Communication-efficient federated learning

Federated learning (FL) enables edge devices, such as Internet of Things devices (e.g., sensors), servers, and institutions (e.g., hospitals), to collaboratively train a machine learning (ML) model without sharing their private data. FL requires devices to exchange their ML parameters iteratively, and thus the time it requires to jointly learn a reliable model depends not only on the number of training steps but also on the ML parameter transmission time per step. In practice, FL parameter transmissions are often carried out by a multitude of participating devices over resource-limited communication networks, for example, wireless networks with limited bandwidth and power. Therefore, the repeated FL parameter transmission from edge devices induces a notable delay, which can be larger than the ML model training time by orders of magnitude. Hence, communication delay constitutes a major bottleneck in FL. Here, a communication-efficient FL framework is proposed to jointly improve the FL convergence time and the training loss. In this framework, a probabilistic device selection scheme is designed such that the devices that can significantly improve the convergence speed and training loss have higher probabilities of being selected for ML model transmission. To further reduce the FL convergence time, a quantization method is proposed to reduce the volume of the model parameters exchanged among devices, and an efficient wireless resource allocation scheme is developed. Simulation results show that the proposed FL framework can improve the identification accuracy and convergence time by up to 3.6% and 87% compared to standard FL.

Machine learning (ML) uses data to realize intelligent and autonomous decision-making and inference. ML algorithms have been used in a wide variety of areas, such as computer vision, natural language processing, medical imaging, and communications (1–4). Data are often collected on devices at the edges of networks: Images and text messages are often generated and stored on smartphones; biomedical signals are collected by medical and wearable devices, and often stored on hospital servers; various forms of signals are recorded by Internet of Things systems and sensors.As massive amounts of data are typically required to train an ML model, such as deep neural networks, centralized ML algorithms must collect training data from edge devices for training purposes. For example, to train an ML model for medical diagnosis, a central controller (CC) must collect the medical data from multiple hospitals. Nonetheless, in some applications, such as the aforementioned example of training an ML diagnosis system for medical data, the edge devices may not be willing to share their data, due to privacy concerns and regulations. Furthermore, conveying large volumes of aggregated data by many mobile devices may induce a notable burden on the communication infrastructure. These considerations gave rise to the need for ML algorithms that train an ML model in a distributed fashion, such that edge devices can contribute to the learning procedure without sharing their data.Federated learning (FL) proposed in ref. 5 is a distributed learning algorithm that enables edge devices to jointly train a common ML model without being required to share their data. The FL procedure relies on the ability of each device to train an ML model locally, based on its data, while having the devices iteratively exchanging and synchronizing their local ML model parameters with each other in a manner orchestrated by a CC unit (6). Due to its unique features, FL has been applied in a wide variety of practical applications such as mobile keyboard prediction [e.g., Google (7)], speaker and command recognition [e.g., Apple (8)], and data silos for insurance companies [e.g., WeBank (9)].However, implementation of FL in practical applications faces several challenges which stem from its distributed operation, which is fundamentally different from traditional centralized ML algorithms (10). These challenges include 1) communication overhead induced by the repetitive model parameter exchanges; 2) device hardware heterogeneity, as each device may have different computational capabilities; 3) data heterogeneity, as each device can access a relatively small and personalized dataset and may thus train an ML model which is biased toward its own data; and 4) privacy and security issues, which follow from the fact that the learning procedure is carried out over multiple individual devices. Among these challenges, the communication overhead constitutes a major bottleneck due to the following reasons. First, FL is trained by an iterative process, and hence the time it takes to learn, that is, the convergence time, depends on both the optimization procedure, for example, the number of training steps, as well as the FL parameter transmission delay per training step. Second, FL training is potentially implemented by millions of edge devices, and each device must iteratively share its large-size FL parameters with a CC. Therefore, for FL implemented over a realistic network with limited computational and communication resources, its FL parameter transmission delay may be much larger than the time it takes the devices to train their local ML models. Therefore, it is necessary to design a communication-efficient FL framework that can significantly improve both convergence speed and model accuracy, thus allowing its application to training large-scale ML models over millions of edge devices.A number of existing works, including refs. 11–23, have studied the design of communication-efficient FL algorithms. However, the majority of these works focus on optimization of FL in a single aspect such as device selection and scheduling (11–13), FL model parameter update and transmission (14–18), or network resource management (20–22). In this study, we propose a communication-efficient FL framework that tackles multiple causes for communication delay, by jointly optimizing the device selection, FL model parameter transmission, and network resource management. In particular, we first propose a probabilistic device selection scheme which allows the devices that can significantly improve the convergence speed and training loss to have higher probabilities for ML model transmission. Then, a quantization method is designed to reduce the data size of the ML parameters exchanged among devices, thus improving FL convergence speed. In addition, for the selected devices, a wireless resource allocation scheme is developed to further improve their transmission data rates, thus reducing the FL transmission delay at each learning step. Finally, we analyze the convergence of our proposed FL framework. Simulation results based on real-world data demonstrate the performance of our proposed FL framework and its ability to allow accurate and fast collaborative training of multiple edge devices in a federated manner.     

Pathogenic potential assessment of the Shiga toxin–producing Escherichia coli by a source attribution–considered machine learning model

Instead of conventional serotyping and virulence gene combination methods, methods have been developed to evaluate the pathogenic potential of newly emerging pathogens. Among them, the machine learning (ML)–based method using whole-genome sequencing (WGS) data are getting attention because of the recent advances in ML algorithms and sequencing technologies. Here, we developed various ML models to predict the pathogenicity of Shiga toxin–producing Escherichia coli (STEC) isolates using their WGS data. The input dataset for the ML models was generated using distinct gene repertoires from positive (pathogenic) and negative (nonpathogenic) control groups in which each STEC isolate was designated based on the source attribution, the relative risk potential of the isolation sources. Among the various ML models examined, a model using the support vector machine (SVM) algorithm, the SVM model, discriminated between the two control groups most accurately. The SVM model successfully predicted the pathogenicity of the isolates from the major sources of STEC outbreaks, the isolates with the history of outbreaks, and the isolates that cannot be assessed by conventional methods. Furthermore, the SVM model effectively differentiated the pathogenic potentials of the isolates at a finer resolution. Permutation importance analyses of the input dataset further revealed the genes important for the estimation, proposing the genes potentially essential for the pathogenicity of STEC. Altogether, these results suggest that the SVM model is a more reliable and broadly applicable method to evaluate the pathogenic potential of STEC isolates compared with conventional methods.

Emerging pathogens causing an increasing number of outbreaks are now considered as the major risk to public health (1). The exact assessment of the pathogenic potential of pathogens is required to predict and manage their health risk in advance (2). Conventional methods such as serotyping and virulence gene combinations have been used to assess the bacterial pathogenic potentials (3, 4). However, these conventional assessment methods are not reliable for evaluating the pathogenic potential of emerging pathogens, because the same serotype may carry different virulence genes, and/or contribution of unknown virulence genes to the bacterial pathogenicity is still possible (4, 5). Therefore, the development of methods assessing their pathogenic potential is required to cope with the public health risks caused by newly emerging pathogens.Shiga toxin–producing Escherichia coli (STEC) causes a wide range of human illnesses ranging from mild diarrhea to hemolytic uremic syndrome, which often results in permanent kidney failure (6). In addition to the O157 serotype STEC, emerging non-O157 serotype STECs have been identified as causative agents for the increasing outbreaks lately (5, 7). However, the relationships between the non-O157 serotypes and their pathogenicity have not been defined yet, and thus, predicting the pathogenic potential of the non-O157 serotype STECs has limitations (4, 5). It has been reported that virulence genes such as stx2 and eae are required for the pathogenesis of STEC (4, 5, 7–9). However, the emerging highly pathogenic STEC isolates carry novel virulence genes (4, 5), and indeed, a STEC isolate with a novel combination of stx2 and aggR had caused a huge outbreak in Europe in 2011 (7, 10).Recently, advances in next-generation sequencing technologies have enabled us to exploit whole-genome sequencing (WGS) data (11, 12). Although the WGS data of pathogens can provide rich information about various genetic features of the pathogens, these data are too complex to gain valuable insights on their pathogenicity by using traditional statistical methods (12, 13). In contrast, machine learning (ML) algorithms have notable performance in the analysis of the complex WGS data (12, 13) and therefore have been exploited lately to find out the connection between genetic features and pathogenicity of some pathogens (12, 14–17). The ML algorithms include two broad categories: unsupervised and supervised. The unsupervised ML algorithms, such as phylogenetic tree analysis, principal component analysis (PCA), and Gaussian mixture model (GMM), recognize the inherent patterns in a dataset without the concept of output and then discriminate the given dataset using the inherent patterns (17, 18). On the other hand, the supervised ML algorithms such as Gaussian Naive Bayes (GaussianNB), decision trees (DTs), random forest (RF), and support vector machine (SVM) predict an output from an input data. However, these supervised ML algorithms need to be trained on known input–output pairs until they can predict the correct output using the given input data (17).In this study, we built various ML models and compared their performances of evaluating the pathogenicity of STEC isolates. We subsequently developed an ML model using the SVM algorithm, the SVM model, which can evaluate the pathogenic potential of the STEC isolates the most accurately among the tested ML models. Because the SVM model can also estimate the pathogenic potential of STEC isolates of which the pathogenicity cannot be estimated by conventional methods, the model is more widely applicable to predict the risk of STEC isolates. Moreover, permutation importance analyses discovered the genes important for the evaluation of the SVM model and identified the genes potentially contributing to the pathogenicity of the STEC isolates.     

Evolution and global charge conservation for polarization singularities emerging from non-Hermitian degeneracies

Core concepts in singular optics, especially the polarization singularities, have rapidly penetrated the surging fields of topological and non-Hermitian photonics. For open photonic structures with non-Hermitian degeneracies in particular, polarization singularities would inevitably encounter another sweeping concept of Berry phase. Several investigations have discussed, in an inexplicit way, connections between both concepts, hinting at that nonzero topological charges for far-field polarizations on a loop are inextricably linked to its nontrivial Berry phase when degeneracies are enclosed. In this work, we reexamine the seminal photonic crystal slab that supports the fundamental two-level non-Hermitian degeneracies. Regardless of the invariance of nontrivial Berry phase (concerning near-field Bloch modes defined on the momentum torus) for different loops enclosing both degeneracies, we demonstrate that the associated far polarization fields (defined on the momentum sphere) exhibit topologically inequivalent patterns that are characterized by variant topological charges, including even the trivial scenario of zero charge. Moreover, the charge carried by the Fermi arc actually is not well defined, which could be different on opposite bands. It is further revealed that for both bands, the seemingly complex evolutions of polarizations are bounded by the global charge conservation, with extra points of circular polarizations playing indispensable roles. This indicates that although not directly associated with any local charges, the invariant Berry phase is directly linked to the globally conserved charge, physical principles underlying which have all been further clarified by a two-level Hamiltonian with an extra chirality term. Our work can potentially trigger extra explorations beyond photonics connecting Berry phase and singularities.

Pioneered by Pancharatnam, Berry, Nye, and others (1–10), Berry phase and singularities have become embedded languages all across different branches of photonics. Optical Berry phase is largely manifested through either polarization evolving Pancharatnam–Berry phase or the spin-redirection Bortolotti–Rytov–Vladimirskii–Berry phase (2, 4, 5, 11–15); while optical singularities are widely observed as singularities of intensity (caustics) (6), phase (vortices) (7), or polarization (8–10). As singularities for complex vectorial waves, polarization singularities are skeletons of electromagnetic waves and are vitally important for understanding various interference effects underlying many applications (16–20).There is a superficial similarity between the aforementioned two concepts: Both the topological charge of polarization field [Hopf index of line field (21, 22)] and Berry phase are defined on a closed circuit. Despite this, it is quite unfortunate that almost no definite connections have been established between them in optics. This is fully understandable: Berry phase is defined on the Pancharatnam connection (parallel transport) that decides the phase contrast between neighboring states on the loop (3, 4); while the polarization charge reflects accumulated orientation rotations of polarization ellipses, which has no direct relevance to the overall phase of each state. This explains why in pioneering works where both concepts were present (23–27), their interplay was rarely elaborated on.Spurred by studies into bound states in the continuum, polarization singularities have gained enormous renewed interest in open periodic photonic structures, manifested in different morphologies with both fundamental and higher-order half-integer charges (28–50). Simultaneously, the significance of Berry phase has been further reinforced in surging fields of topological and non-Hermitian photonics (1, 23, 26, 51–56). In open periodic structures involving band degeneracies, Berry phase and polarization singularity would inevitably meet, which sparks the influential work on non-Hermitian degeneracy (36) and several other following studies (40, 43, 45) discussing both concepts simultaneously. Although not claimed explicitly, those works hint that nontrivial Berry phase produces nonzero polarization charges.Aiming to bridge Berry phase and polarization singularity, we reexamine the seminal photonic crystal slab (PCS) that supports elementary two-level non-Hermitian degeneracies. It is revealed that with an invariant nontrivial π Berry phase, the corresponding polarization fields on different isofrequency contours enclosing both non-Hermitian degenerate points (or equivalently exceptional points [EPs]) (26) exhibit diverse patterns characterized by different polarization charges, even including the trivial zero charge. It is further revealed that the charge carried by the Fermi arc is actually not well defined, which could be different on opposite bands. We also discover that such complexity of field evolutions is constrained by global charge conservation for both bands, with extra points of circular polarizations (C points) playing pivotal roles. This reveals the explicit connection between globally conserved charge and the invariant Berry phase, underlying which the physical mechanisms have been further clarified by a two-level Hamiltonian with an extra chirality term (25). We show that such an unexpected connection is generically manifest in various structures, despite the fact that Berry phase and polarization charge actually characterize different entities of near-field Bloch modes and their projected far polarization fields, respectively: Bloch modes are defined on the momentum torus and can be folded into the irreducible Brillouin zone; while polarization fields are defined on the momentum sphere, due to the involvement of out-of-plane wave vectors along which there is no periodicity. Our study can spur further investigations in other subjects beyond photonics to explore conceptual interconnectedness, where both the concepts of Berry phase and singularities are present.     

Circular swimming motility and disordered hyperuniform state in an algae system

Active matter comprises individually driven units that convert locally stored energy into mechanical motion. Interactions between driven units lead to a variety of nonequilibrium collective phenomena in active matter. One of such phenomena is anomalously large density fluctuations, which have been observed in both experiments and theories. Here we show that, on the contrary, density fluctuations in active matter can also be greatly suppressed. Our experiments are carried out with marine algae (Effrenium voratum), which swim in circles at the air–liquid interfaces with two different eukaryotic flagella. Cell swimming generates fluid flow that leads to effective repulsions between cells in the far field. The long-range nature of such repulsive interactions suppresses density fluctuations and generates disordered hyperuniform states under a wide range of density conditions. Emergence of hyperuniformity and associated scaling exponent are quantitatively reproduced in a numerical model whose main ingredients are effective hydrodynamic interactions and uncorrelated random cell motion. Our results demonstrate the existence of disordered hyperuniform states in active matter and suggest the possibility of using hydrodynamic flow for self-assembly in active matter.

Active matter exists over a wide range of spatial and temporal scales (1–6) from animal groups (7, 8) to robot swarms (9–11), to cell colonies and tissues (12–16), to cytoskeletal extracts (17–20), to man-made microswimmers (21–25). Constituent particles in active matter systems are driven out of thermal equilibrium at the individual level; they interact to develop a wealth of intriguing collective phenomena, including clustering (13, 22, 24), flocking (11, 26), swarming (12, 13), spontaneous flow (14, 20), and giant density fluctuations (10, 11). Many of these observed phenomena have been successfully described by particle-based or continuum models (1–6), which highlight the important roles of both individual motility and interparticle interactions in determining system dynamics.Current active matter research focuses primarily on linearly swimming particles which have a symmetric body and self-propel along one of the symmetry axes. However, a perfect alignment between the propulsion direction and body axis is rarely found in reality. Deviation from such a perfect alignment leads to a persistent curvature in the microswimmer trajectories; examples of such circle microswimmers include anisotropic artificial micromotors (27, 28), self-propelled nematic droplets (29, 30), magnetotactic bacteria and Janus particles in rotating external fields (31, 32), Janus particle in viscoelastic medium (33), and sperm and bacteria near interfaces (34, 35). Chiral motility of circle microswimmers, as predicted by theoretical and numerical investigations, can lead to a range of interesting collective phenomena in circular microswimmers, including vortex structures (36, 37), localization in traps (38), enhanced flocking (39), and hyperuniform states (40). However, experimental verifications of these predictions are limited (32, 35), a situation mainly due to the scarcity of suitable experimental systems.Here we address this challenge by investigating marine algae Effrenium voratum (41, 42). At air–liquid interfaces, E. voratum cells swim in circles via two eukaryotic flagella: a transverse flagellum encircling the cellular anteroposterior axis and a longitudinal one running posteriorly. Over a wide range of densities, circling E. voratum cells self-organize into disordered hyperuniform states with suppressed density fluctuations at large length scales. Hyperuniformity (43, 44) has been considered as a new form of material order which leads to novel functionalities (45–49); it has been observed in many systems, including avian photoreceptor patterns (50), amorphous ices (51), amorphous silica (52), ultracold atoms (53), soft matter systems (54–61), and stochastic models (62–64). Our work demonstrates the existence of hyperuniformity in active matter and shows that hydrodynamic interactions can be used to construct hyperuniform states.     

Cholinergic suppression of hippocampal sharp-wave ripples impairs working memory

Learning and memory are assumed to be supported by mechanisms that involve cholinergic transmission and hippocampal theta. Using G protein–coupled receptor-activation–based acetylcholine sensor (GRAB_ACh3.0) with a fiber-photometric fluorescence readout in mice, we found that cholinergic signaling in the hippocampus increased in parallel with theta/gamma power during walking and REM sleep, while ACh3.0 signal reached a minimum during hippocampal sharp-wave ripples (SPW-R). Unexpectedly, memory performance was impaired in a hippocampus-dependent spontaneous alternation task by selective optogenetic stimulation of medial septal cholinergic neurons when the stimulation was applied in the delay area but not in the central (choice) arm of the maze. Parallel with the decreased performance, optogenetic stimulation decreased the incidence of SPW-Rs. These findings suggest that septo–hippocampal interactions play a task-phase–dependent dual role in the maintenance of memory performance, including not only theta mechanisms but also SPW-Rs.

The neurotransmitter acetylcholine is thought to be critical for hippocampus-dependent declarative memories (1, 2). Reduction in cholinergic neurotransmission, either in Alzheimer’s disease or in experiments with cholinergic antagonists, such as scopolamine, impairs memory function (3–8). Acetylcholine may bring about its beneficial effects on memory encoding by enhancing theta rhythm oscillations, decreasing recurrent excitation, and increasing synaptic plasticity (9–11). Conversely, drugs which activate cholinergic receptors enhance learning and, therefore, are a neuropharmacological target for the treatment of memory deficits in Alzheimer’s disease (5, 12, 13).The contribution of cholinergic mechanisms in the acquisition of long-term memories and the role of the hippocampal–entorhinal–cortical interactions are well supported by experimental data (5, 12, 13). In addition, working memory or “short-term” memory is also supported by the hippocampal–entorhinal–prefrontal cortex (14–16). Working memory in humans is postulated to be a conscious process to “keep things in mind” transiently (16). In rodents, matching to sample task, spontaneous alternation between reward locations, and the radial maze task have been suggested to function as a homolog of working memory [“working memory like” (17)].Cholinergic activity is a critical requirement for working memory (18, 19) and for sustaining theta oscillations (10, 20–22). In support of this contention, theta–gamma coupling and gamma power are significantly higher in the choice arm of the maze, compared with those in the side arms where working memory is no longer needed for correct performance (23–26). It has long been hypothesized that working memory is maintained by persistent firing of neurons, which keep the presented items in a transient store in the prefrontal cortex and hippocampal–entorhinal system (27–31), although the exact mechanisms are debated (32–37). An alternative hypothesis holds that items of working memory are stored in theta-nested gamma cycles (38). Common in these models of working memory is the need for an active, cholinergic system–dependent mechanism (39–41). However, in spontaneous alternation tasks, the animals are not moving continuously during the delay, and theta oscillations are not sustained either. During the immobility epochs, theta is replaced by intermittent sharp-wave ripples (SPW-R), yet memory performance does not deteriorate. On the contrary, artificial blockade of SPW-Rs can impair memory performance (42, 43), and prolongation of SPW-Rs improves performance (44). Under the cholinergic hypothesis of working memory, such a result is unexpected.To address the relationship between cholinergic/theta versus SPW-R mechanism in spontaneous alternation, we used a G protein–coupled receptor-activation–based acetylcholine sensor (GRAB_ACh3.0) (45) to monitor acetylcholine (ACh) activity during memory performance in mice. In addition, we optogenetically enhanced cholinergic tone, which suppresses SPW-Rs by a different mechanism than electrically or optogenetically induced silencing of neurons in the hippocampus (43, 44). We show that cholinergic signaling in the hippocampus increases in parallel with theta power/score during walking and rapid eye movement (REM) sleep and reaches a transient minimum during SPW-Rs. Selective optogenetic stimulation of medial septal cholinergic neurons decreased the incidence of SPW-Rs during non-REM sleep (46–48), as well as during the delay epoch of a working memory task and impaired memory performance. These findings demonstrate that memory performance is supported by complementary theta and SPW-R mechanisms.     

Rules of contact inhibition of locomotion for cells on suspended nanofibers

Contact inhibition of locomotion (CIL), in which cells repolarize and move away from contact, is now established as a fundamental driving force in development, repair, and disease biology. Much of what we know of CIL stems from studies on two-dimensional (2D) substrates that do not provide an essential biophysical cue—the curvature of extracellular matrix fibers. We discover rules controlling outcomes of cell–cell collisions on suspended nanofibers and show them to be profoundly different from the stereotyped CIL behavior on 2D substrates. Two approaching cells attached to a single fiber do not repolarize upon contact but rather usually migrate past one another. Fiber geometry modulates this behavior; when cells attach to two fibers, reducing their freedom to reorient, only one cell repolarizes on contact, leading to the cell pair migrating as a single unit. CIL outcomes also change when one cell has recently divided and moves with high speed—cells more frequently walk past each other. Our computational model of CIL in fiber geometries reproduces the core qualitative results of the experiments robustly to model parameters. Our model shows that the increased speed of postdivision cells may be sufficient to explain their increased walk-past rate. We also identify cell–cell adhesion as a key mediator of collision outcomes. Our results suggest that characterizing cell–cell interactions on flat substrates, channels, or micropatterns is not sufficient to predict interactions in a matrix—the geometry of the fiber can generate entirely new behaviors.

Cell migration is an essential component of various physiological processes such as morphogenesis, wound healing, and metastasis (1). Cell–cell interactions in which cell–cell contact reorients cell polarity are necessary for the correct function of many developmental events (2). One of the earliest such interactions known was termed “contact inhibition of locomotion” (CIL) by Abercombie and Heaysman over five decades ago in chick fibroblasts cultured on flat two-dimensional (2D) substrates (2–4). In CIL, two approaching cells isolated from the rest of the cell population first make contact, followed by protrusion inhibition at the site of contact, which leads to cell repolarization through formation of new protrusions away from the site of contact. Subsequently, cells migrate away from each other in the direction of newly formed protrusions (1). This sequence can, however, be altered in specific conditions such as metastasis in which a loss of CIL allows malignant cells to invade fibroblast cultures—this is a loss of CIL between different cell types (heterotypic CIL) (4, 5). Recent work has also begun to identify the molecular players that initiate and regulate CIL, including Rac activity, microtubules, Eph/Ephrin binding, and E- and N-cadherin expression (6–10).CIL is most commonly studied and analyzed on flat 2D substrates using several invasion and collision assays (2, 3, 11). By contrast, cells traveling in matrix in vivo are constrained to move along narrow fibers. A common shortcoming in featureless 2D assays is thus the inability to study CIL under natural constraints (11–13). Recently, micropatterned substrates have been used to understand restricted motility, developing one-dimensional (1D) collision assays where cell migration is constrained to straight lines, allowing for a greater occurrence of cell–cell collisions to quantify rates and outcomes of different types of cell–cell interactions (11, 13–15). These interactions do not necessarily resemble the stereotyped CIL behavior. Broadly, experiments and simulations (16–18) have observed the following: 1) the classical stereotype of CIL with two cells contacting head-on, with both cells repolarizing (referred to as “reversal” or “mutual CIL”); 2) after a head-on collision, only one cell reverses (“training” or “nonmutual CIL”); and 3) cells manage to crawl past or over one another, exchanging positions (“walk past” or “sliding”). Within the well-studied neural-crest cell explants, walk past is extremely rare (11), but it can occur in epithelial cells, especially in those that have been metastatically transformed or that have decreased E-cadherin expression (15).Both 2D substrates and micropatterned stripes provide controllable and reproducible environments but neither fully models the details of in vivo native cellular environments, which consist of extracellular matrices (ECM) of fibrous proteins, with these fibers having different radii. Our earlier in vitro recapitulation of the effects of fiber curvature showed that both protrusive and migratory behavior is sensitive to fiber diameter (19–21). Furthermore, we have shown that suspended, flat 2D ribbons do not capture the protrusive behavior observed on suspended round fibers (19); thus, we wanted to inquire if the CIL rules developed on 1D collision and 2D assays extend to contextually relevant fibrous environments. To understand CIL in fibrous environments that mimic native ECM, we use suspended and aligned nanofiber networks to study CIL behavior in NIH/3T3 fibroblast cell–cell pairs exhibiting two distinct elongated morphologies: spindle, attached to a single fiber and parallel cuboidal, attached to two fibers (22). We further investigate the effect of cell division on CIL by studying the encounters of cells that have recently divided (daughter cells) with other cells; these recently divided cells are much faster, consistent with earlier work (23). Our work allows us to determine the types and rates of cell–cell contact outcomes—the “rules of CIL”—in a biologically relevant system with a controlled geometry. These rules are radically different from the known stereotypical behavior in 2D assays, but the essential features of these rules emerge robustly from a minimal computational model of CIL in confined geometries.     

Chemokine CCL5 promotes robust optic nerve regeneration and mediates many of the effects of CNTF gene therapy

Ciliary neurotrophic factor (CNTF) is a leading therapeutic candidate for several ocular diseases and induces optic nerve regeneration in animal models. Paradoxically, however, although CNTF gene therapy promotes extensive regeneration, recombinant CNTF (rCNTF) has little effect. Because intraocular viral vectors induce inflammation, and because CNTF is an immune modulator, we investigated whether CNTF gene therapy acts indirectly through other immune mediators. The beneficial effects of CNTF gene therapy remained unchanged after deleting CNTF receptor alpha (CNTFRα) in retinal ganglion cells (RGCs), the projection neurons of the retina, but were diminished by depleting neutrophils or by genetically suppressing monocyte infiltration. CNTF gene therapy increased expression of C-C motif chemokine ligand 5 (CCL5) in immune cells and retinal glia, and recombinant CCL5 induced extensive axon regeneration. Conversely, CRISPR-mediated knockdown of the cognate receptor (CCR5) in RGCs or treating wild-type mice with a CCR5 antagonist repressed the effects of CNTF gene therapy. Thus, CCL5 is a previously unrecognized, potent activator of optic nerve regeneration and mediates many of the effects of CNTF gene therapy.

Like most pathways in the mature central nervous system (CNS), the optic nerve cannot regenerate once damaged due in part to cell-extrinsic suppressors of axon growth (1, 2) and the low intrinsic growth capacity of adult retinal ganglion cells (RGCs), the projection neurons of the eye (3–5). Consequently, traumatic or ischemic optic nerve injury or degenerative diseases such as glaucoma lead to irreversible visual losses. Experimentally, some degree of regeneration can be induced by intraocular inflammation or growth factors expressed by inflammatory cells (6–10), altering the cell-intrinsic growth potential of RGCs (3–5), enhancing physiological activity (11, 12), chelating free zinc (13, 14), and other manipulations (15–19). However, the extent of regeneration achieved to date remains modest, underlining the need for more effective therapies.Ciliary neurotrophic factor (CNTF) is a leading therapeutic candidate for glaucoma and other ocular diseases (20–23). Activation of the downstream signal transduction cascade requires CNTF to bind to CNTF receptor-α (CNTFRα) (24), which leads to recruitment of glycoprotein 130 (gp130) and leukemia inhibitory factor receptor-β (LIFRβ) to form a tripartite receptor complex (25). CNTFRα anchors to the plasma membrane through a glycosylphosphatidylinositol linkage (26) and can be released and become soluble through phospholipase C-mediated cleavage (27). CNTF has been reported to activate STAT3 phosphorylation in retinal neurons, including RGCs, and to promote survival, but it is unknown whether these effects are mediated by direct action of CNTF on RGCs via CNTFRα (28). Our previous studies showed that CNTF promotes axon outgrowth from neonate RGCs in culture (29) but fails to do so in cultured mature RGCs (8) or in vivo (6). Although some studies report that recombinant CNTF (rCNTF) can promote optic nerve regeneration (20, 30, 31), others find little or no effect unless SOCS3 (suppressor of cytokine signaling-3), an inhibitor of the Jak-STAT pathway, is deleted in RGCs (5, 6, 32). In contrast, multiple studies show that adeno-associated virus (AAV)-mediated expression of CNTF in RGCs induces strong regeneration (33–40). The basis for the discrepant effects of rCNTF and CNTF gene therapy is unknown but is of considerable interest in view of the many promising clinical and preclinical outcomes obtained with CNTF to date.Because intravitreal virus injections induce inflammation (41), we investigated the possibility that CNTF, a known immune modulator (42–44), might act by elevating expression of other immune-derived factors. We report here that the beneficial effects of CNTF gene therapy in fact require immune system activation and elevation of C-C motif chemokine ligand 5 (CCL5). Depletion of neutrophils, global knockout (KO) or RGC-selective deletion of the CCL5 receptor CCR5, or a CCR5 antagonist all suppress the effects of CNTF gene therapy, whereas recombinant CCL5 (rCCL5) promotes axon regeneration and increases RGC survival. These studies point to CCL5 as a potent monotherapy for optic nerve regeneration and to the possibility that other applications of CNTF and other forms of gene therapy might similarly act indirectly through other factors.     

Tropospheric carbonyl sulfide mass balance based on direct measurements of sulfur isotopes

Robust estimates for the rates and trends in terrestrial gross primary production (GPP; plant CO₂ uptake) are needed. Carbonyl sulfide (COS) is the major long-lived sulfur-bearing gas in the atmosphere and a promising proxy for GPP. Large uncertainties in estimating the relative magnitude of the COS sources and sinks limit this approach. Sulfur isotope measurements (³⁴S/³²S; δ³⁴S) have been suggested as a useful tool to constrain COS sources. Yet such measurements are currently scarce for the atmosphere and absent for the marine source and the plant sink, which are two main fluxes. Here we present sulfur isotopes measurements of marine and atmospheric COS, and of plant-uptake fractionation experiments. These measurements resulted in a complete data-based tropospheric COS isotopic mass balance, which allows improved partition of the sources. We found an isotopic (δ³⁴S ± SE) value of 13.9 ± 0.1‰ for the troposphere, with an isotopic seasonal cycle driven by plant uptake. This seasonality agrees with a fractionation of −1.9 ± 0.3‰ which we measured in plant-chamber experiments. Air samples with strong anthropogenic influence indicated an anthropogenic COS isotopic value of 8 ± 1‰. Samples of seawater-equilibrated-air indicate that the marine COS source has an isotopic value of 14.7 ± 1‰. Using our data-based mass balance, we constrained the relative contribution of the two main tropospheric COS sources resulting in 40 ± 17% for the anthropogenic source and 60 ± 20% for the oceanic source. This constraint is important for a better understanding of the global COS budget and its improved use for GPP determination.

The Earth system is going through rapid changes as the climate warms and CO₂ level rises. This rise in CO₂ is mitigated by plant uptake; hence, it is important to estimate global and regional photosynthesis rates and trends (1). Yet, robust tools for investigating these processes at a large scale are scarce (2). Recent studies suggest that carbonyl sulfide (COS) could provide an improved constraint on terrestrial photosynthesis (gross primary production, GPP) (2–12). COS is the major long-lived sulfur-bearing gas in the atmosphere and the main supplier of sulfur to the stratospheric sulfate aerosol layer (13), which exerts a cooling effect on the Earth’s surface and regulates stratospheric ozone chemistry (14).During terrestrial photosynthesis, COS diffuses into leaf stomata and is consumed by photosynthetic enzymes in a similar manner to CO₂ (3–5). Contrary to CO₂, COS undergoes rapid and irreversible hydrolysis mainly by the enzyme carbonic-anhydrase (6, 7). Thus, COS can be used as a proxy for the one-way flux of CO₂ removal from the atmosphere by terrestrial photosynthesis (2, 8–11). However, the large uncertainties in estimating the COS sources weaken this approach (10–12, 15). Tropospheric COS has two main sources: the oceans and anthropogenic emissions, and one main sink–terrestrial plant uptake (8, 10–13). Smaller sources include biomass burning, soil emissions, wetlands, volcanoes, and smaller sinks include OH destruction, stratospheric destruction, and soil uptake (12). The largest source of COS to the atmosphere is the ocean, both as direct COS emission, and as indirect carbon disulfide (CS₂) and dimethylsulfide (DMS) emissions that are rapidly oxidized to COS (10, 16–20). Recent studies suggest oceanic COS emissions are in the range of 200–4,000 GgS/y (19–22). The second major COS source is the anthropogenic source, which is dominated by indirect emissions derived from CS₂ oxidation, mainly from the use of CS₂ as an industrial solvent. Direct emissions of COS are mainly derived from coal and fuel combustion (17, 23, 24). Recent studies suggest that anthropogenic emissions are in the range of 150–585 GgS/y (23, 24). The terrestrial plant uptake is estimated to be in the range of 400–1,360 GgS/y (11). Measurements of sulfur isotope ratios (δ³⁴S) in COS may be used to track COS sources and thus reduce the uncertainties in their flux estimations (15, 25–27). However, the isotopic mass balance approach works best if the COS end members are directly measured and have a significantly different isotopic signature. Previous δ³⁴S measurements of atmospheric COS are scarce and there have been no direct measurements of two important components: the δ³⁴S of oceanic COS emissions, and the isotopic fractionation of COS during plant uptake (15, 25–27). In contrast to previous studies that used assessments for these isotopic values, our aim was to directly measure the isotopic values of these missing components, and to determine the tropospheric COS δ³⁴S variability over space and time.     

A biophysical limit for quorum sensing in biofilms

Bacteria grow on surfaces in complex immobile communities known as biofilms, which are composed of cells embedded in an extracellular matrix. Within biofilms, bacteria often interact with members of their own species and cooperate or compete with members of other species via quorum sensing (QS). QS is a process by which microbes produce, secrete, and subsequently detect small molecules called autoinducers (AIs) to assess their local population density. We explore the competitive advantage of QS through agent-based simulations of a spatial model in which colony expansion via extracellular matrix production provides greater access to a limiting diffusible nutrient. We note a significant difference in results based on whether AI production is constitutive or limited by nutrient availability: If AI production is constitutive, simple QS-based matrix-production strategies can be far superior to any fixed strategy. However, if AI production is limited by nutrient availability, QS-based strategies fail to provide a significant advantage over fixed strategies. To explain this dichotomy, we derive a biophysical limit for the dynamic range of nutrient-limited AI concentrations in biofilms. This range is remarkably small (less than 10-fold) for the realistic case in which a growth-limiting diffusible nutrient is taken up within a narrow active growth layer. This biophysical limit implies that for QS to be most effective in biofilms AI production should be a protected function not directly tied to metabolism.

Bacteria are capable of communicating with their neighbors through a process known as quorum sensing (QS). QS depends on the secretion and detection of small, diffusible molecules known as autoinducers (AIs), whose concentration increases with increasing cell density (1, 2). QS allows bacteria to control processes that are unproductive when undertaken by an individual but effective when undertaken by all members of the group and thus leads to a competitive advantage for bacterial communities that employ QS (1–6).QS is known to promote and regulate bacterial biofilms: immobile communities of cells densely packed in an extracellular matrix (7). QS has been demonstrated to be critical to proper biofilm formation (8–13). For example, Pseudomonas aeruginosa mutants that do not synthesize AIs terminate biofilm formation at an early stage (14). Given that the interior of biofilms is known to be nutrient-deficient (15), it is an open question to what extent these interior cells participate in AI production. Indeed, in many cases AI production relies on central metabolic compounds. For example, a substrate for synthesis of the ubiquitous acyl-HSL AIs is produced by one-carbon metabolism, which is highly dependent on nutrient availability (16–18). Thus, we sought to understand whether QS can be advantageous to cells in a biofilm if AI production depends on access to nutrients.One context in which QS has been found to afford a competitive advantage in biofilms is via regulation of production of the extracellular matrix, which is composed of biopolymers, including polysaccharides, proteins, nucleic acids, and lipids. Advantages provided by the matrix include adhering cells to each other and to a substrate, creating a protective barrier against chemicals and predators, and facilitating horizontal gene transfer.In simple models of biofilms that incorporate realistic reaction-diffusion effects, Xavier and Foster (19) found that matrix production allows cells to push descendants outward from a surface into a more O2-rich environment. Consequently, they found that matrix production provides a strong competitive advantage to cell lineages by suffocating neighboring nonproducing cells (19). Building upon this work, Nadell et al. (20) showed that strategies that employ QS to deactivate matrix production in mature biofilms can yield a further advantage by redirecting resources into reproduction, and this scenario has been replicated and further developed (21–23). Notably, all these models assume constitutive AI production with no dependence on nutrient availability (20–23). We were therefore inspired to ask whether QS still provide an advantage in regulating matrix production if AI production is limited by nutrient availability.To this end, we simulated competitions among biofilm-forming cells, comparing matrix-production strategies that employ QS with strategies that do not. While QS cells that constitutively produce AI could outcompete all fixed strategies, we found, surprisingly, that QS cells which produce AI in a nutrient-dependent manner have essentially no advantage over non-QS cells. We trace this result to a novel biophysical limit on the dynamic range (DR) of AI concentrations if AI production is nutrient-limited. This biophysical limit applies to all bacterial systems that employ QS. These results suggest that for QS to be effective in biofilms and other conditions where nutrients are limiting, cells must privilege AI production despite the metabolic cost. From this perspective, autoinduction itself, i.e., positive feedback on AI production from AI sensing, can be viewed as one way that cells decouple AI production from metabolism.     

Cognitive aging is associated with redistribution of synaptic weights in the hippocampus

Behaviors that rely on the hippocampus are particularly susceptible to chronological aging, with many aged animals (including humans) maintaining cognition at a young adult-like level, but many others the same age showing marked impairments. It is unclear whether the ability to maintain cognition over time is attributable to brain maintenance, sufficient cognitive reserve, compensatory changes in network function, or some combination thereof. While network dysfunction within the hippocampal circuit of aged, learning-impaired animals is well-documented, its neurobiological substrates remain elusive. Here we show that the synaptic architecture of hippocampal regions CA1 and CA3 is maintained in a young adult-like state in aged rats that performed comparably to their young adult counterparts in both trace eyeblink conditioning and Morris water maze learning. In contrast, among learning-impaired, but equally aged rats, we found that a redistribution of synaptic weights amplifies the influence of autoassociational connections among CA3 pyramidal neurons, yet reduces the synaptic input onto these same neurons from the dentate gyrus. Notably, synapses within hippocampal region CA1 showed no group differences regardless of cognitive ability. Taking the data together, we find the imbalanced synaptic weights within hippocampal CA3 provide a substrate that can explain the abnormal firing characteristics of both CA3 and CA1 pyramidal neurons in aged, learning-impaired rats. Furthermore, our work provides some clarity with regard to how some animals cognitively age successfully, while others’ lifespans outlast their “mindspans.”

Aging is the biggest risk factor for Alzheimer’s disease, but many aged individuals nevertheless retain the ability to perform cognitive tasks with young adult (YA)-like competency, and are thus resilient to age-related cognitive decline and dementias (1, 2). The mechanisms of such resilience are unknown, but are thought to involve neural or cognitive reserve, brain network adaptations, or simply the ability to maintain cognitive brain circuits in a YA-like state (3–5). Much of the cellular and functional insight into the concept or risk of/resilience against age-related cognitive impairments has come from animal models of normal/nonpathological aging (6–10). Many of these studies have shown that circuit function abnormalities are associated with behavioral impairments. The cellular and structural bases for such functional aberrations, however, remain largely unknown.Two of the most well-studied cognitive domains that show susceptibility to chronological aging in both rodents and nonhuman primates are working memory and spatial/temporal memory (6–10). Importantly, these cognitive domains engage anatomically distinct neurocognitive systems, with the former relying on prefrontal/orbitofrontal cortical circuits and the latter relying on hippocampal circuitry. Interestingly, although behavioral deficits in these two domains (in the case of rat models of cognitive aging) begin to emerge, worsen, and become increasingly prevalent between 12 and 18 mo of age in most strains (reviewed in refs. 9 and 11), cognitive aging within hippocampus-dependent forms of learning and memory are relatively independent of those that engage the prefrontal/orbitofrontal cortical neural systems (6–9, 12–15).Neither the mechanisms underlying the conservation of memory function across chronological aging nor those contributing to the age-related emergence and exacerbation of memory impairments are clearly understood for either neurocognitive system. It is clear, however, that neither frank neuronal loss (16, 17) nor overall synapse loss (18) contributes to cognitive aging within the medial temporal lobe/hippocampal memory system. Rather, the intriguing idea that has emerged from work in both the hippocampal and the prefrontal/orbitofrontal cortical memory systems is that there are functional alterations in the synaptic connections in individual microcircuits embedded within these larger neuroanatomical systems (6–10, 19–31).Axospinous synapses (including those in hippocampal and cortical circuits) are characterized on the basis of the three-dimensional morphology of their postsynaptic densities (PSDs) (20, 32–34). The most-abundant axospinous synaptic subtype has a continuous, macular, disk-shaped PSD, as compared to the less-abundant perforated synaptic subtype, which has at least one discontinuity in its PSD (34). In addition to differing substantially with regard to relative frequency, perforated and nonperforated synapses also harbor major differences in size and synaptic AMPA-type and NMDA-type receptor expression levels (AMPAR and NMDAR, respectively) (34–38). There is also evidence that perforated and nonperforated synapses are differentially involved in synaptic plasticity (39–44) and in preservation of—or reductions in—memory function during chronological aging (6, 20, 45). Layered onto these general distinctions between perforated and nonperforated synapses are more specific differences in their characteristics when considered within neural circuits. For example, perforated synapses have a stronger and more consistent influence on neuronal computation within hippocampal region CA1 than their nonperforated counterparts, which nevertheless outnumber the former by a roughly 9-to-1 ratio (34, 46, 47).These and other circuit-specific differences necessitate a circuit-based approach to understanding the synaptic bases underlying the retention or loss of YA-like memory function in the aging brain. In many ways, the hippocampal system is particularly convenient for such circuit-based approaches (48, 49). Information about the internal and external world is funneled to the parahippocampal system and then relayed via the entorhinal cortex to the dentate gyrus, the first component of the so-called trisynaptic circuit in the hippocampus proper. Granule cells in the dentate gyrus then transmit their computations to hippocampal region CA3 via the mossy fibers, which form very large and anatomically distinct synapses called mossy fiber bouton–thorny excrescence synaptic complexes in the stratum lucidum (SL). CA3 pyramidal neurons then integrate information from their autoassociational connections in the stratum radiatum (SR) and stratum oriens (SO), with both direct entorhinal inputs in stratum lacunosum-moleculare (SLM) and the dentate gyrus inputs in the SL, and convey this information to hippocampal CA1 pyramidal neurons. Neurons in hippocampal CA1 then integrate this information in their basal and apical SR dendrites with direct entorhinal cortical inputs in their most distal, tufted dendrites in the SLM, and represent the first and largest extrahippocampal output from the hippocampus proper. Thus, the computations performed both within individual hippocampal subregions and between them as an interconnected neurocognitive system are complex, and involve a combination of intrinsic (i.e., membrane-bound ion channels that regulate membrane excitability) and synaptic (i.e., ligand-gated ion channels expressed at both excitatory and inhibitory synapses) influences. Additionally, age-related changes at any level of these complex circuits will have downstream consequences on the accuracy/reliability of the information being relayed to extrahippocampal regions via CA1 pyramidal neurons.Given the amount of evidence supporting a possible synaptic explanation for age-related learning and memory impairments in hippocampus-dependent forms of cognition (6–10), we combined patch-clamp physiology, serial section conventional and immunogold electron microscopy (EM), quantitative Western blot analyses, and behavioral characterization using two hippocampus-dependent forms of learning in YA (6- to 8-mo old) and aged rats (28- to 29-mo old) to examine two interconnected hippocampal regions implicated in cognitive aging: Regions CA1 and CA3. We focused on CA1 and CA3 because of their central location in the hippocampal circuit (48–50), their similar laminar dendritic structure (48–50), and their well-documented age-related changes in place field specificity and reliability (51–56).We find that the synaptic architecture and balance of synaptic weights in YA and aged, learning-unimpaired (AU) rats is remarkably similar, but that both are different in aged, learning-impaired (AI) rats. Moreover, this restructuring among “unsuccessful” cognitive agers has an intriguing specificity: It involves only AMPARs, only perforated axospinous synapses, and only hippocampal region CA3, which together shift the balance of synaptic weights that drive action potential output in CA3 pyramidal neurons maladaptively toward an overemphasis of the autoassociational synapses that interconnect CA3 pyramidal neurons.     

LGI1–ADAM22–MAGUK configures transsynaptic nanoalignment for synaptic transmission and epilepsy prevention

Physiological functioning and homeostasis of the brain rely on finely tuned synaptic transmission, which involves nanoscale alignment between presynaptic neurotransmitter-release machinery and postsynaptic receptors. However, the molecular identity and physiological significance of transsynaptic nanoalignment remain incompletely understood. Here, we report that epilepsy gene products, a secreted protein LGI1 and its receptor ADAM22, govern transsynaptic nanoalignment to prevent epilepsy. We found that LGI1–ADAM22 instructs PSD-95 family membrane-associated guanylate kinases (MAGUKs) to organize transsynaptic protein networks, including NMDA/AMPA receptors, Kv₁ channels, and LRRTM4–Neurexin adhesion molecules. Adam22^ΔC5/ΔC5 knock-in mice devoid of the ADAM22–MAGUK interaction display lethal epilepsy of hippocampal origin, representing the mouse model for ADAM22-related epileptic encephalopathy. This model shows less-condensed PSD-95 nanodomains, disordered transsynaptic nanoalignment, and decreased excitatory synaptic transmission in the hippocampus. Strikingly, without ADAM22 binding, PSD-95 cannot potentiate AMPA receptor-mediated synaptic transmission. Furthermore, forced coexpression of ADAM22 and PSD-95 reconstitutes nano-condensates in nonneuronal cells. Collectively, this study reveals LGI1–ADAM22–MAGUK as an essential component of transsynaptic nanoarchitecture for precise synaptic transmission and epilepsy prevention.

Epilepsy, characterized by unprovoked, recurrent seizures, affects 1 to 2% of the population worldwide. Many genes that cause inherited epilepsy when mutated encode ion channels, and dysregulated synaptic transmission often causes epilepsy (1, 2). Although antiepileptic drugs have mainly targeted ion channels, they are not always effective and have adverse effects. It is therefore important to clarify the detailed processes for synaptic transmission and how they are affected in epilepsy.Recent superresolution imaging of the synapse reveals previously overlooked subsynaptic nano-organizations and pre- and postsynaptic nanodomains (3–6), and mathematical simulation suggests their nanometer-scale coordination in individual synapses for efficient synaptic transmission: presynaptic neurotransmitter release machinery and postsynaptic receptors precisely align across the synaptic cleft to make “transsynaptic nanocolumns” (7, 8).So far, numerous transsynaptic cell-adhesion molecules have been identified (9–12), including presynaptic Neurexins and type IIa receptor protein tyrosine phosphatases (PTPδ, PTPσ, and LAR) and postsynaptic Neuroligins, LRRTMs, NGL-3, IL1RAPL1, Slitrks, and SALMs. Neurexins–Neuroligins have attracted particular attention because of their synaptogenic activities when overexpressed and their genetic association with neuropsychiatric disorders (e.g., autism). Another type of transsynaptic adhesion complex mediated by synaptically secreted Cblns (e.g., Neurexin–Cbln1–GluD2) promotes synapse formation and maintenance (13–15). Genetic studies in Caenorhabditis elegans show that secreted Ce-Punctin, the ortholog of the mammalian ADAMTS-like family, specifies cholinergic versus GABAergic identity of postsynaptic domains and functions as an extracellular synaptic organizer (16). However, the molecular identity and in vivo physiological significance of transsynaptic nanocolumns remain incompletely understood.LGI1, a neuronal secreted protein, and its receptor ADAM22 have recently emerged as major determinants of brain excitability (17) as 1) mutations in the LGI1 gene cause autosomal dominant lateral temporal lobe epilepsy (18); 2) mutations in the ADAM22 gene cause infantile epileptic encephalopathy with intractable seizures and intellectual disability (19, 20); 3) Lgi1 or Adam22 knockout mice display lethal epilepsy (21–24); and 4) autoantibodies against LGI1 cause limbic encephalitis characterized by seizures and amnesia (25–28). Functionally, LGI1–ADAM22 regulates AMPA receptor (AMPAR) and NMDA receptor (NMDAR)-mediated synaptic transmission (17, 22, 29) and Kv₁ channel-mediated neuronal excitability (30, 31). Recent structural analysis shows that LGI1 and ADAM22 form a 2:2 heterotetrameric assembly (ADAM22–LGI1–LGI1–ADAM22) (32), suggesting the transsynaptic configuration.In this study, we identify ADAM22-mediated synaptic protein networks in the brain, including pre- and postsynaptic MAGUKs and their functional bindings to transmembrane proteins (NMDA/AMPA glutamate receptors, voltage-dependent ion channels, cell-adhesion molecules, and vesicle-fusion machinery). ADAM22 knock-in mice lacking the MAGUK-binding motif show lethal epilepsy of hippocampal origin. In this mouse, postsynaptic PSD-95 nano-assembly as well as nano-scale alignment between pre- and postsynaptic proteins are significantly impaired. Importantly, PSD-95 is no longer able to modulate AMPAR-mediated synaptic transmission without binding to ADAM22. These findings establish that LGI1–ADAM22 instructs MAGUKs to organize transsynaptic nanocolumns and guarantee the stable brain activity.     

CYK-1/Formin activation in cortical RhoA signaling centers promotes organismal left–right symmetry breaking

Proper left–right symmetry breaking is essential for animal development, and in many cases, this process is actomyosin-dependent. In Caenorhabditis elegans embryos active torque generation in the actomyosin layer promotes left–right symmetry breaking by driving chiral counterrotating cortical flows. While both Formins and Myosins have been implicated in left–right symmetry breaking and both can rotate actin filaments in vitro, it remains unclear whether active torques in the actomyosin cortex are generated by Formins, Myosins, or both. We combined the strength of C. elegans genetics with quantitative imaging and thin film, chiral active fluid theory to show that, while Non-Muscle Myosin II activity drives cortical actomyosin flows, it is permissive for chiral counterrotation and dispensable for chiral symmetry breaking of cortical flows. Instead, we find that CYK-1/Formin activation in RhoA foci is instructive for chiral counterrotation and promotes in-plane, active torque generation in the actomyosin cortex. Notably, we observe that artificially generated large active RhoA patches undergo rotations with consistent handedness in a CYK-1/Formin–dependent manner. Altogether, we conclude that CYK-1/Formin–dependent active torque generation facilitates chiral symmetry breaking of actomyosin flows and drives organismal left–right symmetry breaking in the nematode worm.

The emergence of left–right asymmetry is essential for normal animal development and, in the majority of animal species, one type of handedness is dominant (1). The actin cytoskeleton plays an instrumental role in establishing the left–right asymmetric body plan of invertebrates like fruit flies (2–6), nematodes (7–11), and pond snails (12–15). Moreover, an increasing number of studies showed that vertebrate left–right patterning also depends on a functional actomyosin cytoskeleton (13, 16–22). Actomyosin-dependent chiral behavior has even been reported in isolated cells (23–28) and such cell-intrinsic chirality has been shown to promote left–right asymmetric morphogenesis of tissues (29, 30), organs (21, 31), and entire embryonic body plans (12, 13, 32, 33). Active force generation in the actin cytoskeleton is responsible for shaping cells and tissues during embryo morphogenesis. Torques are rotational forces with a given handedness and it has been proposed that in plane, active torque generation in the actin cytoskeleton drives chiral morphogenesis (7, 8, 34, 35).What could be the molecular origin of these active torques? The actomyosin cytoskeleton consists of actin filaments, actin-binding proteins, and Myosin motors. Actin filaments are polar polymers with a right-handed helical pitch and are therefore chiral themselves (36, 37). Due to the right-handed pitch of filamentous actin, Myosin motors can rotate actin filaments along their long axis while pulling on them (33, 38–42). Similarly, when physically constrained, members of the Formin family rotate actin filaments along their long axis while elongating them (43). In both cases the handedness of this rotation is determined by the helical nature of the actin polymer. From this it follows that both Formins and Myosins are a potential source of molecular torque generation that could drive cellular and organismal chirality. Indeed, chiral processes across different length scales, and across species, are dependent on Myosins (19), Formins (13–15, 26), or both (7, 8, 21, 44). It is, however, unclear how Formins and Myosins contribute to active torque generation and the emergence chiral processes in developing embryos.In our previous work we showed that the actomyosin cortex of some Caenorhabditis elegans embryonic blastomeres undergoes chiral counterrotations with consistent handedness (7, 35). These chiral actomyosin flows can be recapitulated using active chiral fluid theory that describes the actomyosin layer as a thin-film, active gel that generates active torques (7, 45, 46). Chiral counterrotating cortical flows reorient the cell division axis, which is essential for normal left–right symmetry breaking (7, 47). Moreover, cortical counterrotations with the same handedness have been observed in Xenopus one-cell embryos (32), suggesting that chiral counterrotations are conserved among distant species. Chiral counterrotating actomyosin flow in C. elegans blastomeres is driven by RhoA signaling and is dependent on Non-Muscle Myosin II motor proteins (7). Moreover, the Formin CYK-1 has been implicated in actomyosin flow chirality during early polarization of the zygote as well as during the first cytokinesis (48, 49). Despite having identified a role for Myosins and Formins, the underlying mechanism by which active torques are generated remains elusive.Here we show that the Diaphanous-like Formin, CYK-1/Formin, is a critical determinant for the emergence of actomyosin flow chirality, while Non-Muscle Myosin II (NMY-2) plays a permissive role. Our results show that cortical CYK-1/Formin is recruited by active RhoA signaling foci and promotes active torque generation, which in turn tends to locally rotate the actomyosin cortex clockwise. In the highly connected actomyosin meshwork, a gradient of these active torques drives the emergence of chiral counterrotating cortical flows with uniform handedness, which is essential for proper left–right symmetry breaking. Together, these results provide mechanistic insight into how Formin-dependent torque generation drives cellular and organismal left–right symmetry breaking.     

INAUGURAL ARTICLE by a Recently Elected Academy Member:The effect of parathyroid hormone on osteogenesis is mediated partly by osteolectin

We previously described a new osteogenic growth factor, osteolectin/Clec11a, which is required for the maintenance of skeletal bone mass during adulthood. Osteolectin binds to Integrin α11 (Itga11), promoting Wnt pathway activation and osteogenic differentiation by leptin receptor⁺ (LepR⁺) stromal cells in the bone marrow. Parathyroid hormone (PTH) and sclerostin inhibitor (SOSTi) are bone anabolic agents that are administered to patients with osteoporosis. Here we tested whether osteolectin mediates the effects of PTH or SOSTi on bone formation. We discovered that PTH promoted Osteolectin expression by bone marrow stromal cells within hours of administration and that PTH treatment increased serum osteolectin levels in mice and humans. Osteolectin deficiency in mice attenuated Wnt pathway activation by PTH in bone marrow stromal cells and reduced the osteogenic response to PTH in vitro and in vivo. In contrast, SOSTi did not affect serum osteolectin levels and osteolectin was not required for SOSTi-induced bone formation. Combined administration of osteolectin and PTH, but not osteolectin and SOSTi, additively increased bone volume. PTH thus promotes osteolectin expression and osteolectin mediates part of the effect of PTH on bone formation.

The maintenance and repair of the skeleton require the generation of new bone cells throughout adult life. Osteoblasts are relatively short-lived cells that are constantly regenerated, partly by skeletal stem cells within the bone marrow (1). The main source of new osteoblasts in adult bone marrow is leptin receptor-expressing (LepR⁺) stromal cells (2–4). These cells include the multipotent skeletal stem cells that give rise to the fibroblast colony-forming cells (CFU-Fs) in the bone marrow (2), as well as restricted osteogenic progenitors (5) and adipocyte progenitors (6–8). LepR⁺ cells are a major source of osteoblasts for fracture repair (2) and growth factors for hematopoietic stem cell maintenance (9–11).One growth factor synthesized by LepR⁺ cells, as well as osteoblasts and osteocytes, is osteolectin/Clec11a, a secreted glycoprotein of the C-type lectin domain superfamily (5, 12, 13). Osteolectin is an osteogenic factor that promotes the maintenance of the adult skeleton by promoting the differentiation of LepR⁺ cells into osteoblasts. Osteolectin acts by binding to integrin α11β1, which is selectively expressed by LepR⁺ cells and osteoblasts, activating the Wnt pathway (12). Deficiency for either Osteolectin or Itga11 (the gene that encodes integrin α11) reduces osteogenesis during adulthood and causes early-onset osteoporosis in mice (12, 13). Recombinant osteolectin promotes osteogenic differentiation by bone marrow stromal cells in culture and daily injection of mice with osteolectin systemically promotes bone formation.Osteoporosis is a progressive condition characterized by reduced bone mass and increased fracture risk (14). Several factors contribute to osteoporosis development, including aging, estrogen insufficiency, mechanical unloading, and prolonged glucocorticoid use (14). Existing therapies include antiresorptive agents that slow bone loss, such as bisphosphonates (15, 16) and estrogens (17), and anabolic agents that increase bone formation, such as parathyroid hormone (PTH) (18), PTH-related protein (19), and sclerostin inhibitor (SOSTi) (20). While these therapies increase bone mass and reduce fracture risk, they are not a cure.PTH promotes both anabolic and catabolic bone remodeling (21–24). PTH is synthesized by the parathyroid gland and regulates serum calcium levels, partly by regulating bone formation and bone resorption (23–25). PTH1R is a PTH receptor (26, 27) that is strongly expressed by LepR⁺ bone marrow stromal cells (8, 28–30). Recombinant human PTH (Teriparatide; amino acids 1 to 34) and synthetic PTH-related protein (Abaloparatide) are approved by the US Food and Drug Administration (FDA) for the treatment of osteoporosis (19, 31). Daily (intermittent) administration of PTH increases bone mass by promoting the differentiation of osteoblast progenitors, inhibiting osteoblast and osteocyte apoptosis, and reducing sclerostin levels (32–35). PTH promotes osteoblast differentiation by activating Wnt and BMP signaling in bone marrow stromal cells (28, 36, 37), although the mechanisms by which it regulates Wnt pathway activation are complex and uncertain (38).Sclerostin is a secreted glycoprotein that inhibits Wnt pathway activation by binding to LRP5/6, a widely expressed Wnt receptor (7, 8), reducing bone formation (39, 40). Sclerostin is secreted by osteocytes (8, 41), negatively regulating bone formation by inhibiting the differentiation of osteoblasts (41, 42). SOSTi (Romosozumab) is a humanized monoclonal antibody that binds sclerostin, preventing binding to LRP5/6 and increasing Wnt pathway activation and bone formation (43). It is FDA-approved for the treatment of osteoporosis (20, 44) and has activity in rodents in addition to humans (45, 46).The discovery that osteolectin is a bone-forming growth factor raises the question of whether it mediates the effects of PTH or SOSTi on osteogenesis.     

Identifying hydrophobic protein patches to inform protein interaction interfaces

Interactions between proteins lie at the heart of numerous biological processes and are essential for the proper functioning of the cell. Although the importance of hydrophobic residues in driving protein interactions is universally accepted, a characterization of protein hydrophobicity, which informs its interactions, has remained elusive. The challenge lies in capturing the collective response of the protein hydration waters to the nanoscale chemical and topographical protein patterns, which determine protein hydrophobicity. To address this challenge, here, we employ specialized molecular simulations wherein water molecules are systematically displaced from the protein hydration shell; by identifying protein regions that relinquish their waters more readily than others, we are then able to uncover the most hydrophobic protein patches. Surprisingly, such patches contain a large fraction of polar/charged atoms and have chemical compositions that are similar to the more hydrophilic protein patches. Importantly, we also find a striking correspondence between the most hydrophobic protein patches and regions that mediate protein interactions. Our work thus establishes a computational framework for characterizing the emergent hydrophobicity of amphiphilic solutes, such as proteins, which display nanoscale heterogeneity, and for uncovering their interaction interfaces.

Protein–protein interactions play a crucial role in numerous biological processes, ranging from signal transduction and immune response to protein aggregation and phase behavior (1–3). Consequently, being able to understand, predict, and modulate protein interactions has important implications for understanding cellular processes and mitigating the progression of disease (4, 5). A necessary first step toward this ambitious goal is uncovering the interfaces through which proteins interact (6–8). In principle, identifying hydrophobic protein regions, which interact weakly with water, should be a promising strategy for uncovering protein interaction interfaces (9, 10). Indeed, the release of weakly interacting hydration waters from hydrophobic regions can drive protein interactions, as well as other aqueous assemblies (11–13). However, even when the structure of a protein is available at atomistic resolution, it is challenging to identify its hydrophobic patches because they are not uniformly nonpolar, but display variations in polarity and charge at the nanoscale. Moreover, the emergent hydrophobicity of a protein patch stems from the collective response of protein hydration waters to the nanoscale chemical and topographical patterns displayed by the patch (14–20) and cannot be captured by simply counting the number of nonpolar groups in the patch, or even through more involved additive approaches, such as hydropathy scales or surface-area models (21–28).To address this challenge, we build upon seminal theoretical advances and molecular simulation studies, which have shown that near a hydrophobic surface, it is easier to disrupt surface–water interactions and form interfacial cavities (29–34). To uncover protein regions that have the weakest interactions with water, here, we employ specialized molecular simulations, wherein protein–water interactions are disrupted by systematically displacing water molecules from the protein hydration shell (35–37). By identifying the protein patches that nucleate cavities most readily in our simulations, we are then able to uncover the most hydrophobic protein regions. Interestingly, we find that both hydrophobic and hydrophilic protein patches are highly heterogeneous and contain comparable numbers of nonpolar and polar atoms. Our results thus highlight the nontrivial relationship between the chemical composition of protein patches and their emergent hydrophobicity (24–26), and further emphasize the importance of accounting for the collective solvent response in characterizing protein hydrophobicity (16). We then interrogate whether the most hydrophobic protein patches, which nucleate cavities readily, are also likely to mediate protein interactions. Across five proteins that participate in either homodimer or heterodimer formation, we find that roughly 60 to 70% of interfacial contacts and only about 10 to 20% of noncontacts nucleate cavities. Our work thus provides a versatile computational framework for characterizing hydrophobicity and uncovering interaction interfaces of not just proteins, but also of other complex amphiphilic solutes, such as cavitands, dendrimers, and patchy nanoparticles (38–41).     

Periodic bouncing of a plasmonic bubble in a binary liquid by competing solutal and thermal Marangoni forces

The physicochemical hydrodynamics of bubbles and droplets out of equilibrium, in particular with phase transitions, display surprisingly rich and often counterintuitive phenomena. Here we experimentally and theoretically study the nucleation and early evolution of plasmonic bubbles in a binary liquid consisting of water and ethanol. Remarkably, the submillimeter plasmonic bubble is found to be periodically attracted to and repelled from the nanoparticle-decorated substrate, with frequencies of around a few kilohertz. We identify the competition between solutal and thermal Marangoni forces as the origin of the periodic bouncing. The former arises due to the selective vaporization of ethanol at the substrate’s side of the bubble, leading to a solutal Marangoni flow toward the hot substrate, which pushes the bubble away. The latter arises due to the temperature gradient across the bubble, leading to a thermal Marangoni flow away from the substrate, which sucks the bubble toward it. We study the dependence of the frequency of the bouncing phenomenon from the control parameters of the system, namely the ethanol fraction and the laser power for the plasmonic heating. Our findings can be generalized to boiling and electrolytically or catalytically generated bubbles in multicomponent liquids.

Bubbles and bubble nucleation are ubiquitous in nature and technology, e.g., in boiling, electrolysis, and catalysis, where the phenomena connected with them have tremendous relevance for energy conversion, or in flotation, sonochemistry, cavitation, ultrasonic cleaning, and biomedical applications of ultrasound and bubbles. This also includes plasmonic bubbles, i.e., bubbles nucleating at liquid-immersed metal nanoparticles under laser irradiation, due to which an enormous amount of heat is produced because of a surface plasmon resonance (1–5). For an overview on the fundamentals of bubbles and their applications we refer to our recent review article (6). In general, in these applications the bubble nucleation does not occur in a pure liquid, but in multicomponent liquids. Because of that, various additional forces and effects come into play (7), which are not relevant in pure liquids. Examples are the Soret effect (8–10) or body forces arising due to density gradients. Once the multicomponent systems have interfaces, solutal Marangoni forces (11) become relevant. The phenomena become even richer once phase transitions occur in such systems, e.g., solidification (12, 13), evaporation (14–21), or dissolution of multicomponent droplets (22–25); or nucleation of a new phase such as in the so-called ouzo effect (26, 27); or in boiling (28), electrolysis (29, 30), or catalysis (31, 32). Similarly, also chemical reactions occurring at the interface in a multicomponent liquid lead to spectacular effects, such as swimming droplets (33, 34), phoretic self-propulsion (35–38), or pattern formation in electroconvection (39). The whole field could be summarized as physicochemical hydrodynamics, and although this is a classical subject (40), it received increasing attention in recent years due to its relevance for various applications, due to new experimental and numerical possibilities, and due to the beauty of the often surprising and counterintuitive phenomena. For recent reviews on physicochemical hydrodynamics, we refer to refs. 41 and 42.To exactly analyze the various competing forces playing a role in physicochemical hydrodynamical systems, one has to strive to have simple and clean geometries, allowing for precise measurements and a theoretical and numerical approach. For example, in refs. 43 and 44 we analyzed the competition between solutal Marangoni forces, gravity, and thermal diffusion by studying an oil droplet in a stably stratified liquid consisting of ethanol and water, imposing density and surface tension gradients on the droplet. Depending on the control parameters, the droplet was either stably levitating or jumping up and down, with a very low frequency of ∼0.02 Hz. Similar droplet and bubble oscillations originating from the competition between solutal Marangoni forces and gravity were observed in ref. 45.In this paper, we report and analyze another controlled physicochemical hydrodynamic bouncing phenomenon, even involving phase transitions, namely that of a nucleating plasmonic bubble (2, 4, 5), but now in an initially homogeneous binary liquid, for which the delay of bubble nucleation after turning on the laser depends on the composition of the binary liquid and the amount of dissolved gas (46) (next, of course, to the power of the employed laser). As in refs. 43 and 44, we will again see a bouncing behavior, but this time on a much faster timescale, corresponding to frequencies of ∼103 Hz. We will use this controlled physicochemical hydrodynamic system out of equilibrium to probe the competition between solutal and thermal Marangoni forces. That, in the presence of concentration gradients, the latter can compete with the former ones only is possible because of the very high-temperature gradients in the system of a nucleating plasmonic bubble. Under more standard conditions, such as for the evaporation of a binary droplet, the solutal Marangoni forces tend to be much stronger than the thermal ones (16).We note that plasmonic bubbles are in themselves very interesting with potential applications in biomedical diagnosis and therapy, micro- and nanomanipulation, and catalysis (1, 47–49). Also note that plasmonic bubbles directly after nucleation are pure vapor bubbles (50) originating from evaporation of the surrounding liquid, but during their expansion they are invaded by dissolved gas from the surrounding liquid (46, 51–53), which in the long term crucially determines their dynamics and lifetime.The key idea of this study here will build on the selective heating of the liquid surrounding the plasmonic bubble, namely on the side of the plasmonic nanoparticles. This leads to very strong temperature gradients across the bubble and thus to thermal Marangoni forces and at the same time to strong concentration gradients, as the evaporation of the surrounding binary liquid is selective, favoring the liquid with the lower boiling point. Thus, also solutal Marangoni forces along the bubble–liquid interface emerge. As we will see, which of these two different Marangoni forces is stronger depends on time and bubble position, leading to an oscillatory or bouncing bubble behavior.