen, 2006 ). For example, the genus Durusdinium (formerly known as clade D) has been associated with high bleaching resistance under both heat and cold stress ( Stat & Gates, 2011 ; Silverstein, Cunning & Baker, 2017 ), but typically at the cost of slower coral growth compared to the more thermally sensitive genus Cladocopium (formerly known as clade C) ( Jones & Berkelmans, 2010 ; LaJeunesse et al., 2018 ). Symbiodiniaceae genus composition can vary spatially, with corals in more thermally extreme sites generally being more likely to host relatively more Durusdinium than other genera ( Oliver & Palumbi, 2009 ; Keshavmurthy et al., 2012 ; but see Howells et al., 2020 for an example of thermally tolerant Cladocopium ). Further, species within a genus of Symbiodiniaceae can also show different responses to high temperature, which can vary based on coral host-algal symbiont pairing ( Hoadley et al., 2019 ). In general, a challenge to pinpointing potential mechanisms of thermal tolerance in corals is the lack of long-term, in situ physiological/molecular data under unstressed conditions ( e.g. , relatively poor understanding of seasonal trends and natural fluctuations). Such data are needed to put monitoring studies ( e.g. , pre- and post-bleaching surveys) and short-term laboratory experiments in context. This is especially pertinent when assessing the potential role of coral traits and reef site characteristics ( e.g. , temperature regime) on the acquisition of thermal resistance and resilience. We assessed the seasonal dynamics of algal symbionts hosted by corals from a chronically warmed and thermally variable reef compared to a thermally stable reef in southern Taiwan over 15 months. We examined the dynamics of Symbiodiniaceae genus associations and photochemical efficiency across three common coral species with different life-history traits: Acropora nana , Pocillopora acuta , and Porites lutea , at both sites. We predicted that in the absence of large thermal anomalies, Symbiodiniaceae genus/genera associations would remain stable across seasons but would show site- and species-specific differences due to the distinct thermal regimes of the reef sites and species life-histories, respectively. We predicted that photochemical efficiency would vary across seasons, but that seasonal trends among species would be similar.",
"discussion": "Discussion Our comparison of algal symbiont dynamics in A. nana , P. acuta , and P. lutea colonies at a warmed and thermally variable reef (Outlet reef) and a thermally stable reef (Wanlitung reef) showed site- and species-specific patterns. In general, Durusdinium was detected earlier in the qPCR cycle than Cladocopium in colonies from the warmed and thermally variable reef than at the thermally stable reef. Symbiodiniaceae associations remained relatively consistent across seasons. Reef site patterns in Symbiodiniaceae associations were not clearly mirrored by similar site patterns in photochemical efficiency. Intraspecific F v /F m differed between sites in only one or two seasons for each coral species, with corals from the warmed and thermally variable reef periodically having higher F v /F m . Interspecific differences showed similar patterns at both sites, with A. nana and P. acuta typically having higher F v /F m than P. lutea . Although there was some subtle variation in F v /F m over time for each species at each site, no consistent seasonal trends in F v /F m were observed. Temperature and nutrients: reef site comparison The main physical difference between the two reef sites, based on the parameters assessed in this study, was the thermal regime ( Fig. 1 ). Outlet reef had higher maximum temperatures, particularly in summer months, than Wanlitung reef as the former is chronically influenced by warm-water effluent from an adjacent nuclear power plant. However, Outlet reef also had lower daily temperature minima due to cold-water upwelling in Nanwan Bay ( Hsu et al., 2020 ). As upwelling somewhat mitigated the power plant warming, mean daily temperature at Outlet reef was higher than at Wanlitung reef only in winter and spring. Corals that experience high variability or extremes in temperature have been shown to have increased thermal tolerance owing to acclimatization and/or adaptation ( e.g. , genetic adaptations ( Barshis et al., 2013 ), morphological adaptations ( Enríquez et al., 2017 ), hosting thermally tolerant algal symbionts ( Oliver & Palumbi, 2011 ); but also see Le Nohaïc et al., 2017 ; Smith et al., 2017 ; Klepac & Barshis, 2020 for limits on adaptation in corals from variable/extreme reefs). In contrast, reefs with relatively stable thermal regimes tend to have corals that are more susceptible to bleaching under elevated temperatures ( Thomas et al., 2018 ; Safaie et al., 2018 ). In contrast to temperature regime patterns, nutrient concentrations at both sites were similar across the study period ( Table S5 ). Levels of BOD 5 , NO 3 − , NO 2 − , NO 3 − , NH 3, and PO 4 3− measured across seasons did not differ between Outlet and Wanlitung reefs and were relatively low in comparison to other reefs impacted by anthropogenic stressors in southern Taiwan ( Meng et al., 2008 ; Liu et al., 2012 ). Performance patterns of corals between sites are therefore more likely to be attributable to differences in thermal regime than to nutrient levels. Symbiodiniaceae dynamics We found differing patterns in Symbiodiniaceae genera detection between our thermally distinct reef sites ( Figs. 2 – 4 ). Each of the three coral species showed earlier detection (based on qPCR cycle) of the more thermally tolerant Durusdinium algal symbiont at the warmed and thermally variable reef than at the thermally stable reef. The presence/absence (and also the proportional composition) of Symbiodiniaceae genera in coral tissues can change after stress events ( e.g. , bleaching) with more thermally tolerant genera typically replacing thermally sensitive ones ( Jones et al., 2008 ; Cunning, Silverstein & Baker, 2015 , but see Kao et al., 2018 ; Rouzé et al., 2019 , for examples of limited shuffling). It is probable that the chronic warming influence, and associated high summer maximum temperatures, at Outlet reef have resulted in a shift to corals potentially hosting more thermally tolerant Symbiodiniaceae. Indeed, 16 coral genera (including the three genera considered in our study) have been found to predominantly associate with either Durusdinium or a combination of Cladocopium and Durusdinium at Outlet reef ( Keshavmurthy et al., 2014 ). In contrast, the same genera at nearby reefs not influenced by the nuclear power plant, and corals deeper than 7 m at Outlet reef and hence out of range of the warm effluent, predominantly host Cladocopium ( Keshavmurthy et al., 2014 ). Potentially associating with more Durusdinium may allow corals at Outlet reef to resist bleaching under high summer temperature maxima and/or facilitate their capacity to recover from bleaching ( Silverstein, Cunning & Baker, 2015 ). Corals at Wanlitung reef, like many Indo-Pacific coral species that live under more stable conditions, may tend to preferentially host Cladocopium because it is an abundant and species-rich Symbiodiniaceae ( LaJeunesse et al., 2018 ) that is not associated with the energetic trade-offs of hosting the more thermally tolerant Durusdinium ( e.g. , Jones & Berkelmans, 2011 ). We did not explicitly quantify dominant or background levels of Symbiodiniaceae genera in this study, but these observed patterns in qPCR detection would likely benefit from a deeper examination ( e.g. , moving beyond genus-level to species-level assessment is increasingly viewed as important for improving our understanding of coral-algal symbiont dynamics; see Davies et al., 2023 ). In general, Symbiodiniaceae genera associations remained relatively stable over time for each species and reef site (see also Epstein et al., 2019 ). This seasonal stability reflects the fact that temperatures remained relatively moderate ( i.e., within typical site-specific seasonal ranges) throughout our study, with no heatwaves or summer mass bleaching observed. Symbiodiniaceae genus fidelity may also be attributed to acclimatization/adaptation to reef site characteristics ( e.g. , Iglesias-Prieto et al., 2004 ; Howells et al., 2020 ), long-standing co-evolution of host and symbionts ( Thornhill et al., 2014 ; Turnham et al., 2021 ), and/or to a species-specific ‘Symbiodiniaceae signature’ to the host colony ( Rouzé et al., 2019 ). Photochemical efficiency dynamics We examined algal symbiont photochemical efficiency dynamics by tracking F v /F m across coral species and seasons at both reef sites ( Fig. 5 ). Values of F v /F m usually decrease in response to elevated temperature ( Jones et al., 2000 ; Okamoto et al., 2005 ; Silverstein, Cunning & Baker, 2015 ) and can vary seasonally under non-stressed conditions ( Warner et al., 2002 ), often as a response to annual fluctuation in solar irradiance ( Winters, Loya & Beer, 2006 ). However, we found few intraspecific differences in F v /F m between sites, and inconsistent seasonal trends in F v /F m among species. Occasional differences between reef sites and seasons were also found by Carballo-Bolaños et al. (2019) who showed that F v /F m in the brain coral Leptoria phrygia was lower at Wanlitung than Outlet reef (in summer and winter, but not spring), and seasonally variable at Wanlitung reef but not at Outlet reef. In our study, typical seasonal differences in temperature and light may not have been enough to elicit significant changes in F v /F m for our coral species. Future investigation into other photochemical metrics may be useful to better assess photosystem dynamics ( e.g. , Ragni et al., 2010 ) . Alternatively, clear patterns in F v /F m seasonality may have been masked by subtle seasonal fluctuations in Symbiodiniaceae composition, i.e., due to differential photochemical performance across genera ( Kemp et al., 2014 ). We did nevertheless observe relatively consistent interspecific differences in F v /F m . At our sites, during each season, the F v /F m for all coral species was within a typical healthy range for coral species in southern Taiwan ( Putnam, Edmunds & Fan, 2010 ; Mayfield, Fan & Chen, 2013 ; Carballo-Bolaños et al., 2019 ), but A. nana and P. acuta had higher F v /F m than P. lutea in most seasons. Differences in F v /F m among species are not uncommon ( e.g. , higher F v /F m in unstressed Pocillopora meandrina compared to Porites rus ( Putnam & Edmunds, 2008 ); wider thermal breadth in Porites cylindrica than Acropora valenciennesi ( Jurriaans & Hoogenboom, 2020 )). This is likely a result of taxon-specific traits, such as coral tissue thickness ( Anthony & Hoegh-Guldberg, 2003 ), algal symbiont position within the coral tissue ( Edmunds, Putnam & Gates, 2012 ), and/or genus-specific symbiont associations ( Wang et al., 2012 ; Yuyama et al., 2016 )."
} | 4,115 |
30240646 | PMC6123858 | pmc | 114 | {
"abstract": "The success of deep networks and recent industry involvement in brain-inspired computing is igniting a widespread interest in neuromorphic hardware that emulates the biological processes of the brain on an electronic substrate. This review explores interdisciplinary approaches anchored in machine learning theory that enable the applicability of neuromorphic technologies to real-world, human-centric tasks. We find that (1) recent work in binary deep networks and approximate gradient descent learning are strikingly compatible with a neuromorphic substrate; (2) where real-time adaptability and autonomy are necessary, neuromorphic technologies can achieve significant advantages over main-stream ones; and (3) challenges in memory technologies, compounded by a tradition of bottom-up approaches in the field, block the road to major breakthroughs. We suggest that a neuromorphic learning framework, tuned specifically for the spatial and temporal constraints of the neuromorphic substrate, will help guiding hardware algorithm co-design and deploying neuromorphic hardware for proactive learning of real-world data.",
"conclusion": "Concluding Remarks As Carver Mead suggested decades ago, neuromorphic engineering and its applications are held back by our limited understanding of neural circuits' organizing principles, and less so by difficulties in implementation. Interestingly, the foundations of current deep neural networks were laid out during the same period. But at that time, computers could not scale to real problems such as visual recognition and natural language processing tasks. Today, many machine learning and neural networks can solve some of those problems ( LeCun et al., 2015 ). These successes provide renewed interest in taking inspiration from machine learning to guide our understanding of the organizing principles in the brain and applying them to neuromorphic hardware. Researchers at the interface of these fields highlight the possible benefits in cross-fertilizing machine learning and neuroscience ( Hassabis et al., 2017 , Lake et al., 2017 ), in spite of a strong cultural gap between the two fields. This cultural gap is understandable: brain-inspired models, especially those based on spiking neuron models, severely restrict the breadth of computations during learning and inference. With the advent of powerful graphical processing units and dedicated machine learning accelerators, the brain-inspired approach to learning machines is often heavily criticized as being misguided. These criticisms are relevant to bottom-up designs, or metrics purely based on absolute accuracy at standardized benchmarks. However, in cases in which real-time adaptability, autonomy, or privacy is essential, learning must be performed closer to the sensors. In this situation, power becomes a key metric and neuromorphic hardware co-designed with learning algorithms can have significant advantages. Although the traditionally bottom-up approach of neuromorphic engineering is well justified for analysis-by-synthesis research, widespread interest in this field will likely be driven by its technological and economic prospects. In this review, we argued that the technological success of neuromorphic learning machines is not compatible with a purely bottom-up approach using current technologies. With the close analogies with machine learning, we hope to encourage aspiring and experienced neuromorphic hardware engineers to carefully plan learning features by matching neural and synaptic dynamics under task relevant objective functions. Although several challenges remain open in learning with spiking neurons and a software platform for programming neuromorphic learning machine does not yet exist, we expect that matching neural and synaptic dynamics will go a long way in rendering hardware efforts compatible with ongoing algorithmic developments.",
"introduction": "Introduction The harnessing of future big data for societal and economical advances demands an unprecedented amount of computing resources. The difficulties in scaling current computing technologies to meet such demands, combined with a looming end of Moore's law, is spurring widespread interest in novel scalable computing paradigms. One such paradigm is neuromorphic engineering, which strives to reproduce in hardware the brain's cognitive and adaptive abilities by mimicking its architectural and dynamical properties ( Mead, 1990 ). The adaptivity, efficiency, and largely unsurpassed performance of the brain at solving complex cognitive tasks has been a continuing inspiration for designing computing systems ( von Neumann, 1958 ). Although the reasons for the extraordinary robustness, efficiency, and adaptivity of brains are puzzling, their style of computation, supported by massively parallel and self-organizing neural architectures that are fundamentally different from that used in conventional computers, is believed to be a key piece of the puzzle ( Douglas and Martin, 2004 ). The foundational insight of neuromorphic engineering is that the current-voltage dependence in ion channels and transistors operated in the sub-threshold regime are both exponential ( Mead, 1990 ), owing to the same diffusion law governing the transport of their respective carriers. This similarity implies that electronic and biological substrates share constraints on communication, power, and reliability. Thus, neuromorphic hardware designed along these principles has the potential to translate advances in neuroscience research into ultra-low-power computing technologies targeted at producing cognitive function ( Indiveri and Liu, 2015 ). This hardware can in turn be employed as a tool to investigate the organizational principles of the brain by accelerating existing neural simulations ( Zenke and Gerstner, 2014 ) or by analyzing the qualities of the constructed hardware ( Cauwenberghs, 2013 ). Since its inception in the early 1990s, the interest in neuromorphic engineering is rising rapidly ( Schuman et al., 2017 ), and neuromorphic engineering now extends to a wide gamut of software and hardware efforts ( Schuman et al., 2017 ) dedicated at simulating or emulating neural network dynamics. Efforts in neuromorphic engineering resulted in many successful devices ( Indiveri et al., 2011 ). These range from mixed signal ( Benjamin et al., 2014 , Chicca et al., 2013 , Park et al., 2014 , Schemmel et al., 2010 ) systems that emulate the dynamics of spiking neural network models in very-large-scale integration (VLSI) to digital systems ( Merolla et al., 2014 , Davies et al., 2018 , Furber et al., 2014 ) dedicated at simulating the dynamics of spiking networks on a dedicated digital architecture. Neuromorphic systems have been successfully demonstrated in pattern recognition, decision-making, and navigation tasks ( Qiao et al., 2015 , Srinivasa and Cho, 2014 , Neftci et al., 2013 , Serrano-Gotarredona et al., 2009 , Schmuker et al., 2014 , Esser et al., 2016 , Moradi et al., 2018 , Blum et al., 2017 ). Recently, the neuromorphic engineering community has started to dedicate significant effort in embedding synaptic plasticity in their hardware for emulating the adaptive capabilities of the brain ( Azghadi et al., 2014 ). At the system level, neurons in the adult brain communicate principally through sparse, all-or-none events in continuous time ( Gerstner and Kistler, 2002 ). All other internal states such as neurotransmitter concentrations, synaptic states, and membrane potentials are local to the neuron. Such an architecture is highly scalable, thanks to sparse interprocess communication. However, harnessing neuromorphic hardware to solve real-world problems in a reliable fashion proved to be extremely challenging. This is because they require computational strategies that can operate robustly on local information and sparse global communication. Understanding the mechanisms of brain function and devising models and algorithms that operate under such conditions is the key endeavor of computational neuroscience modeling ( Sompolinsky, 2014 ). Although technologies for imaging the brain and analyzing the resulting data are progressing rapidly, the understanding of its organizing principles is still largely incomplete. Our limited understanding of which brain mechanisms are necessary to achieve cognitive function weakens the technological prospects of the traditional bottom-up “brain-as-a-blueprint” approach to neuromorphic engineering. On the other hand, machine learning and deep learning provide relatively well-understood principles for solving problems of practical interest, with the caveat that most state-of-the-art machine learning algorithms rely on information that is not local to the computational building blocks of a neural substrate. In this article, we introduce neuromorphic learning machines as a middle-ground solution between the bottom-up and top-down approaches by reconciling the architecture and dynamics of a neural substrate with the organizing principles of machine learning. This “middle-in” approach is consistent in spirit to Marr's line of inquiry, which strives to study a problem at the levels of theory, algorithm, and hardware ( Marr, 1982 ). As such, our discussion will apply to the more modern and general sense of the term neuromorphic, i.e., that the machines compute with neuron-like units using local information. We will explore the benefits of viewing neuromorphic engineering through the lens of recent advances in artificial neural network and machine learning, i.e., to which extent will these algorithms guide us in neuromorphic hardware design, and what advantages would accrue from such hardware? Through this discussion, we aim to dispel some of the perceived differences and similarities between biologically inspired neural networks and artificial neural networks and provide engineers guidelines for increasing the technological impact of their neuromorphic hardware. In so doing, this review will outline the nature of possible bridges from neurobiology to machine learning and describe modern tools for investigating such bridges."
} | 2,528 |
28406173 | PMC5390294 | pmc | 115 | {
"abstract": "Collective behavior emerging out of self-organization is one of the most striking properties of an animal group. Typically, it is hypothesized that each individual in an animal group tends to align its direction of motion with those of its neighbors. Most previous models for collective behavior assume an explicit alignment rule, by which an agent matches its velocity with that of neighbors in a certain neighborhood, to reproduce a collective order pattern by simple interactions. Recent empirical studies, however, suggest that there is no evidence for explicit matching of velocity, and that collective polarization arises from interactions other than those that follow the explicit alignment rule. We here propose a new lattice-based computational model that does not incorporate the explicit alignment rule but is based instead on mutual anticipation and asynchronous updating. Moreover, we show that this model can realize densely collective motion with high polarity. Furthermore, we focus on the behavior of a pair of individuals, and find that the turning response is drastically changed depending on the distance between two individuals rather than the relative heading, and is consistent with the empirical observations. Therefore, the present results suggest that our approach provides an alternative model for collective behavior.",
"discussion": "Discussion This study presents the first model analysis reproducing the results of a real animal group’s experiment, investigating dynamic individual-level interactions 13 . The experiment sheds light on how the effective social force among individuals depends on their neighbor and/or velocity. To capture the direct behavioral response, the experimenters observed pairwise interactions in the two individuals group. In particular, real animals change their direction depending on their neighbors’ positions rather than the relative angle between their velocities. Nevertheless, many models for collective behavior assumed the alignment rule, by which individuals match their velocity with those of their neighbors. Consequently, these previous models predicted that the direction would increase at a larger relative angle. Furthermore, it is still not clear how this discrepancy between model simulation and experimental results are influenced by the interactions, other than those that follow the explicit alignment rule, between individuals within the group. Individuals in our model do not follow the explicit alignment rule but act based on mutual anticipation. Each individual in our model has its own principal vector and has potential transitions derived from the principal vector. Through mutual anticipation, the individual moves to a site where potential transitions among individuals are concentrated. Then, the principal vector of individual is adjusted based on the anticipated moves of its neighbors. Hence, although the explicit alignment rule does not apply, interactions with neighbors are reflected in the individual’s decision. As shown in Fig. 3 , individuals whose positions and principal vectors are distributed at random in the initial condition can form a coherent swarm with high polarity. As an emerging dynamic group-level property, our model can reproduce the social dynamics forces of individuals observed in a real animal group 13 . To investigate the social force dynamics, we consider the acceleration of the focal individual as the total force and decompose it into a turning component and a speeding component. If individuals follow the alignment rule, the turning force would increase at a larger relative angle between their velocities. In our model, however, we observe that the strength of the turning force increases depending on how far the neighbor is to the right or left from the focal individual, rather than the relative angle, and is consistent with empirical data. Let us consider simple situations. If a neighbor is far from the focal individual on its sides, then the distance left-right is large, and their potential transitions can overlap only at an intermediate area between them, usually not in front of them. Accordingly, the individuals would move toward the area by mutual anticipation, resulting in a large turning force and a large coherence. In contrast, when the distance left-right is close to zero, the potential overlapping transitions cover a larger area, usually including the area in front of the individuals. Consequently, the individuals can move forward in various ways, resulting in approximately zero average turning force. Moreover, the speeding force increases depending on how far the neighbor is in front of or behind the focal individual rather than the relative angle; this is consistent with the empirical data. Such speeding force dynamics cannot be simulated by models with constant speed. However, models considering mutual anticipation between individuals in different configurations can simulate speeding force dynamics in the same manner as the turning force. In this study, we demonstrate emergent group-level properties and social force dynamics due to mutual anticipation and not alignment between individuals in a simulated group. This model was inspired by the observation of soldier crab swarms, although similar patterns are clearly present in human crowds 34 . These results suggest that our approach could provide an alternative model for collective behavior. Finally, we note that there is a possibility that our swarm model defined in a lattice space will lead features not present in the real animal groups. While in our model the direction of individual’s velocity (i.e., its transition in lattice space) cannot change continuously but do discretely, real animals must change their direction continuously. This difference between real animal group and our lattice model could be important when comparing it with the difference between continuous system (such as Heisenberg spin model) and discrete system (such as Ising spin model) in ferromagnets with respect to a continuous symmetry (in this case, rotational invariance) 44 45 46 47 . Were this an equilibrium problem as in spin system, the explicit symmetry breaking by discrete change of the direction (as in Ising spin system) could lead to an orientationally ordered state which could not occur in the continuous symmetry system (e.g., Heisenberg spin system), as implied by the Mermin-Wagner theorem. However, it was revealed that ordered state is possible in non-equilibrium moving systems (e.g., swarm systems) 44 . The fundamental difference between swarm systems and spin systems is that individuals in swarm systems, unlike spins, move with respect to another, so that the interaction network is not fixed in time and individuals change their neighbors perpetually 14 15 . Hydrodynamic theories of flocking have hypothesized that this moving mechanism reinforces correlations between individuals, enhancing global ordering 45 46 47 . Considering the above, while moving swarm systems (of both real animal group and our model) are non-equilibrium, there is a possibility that our discrete model will lead features with respect to ordered state absent in the real animal group. In order to investigate this issue, we will probe continuous space version of our model more sensitively than present discrete version in the future work."
} | 1,834 |
25972778 | PMC4413675 | pmc | 116 | {
"abstract": "Implementing compact, low-power artificial neural processing systems with real-time on-line learning abilities is still an open challenge. In this paper we present a full-custom mixed-signal VLSI device with neuromorphic learning circuits that emulate the biophysics of real spiking neurons and dynamic synapses for exploring the properties of computational neuroscience models and for building brain-inspired computing systems. The proposed architecture allows the on-chip configuration of a wide range of network connectivities, including recurrent and deep networks, with short-term and long-term plasticity. The device comprises 128 K analog synapse and 256 neuron circuits with biologically plausible dynamics and bi-stable spike-based plasticity mechanisms that endow it with on-line learning abilities. In addition to the analog circuits, the device comprises also asynchronous digital logic circuits for setting different synapse and neuron properties as well as different network configurations. This prototype device, fabricated using a 180 nm 1P6M CMOS process, occupies an area of 51.4 mm 2 , and consumes approximately 4 mW for typical experiments, for example involving attractor networks. Here we describe the details of the overall architecture and of the individual circuits and present experimental results that showcase its potential. By supporting a wide range of cortical-like computational modules comprising plasticity mechanisms, this device will enable the realization of intelligent autonomous systems with on-line learning capabilities.",
"conclusion": "5. Conclusions We presented a mixed-signal analog/digital VLSI device for implementing on-line learning spiking neural network architectures with biophysically realistic neuromorphic circuits such as STP synapses, LTP synapses and low-power, low-mismatch adaptive I&F silicon neurons. The proposed architecture exploits digital configuration latches in each synapse and neuron element to guarantee a highly flexible infrastructure for programming, with the same device, diverse spiking neural network architectures. All the operations of the chip are achieved via asynchronous AE streams. These operations include sending events to the chip, configuring the topology of the neuron network, probing internal variables, as well as programming internal properties of synapse and neurons. The parameters for different synapse and neuron behaviors can be fine tuned by programming the temperature-compensated on-chip BG. The ROLLS neuromorphic processor can be used to carry out basic research in computational neuroscience and can be exploited for developing application solutions for practical tasks. In particular, this architecture has been developed to study spike-based adaptation and plasticity mechanism and to use its ability to carry out on-chip on-line learning for solving tasks that require the system to adapt to the changes in its input signals and in the environment it interacts with. Conflict of interest statement The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.",
"introduction": "1. Introduction Recent advances in neural network modeling and theory, combined with advances in technology and computing power, are producing impressive results in a wide range of application domains. For example, large-scale deep-belief neural networks and convolutional networks now represent the state-of-the-art for speech recognition and image segmentation applications (Mohamed et al., 2012 ; Farabet et al., 2013 ). However, the mostly sequential and synchronous clocked nature of conventional computing platforms is not optimally suited for the implementation of these types of massively parallel neural network architectures. For this reason a new generation of custom neuro-computing hardware systems started to emerge. These systems are typically composed of custom Very Large Scale Integration (VLSI) chips that either contain digital processing cores with dedicated memory structures and communication schemes optimized for spiking neural networks architectures (Wang et al., 2013 ; Furber et al., 2014 ; Neil and Liu, 2014 ), or full-custom digital circuit solutions that implement large arrays of spiking neurons with programmable synaptic connections (Merolla et al., 2014 ). While these devices and systems have high potential for solving machine learning tasks and applied research problems, they do not emulate directly the dynamics of real neural systems. At the other end of the spectrum, neuromorphic engineering researchers have been developing hardware implementations of detailed neural models, using mixed signal analog-digital circuits to reproduce faithfully neural and synaptic dynamics, in a basic research effort to understand the principles of neural computation in physical hardware systems (Douglas et al., 1995 ; Liu et al., 2002 ; Chicca et al., 2014 ). By studying the physics of computation of neural systems, and reproducing it through the physics of transistors biased in the subthreshold regime (Liu et al., 2002 ), neuromorphic engineering seeks to emulate biological neural computing systems efficiently, using the least amount of power and silicon real-estate possible. Examples of biophysically realistic neural electronic circuits built following this approach range from models of single neurons (Mahowald and Douglas, 1991 ; Farquhar and Hasler, 2005 ; Hynna and Boahen, 2007 ; van Schaik et al., 2010 ), to models of synaptic dynamics (Liu, 2003 ; Bartolozzi and Indiveri, 2007a ; Xu et al., 2007 ), to auditory/visual sensory systems (Sarpeshkar et al., 1996 ; van Schaik and Meddis, 1999 ; Zaghloul and Boahen, 2004 ; Costas-Santos et al., 2007 ; Liu and Delbruck, 2010 ), to reconfigurable spiking neural network architectures with learning and plasticity (Giulioni et al., 2008 ; Hsieh and Tang, 2012 ; Ramakrishnan et al., 2012 ; Yu et al., 2012 ; Chicca et al., 2014 ). In this paper we propose to combine the basic research efforts with the applied research ones, by presenting a VLSI architecture that can be used to both carry out research experiments in computational neuroscience, and to develop application solutions for practical tasks. The architecture proposed comprises electronic neuromorphic circuits that directly emulate the physics of real neurons and synapses to faithfully reproduce their adaptive and dynamic behavior, together with digital logic circuits that can set both the properties of the individual synapse and neuron elements as well as the topology of the neural network. In particular, this architecture has been developed to implement spike-based adaptation and plasticity mechanisms, and to carry out on-chip on-line learning for tasks that require the system to adapt to the changes in the environment it interacts with. Given these characteristics, including the ability to arbitrarily reconfigure the network topology also at run-time, we named this device the Reconfigurable On-line Learning Spiking Neuromorphic Processor (ROLLS neuromorphic processor). The main novelty of the work proposed, compared to previous analogous approaches (Indiveri et al., 2006 ; Giulioni et al., 2008 ; Ramakrishnan et al., 2012 ; Yu et al., 2012 ) consists in the integration of analog bi-stable learning synapse circuits with asynchronous digital logic cells and in the embedding of these mixed-signal blocks in a large multi-neuron architecture. The combination of analog and digital circuits, with both analog and digital memory elements, within the same block provides the device with an important set of programmable features, including the ability to configure arbitrary network connectivity schemes. At the analog circuit design level, we present improvements in the neuron and spike-based learning synapses over previously proposed ones (Indiveri et al., 2011 ; Chicca et al., 2014 ), which extend their range of behaviors and significantly reduce device mismatch effects. At the system application level we demonstrate, for the first time, both computational neuroscience models of attractor networks and image classification neural networks implemented exclusively on custom mixed-signal analog-digital neuromorphic hardware, with no extra pre- or post-processing done in software. In the next section we describe the ROLLS neuromorphic processor system-level block diagram, highlighting its dynamic and spike-based learning features. In Section 2.2 we describe in detail the circuits that are present in each building block, and in Section 3 we present system level experimental results showcasing examples of both computational neuroscience models and machine vision pattern recognition tasks. Finally, in Sections 4, 5 we discuss the results obtained and summarize our contribution with concluding remarks.",
"discussion": "4. Discussion Unlike conventional von Neumann processors that carry out bit-precise processing and access and store data in a physically separate memory block, the ROLLS neuromorphic processor uses elements in which memory and computation are co-localized. The computing paradigm implemented by these types of neuromorphic processors does not allow for the virtualization of time, with the transfer of partial results back and forth between the computing units and physically separate memory banks at high speeds. Instead, their synapse and neuron circuits process input spikes on demand as they arrive, and produce their output responses in real-time. Consequently, the time constants of the synapses and neurons present in these devices need to be well-matched to the signals the system is designed to process. For the case of real-time behaving systems that must interact with the environment, while processing natural signals in real-time, these time constants turn out to be compatible with the biologically plausible ones that we designed into the ROLLS neuromorphic processor. As we implemented non-linear operations in each synapse (such as short-term depression or long-term plasticity), it is not possible to time-multiplex linear circuits to reduce the area occupied by the synaptic matrix array. As a consequence, our device is essentially a large memory chip with dedicated circuits for each synapse that act both as memory elements and computing ones. This approach is complementary to other recent ones that focus on accelerated neural simulations (Bruederle et al., 2011 ), or that target the real-time emulation of large populations of neurons but with no on-chip learning or adaptive behaviors at the synapse level (Benjamin et al., 2014 ). The device we describe here is ideal for processing sensory signals produced by neuromorphic sensors (Liu and Delbruck, 2010 ) and building autonomous behaving agents. The system level examples demonstrated in Section 3 show how this can be achieved in practice: the hardware attractor network experiment focuses on the idea that the functional units of the cortex are subset of neurons that are repeatedly active together and shows that such units have the capability of storing state-dependent information; the pattern classification experiment demonstrates how it is possible to implement relatively complex sensory processing tasks using event-based neuromorphic sensors. Our results demonstrate the high-degree of programmability of our device as well as its usability in typical application domains. Its properties make it an ideal tool for exploring computational principles of spiking systems consisting of both spiking sensors and cortical-like processing units. This type of tools are an essential resource for understanding how to leverage the physical properties of the electronic substrate as well as the most robust theories of neural computation in light of the design of a new generation of cortex-like processors for real-world applications. The multi-chip system is supported by the use of a newly developed software front-end, PyNCS, which allows rapid integration of heterogeneous spiking neuromorphic devices in unique hardware infrastructure and continuous online monitoring and interaction with the system during execution (Stefanini et al., 2014 ). In order to integrate the DVS and ROLLS in the existing software and hardware infrastructure, it was necessary to list the address specifications for the spiking events and for the configuration events in Neuromorphic Hardware Mark-up Language (NHML) files, the neuromorphic mark-up language used by PyNCS to control the neuromorphic system. The potential of the approach proposed in this work for building intelligent autonomous systems is extremely high, as we develop brain-inspired computing devices embedded with learning capabilities that can interact with the environment in real time. Substantial progress has already been made in the theoretical domain (Schöner, 2007 ; Rutishauser and Douglas, 2009 ), and preliminary results have already been demonstrated also with neuromorphic cognitive systems (Neftci et al., 2013 ) synthesized by the user. The ROLLS neuromorphic processor described in this work can therefore contribute to extending the current state-of-the-art by providing also adaptation and learning mechanisms that could allow these systems to learn the appropriate network properties to implement autonomous cognitive systems."
} | 3,343 |
37759917 | PMC10526461 | pmc | 117 | {
"abstract": "We examine the challenging “marriage” between computational efficiency and biological plausibility—A crucial node in the domain of spiking neural networks at the intersection of neuroscience, artificial intelligence, and robotics. Through a transdisciplinary review, we retrace the historical and most recent constraining influences that these parallel fields have exerted on descriptive analysis of the brain, construction of predictive brain models, and ultimately, the embodiment of neural networks in an enacted robotic agent. We study models of Spiking Neural Networks (SNN) as the central means enabling autonomous and intelligent behaviors in biological systems. We then provide a critical comparison of the available hardware and software to emulate SNNs for investigating biological entities and their application on artificial systems. Neuromorphics is identified as a promising tool to embody SNNs in real physical systems and different neuromorphic chips are compared. The concepts required for describing SNNs are dissected and contextualized in the new no man’s land between cognitive neuroscience and artificial intelligence. Although there are recent reviews on the application of neuromorphic computing in various modules of the guidance, navigation, and control of robotic systems, the focus of this paper is more on closing the cognition loop in SNN-embodied robotics. We argue that biologically viable spiking neuronal models used for electroencephalogram signals are excellent candidates for furthering our knowledge of the explainability of SNNs. We complete our survey by reviewing different robotic modules that can benefit from neuromorphic hardware, e.g., perception (with a focus on vision), localization, and cognition. We conclude that the tradeoff between symbolic computational power and biological plausibility of hardware can be best addressed by neuromorphics, whose presence in neurorobotics provides an accountable empirical testbench for investigating synthetic and natural embodied cognition. We argue this is where both theoretical and empirical future work should converge in multidisciplinary efforts involving neuroscience, artificial intelligence, and robotics.",
"conclusion": "6. Conclusions A key research question lying at the intersection of neuroscience and AI concerns the degree of biological plausibility of the models used to capture cognition in artificial or living systems. Inevitably, due to their biomimicry roots, SNNs play a crucial role in any research aiming to either explain the neuroscientific behavior of biological specimens or provide insight into modeling their cognitive behavior. In this paper, we provided a brief tour of relevant topics to highlight the significance of SNNs in a wide variety of fields from the analysis of brain activities to embodied cognitive behavior of robotic systems. We visited the everlasting tension between the biological plausibility and computational efficiency of neural models and argued that this conflict may have been seized by the invention of neuromorphic hardware. We documented that the existing literature supports the use of SNNs for even traditional EEG signal processing (classification) tasks, which can still benefit from neural modeling of the processes generating the data, notably, to simplify the neural inverse problems. Utilizing these models for neuroprosthetic devices and their integration with brain-computer interfaces were also highlighted to bold the transformative impact that these technologies can have on restoring functionality and improving the lives of individuals with limited motor functions. Where there is a need for capturing the neuronal behavior of biological objects, e.g., in the classification of canonical neuronal models, the increased efficiency of information processing may indeed be taken as further evidence of the computational validity of biological models or as neurocognitive mimetic improvement for an artificial agent. Further, we argued that robotic agents are suitable platforms for testing hypotheses in the field of embodied cognition, due to the availability of the sensing, commanding, and processing information in a physical engineering system. Such a testbed should be also ideal for improving the intelligence, including adaptability and self-awareness, of the robots’ GNC systems. Therefore, it is here that SNNs offer a theoretically robust (leveraging the NFT and the SPA) and computationally efficient (leveraging sparse, distributed event-based information processing) means for studying embodied cognition and consciousness in robotics, addressing at the same time the problem of neurobiological explainability.",
"introduction": "1. Introduction Understanding how living organisms function within their surrounding environment reveals the key properties essential to designing and operating next-generation intelligent systems. We refer to a system as intelligent if it has the capability of fast adapting to the changes in its components, environment, or mission requirements. A central actor enabling organisms’ complex behavioral interactions with the environment is indeed their efficient information processing through neuronal circuitries. Although biological neurons transfer information in the form of trains of impulses or spikes, Artificial Neural Networks (ANNs) that are mostly employed in the current Artificial Intelligence (AI) practices do not necessarily admit a binary behavior. In addition to the obvious scientific curiosity of studying intelligent living organisms, the interest in the biologically plausible neuronal models, i.e., Spiking Neural Networks (SNNs) that can capture a model of organisms’ nervous systems, may be simply justified by their unparalleled energy/computational efficiency. Inspired by the spike-based activity of biological neuronal circuitries, several neuromorphic chips, such as Intel’s Loihi chip [ 1 ], have been emerging. These pieces of hardware process information through distributed spiking activities across analog neurons, which, to some extent, trivially assume biological plausibility; hence, they can inherently leverage neuroscientific advancements in understanding neuronal dynamics to improve our computational solutions for, e.g., robotics and AI applications. Due to their unique processing characteristics, neuromorphic chips fundamentally differ from traditional von Neumann computing machines, where a large energy/time cost is incurred whenever data must be moved from memory to processor and vice versa. On the other hand, coordinating parallel and asynchronous activities of a massive SNN on a neuromorphic chip to perform an algorithm, such as a real-time AI task, is non-trivial and requires new developments of software, compilers, and simulations of dynamical systems. The main advantages of such hardware are (i) computing architectures organized closely to known biological counterparts may offer energy efficiency, (ii) well-known theories in cognitive science can be leveraged to define high-level cognition for artificial systems, and (iii) robotic systems equipped with neuromorphic technologies can establish an empirical testbench to contribute to the research on embodied cognition and to explainability of SNNs. Particularly, this arranged marriage between biological plausibility and computational efficiency perceived from studying embodied human cognition has greatly influenced the field of AI. In any interdisciplinary research, where neural networks are used to construct entirely artificial systems or to simulate living intelligent systems, defining cognition from different disciplines’ points of view becomes a “wicked problem” [ 2 ]. For an engineer, it may be enough to include artificial neural systems capable of state-of-the-art task performance that are designed based on biological principles and are not necessarily faithful to in silico representations of neural circuitry in vivo. On the other hand, a neuroscientist may only be interested in seeking biologically-equivalent representations that are not necessarily computationally efficient. Both describe what they investigate as “cognition”, and both are indeed partially correct (we distinguish between artificial and biological cognition here) until a strong link is drawn from the in vivo observations to in silico realizations. A clear distinction can be observed when considering neuromorphic technologies, where better energy efficiency overshadows the inherent biological plausibility of what can be deemed cognition. Connecting the dots from in vivo to in silico neural systems requires not only an account of physical substrates dynamics supporting cognition, as observed in nature, but also their replication in viable neural computing architectures to reproduce such observations. This perspective provides a platform to test purely embodied versus purely functionalist computational hypotheses of mind: Are certain classes of biologically embodied computations (e.g., sensorimotor processing for perception and locomotion) inherent features of the physical substrates? Or, is the symbolic information processing universal to the substrates? Embodied analog-hardware cognition suggests that certain features of the mind are contingent upon the physical substrates (sensing organs and actuators) in which it is embodied. That is, mental processes are the exclusive province of the brain and they are naturally dependent on and constrained by the physical characteristics of the neural system through which their operations are implemented (the so-called neural dependency). Although symbolic computationalist theories of mind may also have roots in neural activities, they hold the belief that the mind is the consequence of what neural information processing does, i.e., its functionality, not the specific hardware [ 3 ]. Without being bogged down in the endless philosophical debate around the classic mind-body problem, we can pragmatically conclude that a dialectical synthesis of these tensions suggests that a limited class of substrates is capable of performing a limited class of computations and that these classes are constrained by the biological entity and task at hand [ 4 ]. Consistent with this pragmatic approach, the interdisciplinary link between neuroscience and AI may in fact be determining these reciprocal constraints. At this critical junction, robotic experimentation of embodied cognition developed predicated on SNNs becomes the necessary test bench [ 5 ]. Notably, many outstanding questions of the field then collapse into the following: (i) Do biological neuronal models that better reproduce observed empirical data better capture cognition (i.e., intelligent behaviors), when implemented on robotic agents?, and (ii) How do the embodiment of neural networks affect the network dynamics and topology, when interacting with the robot’s physical environment? This article examines some recent multidisciplinary developments in the use of SNNs for both natural and engineering sciences. We summarize and critically review new developments in three separate areas: (i) vision and image processing using Dynamic Vision Sensors (DVS), (ii) SNN-based analysis of the Electroencephalography (EEG), and (iii) robust spiking control systems applicable in robotic Guidance, Navigation, and Control (GNC). By connecting the dots from perception to signal processing for brain data to GNC systems for robotics, we highlight where SNNs may be used to further research goals across disciplines and how they may be constituent to a theory of embodied cognition. A survey of control based on learning-inspired spiking neural networks can be found in [ 6 ] where different learning rules are identified and classified for robotic applications. More recently, [ 7 ] exclusively reviews the application of neuromorphic computing for socially interactive robotics and [ 8 ] focuses on the implementation of neuromorphic hardware in robotics. However, comprehensive research on closing the cognition loop in neurorobotics is still missing in the literature, which is our main agenda in the current paper. The paper is structured as follows: Section 2 provides a deep description of SNN modeling, distinguishing between descriptive models and functional models. Section 3 discusses the bottlenecks in the computational simulation and emulation of SNNs. Section 4 discusses the embodiment of SNN into cognitive robotic systems and reviews multiple perspectives that SNNs can play a crucial role in developing an intelligent cognitive robot. It also explores the consequences of the technological advancements to future brain-inspired computational cognition research in relation to simulating adaptable and reconfigurable neurorobotic systems. Section 5 summarizes the trends in the literature and the limitations of the existing technologies. Some concluding remarks are included in Section 6 ."
} | 3,229 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.