RESUMO
The COVID-19 pandemic continues to pose a substantial threat to human lives and is likely to do so for years to come. Despite the availability of vaccines, searching for efficient small-molecule drugs that are widely available, including in low- and middle-income countries, is an ongoing challenge. In this work, we report the results of an open science community effort, the "Billion molecules against COVID-19 challenge", to identify small-molecule inhibitors against SARS-CoV-2 or relevant human receptors. Participating teams used a wide variety of computational methods to screen a minimum of 1 billion virtual molecules against 6 protein targets. Overall, 31 teams participated, and they suggested a total of 639,024 molecules, which were subsequently ranked to find 'consensus compounds'. The organizing team coordinated with various contract research organizations (CROs) and collaborating institutions to synthesize and test 878 compounds for biological activity against proteases (Nsp5, Nsp3, TMPRSS2), nucleocapsid N, RdRP (only the Nsp12 domain), and (alpha) spike protein S. Overall, 27 compounds with weak inhibition/binding were experimentally identified by binding-, cleavage-, and/or viral suppression assays and are presented here. Open science approaches such as the one presented here contribute to the knowledge base of future drug discovery efforts in finding better SARS-CoV-2 treatments.
Assuntos
COVID-19 , SARS-CoV-2 , Humanos , Pandemias , Bioensaio , Descoberta de DrogasRESUMO
Molecular dynamics (MD) is the method of choice for understanding the structure, function, and interactions of molecules. However, MD simulations are limited by the strong metastability of many molecules, which traps them in a single conformation basin for an extended amount of time. Enhanced sampling techniques, such as metadynamics and replica exchange, have been developed to overcome this limitation and accelerate the exploration of complex free energy landscapes. In this paper, we propose Vendi Sampling, a replica-based algorithm for increasing the efficiency and efficacy of the exploration of molecular conformation spaces. In Vendi sampling, replicas are simulated in parallel and coupled via a global statistical measure, the Vendi Score, to enhance diversity. Vendi sampling allows for the recovery of unbiased sampling statistics and dramatically improves sampling efficiency. We demonstrate the effectiveness of Vendi sampling in improving molecular dynamics simulations by showing significant improvements in coverage and mixing between metastable states and convergence of free energy estimates for four common benchmarks, including Alanine Dipeptide and Chignolin.
RESUMO
The structure-based design of antigens holds promise for developing vaccines with higher efficacy and improved safety profiles. We postulate that abrogation of host receptor interaction bears potential for the improvement of vaccines by preventing antigen-induced modification of receptor function as well as the displacement or masking of the immunogen. Antigen modifications may yet destroy epitopes crucial for antibody neutralization. Here, we present a methodology that integrates deep mutational scans to identify and score SARS-CoV-2 receptor binding domain variants that maintain immunogenicity, but lack interaction with the widely expressed host receptor. Single point mutations were scored in silico, validated in vitro, and applied in vivo. Our top-scoring variant receptor binding domain-G502E prevented spike-induced cell-to-cell fusion, receptor internalization, and improved neutralizing antibody responses by 3.3-fold in rabbit immunizations. We name our strategy BIBAX for body-inert, B-cell-activating vaccines, which in the future may be applied beyond SARS-CoV-2 for the improvement of vaccines by design.
Assuntos
Vacinas contra COVID-19 , COVID-19 , Animais , Coelhos , Anticorpos Neutralizantes , Enzima de Conversão de Angiotensina 2/genética , SARS-CoV-2 , COVID-19/prevenção & controle , Anticorpos AntiviraisRESUMO
Machine learning provides effective computational tools for exploring the chemical space via deep generative models. Here, we propose a new reinforcement learning scheme to fine-tune graph-based deep generative models for de novo molecular design tasks. We show how our computational framework can successfully guide a pretrained generative model toward the generation of molecules with a specific property profile, even when such molecules are not present in the training set and unlikely to be generated by the pretrained model. We explored the following tasks: generating molecules of decreasing/increasing size, increasing drug-likeness, and increasing bioactivity. Using the proposed approach, we achieve a model which generates diverse compounds with predicted DRD2 activity for 95% of sampled molecules, outperforming previously reported methods on this metric.
Assuntos
Desenho de Fármacos , Aprendizado de MáquinaRESUMO
With recent advances in structural biology, including experimental techniques and deep learning-enabled high-precision structure predictions, molecular dynamics methods that scale up to large biomolecular systems are required. Current state-of-the-art approaches in molecular dynamics modeling focus on encoding global configurations of molecular systems as distinct states. This paradigm commands us to map out all possible structures and sample transitions between them, a task that becomes impossible for large-scale systems such as biomolecular complexes. To arrive at scalable molecular models, we suggest moving away from global state descriptions to a set of coupled models that each describe the dynamics of local domains or sites of the molecular system. We describe limitations in the current state-of-the-art global-state Markovian modeling approaches and then introduce Markov field models as an umbrella term that includes models from various scientific communities, including Independent Markov decomposition, Ising and Potts models, and (dynamic) graphical models, and evaluate their use for computational molecular biology. Finally, we give a few examples of early adoptions of these ideas for modeling molecular kinetics and thermodynamics.
Assuntos
Simulação de Dinâmica Molecular , Física , Cadeias de Markov , Cinética , TermodinâmicaRESUMO
Partner recognition in protein binding is critical for all biological functions, and yet, delineating its mechanism is challenging, especially when recognition happens within microseconds. We present a theoretical and experimental framework based on straight-forward nuclear magnetic resonance relaxation dispersion measurements to investigate protein binding mechanisms on sub-millisecond timescales, which are beyond the reach of standard rapid-mixing experiments. This framework predicts that conformational selection prevails on ubiquitin's paradigmatic interaction with an SH3 (Src-homology 3) domain. By contrast, the SH3 domain recognizes ubiquitin in a two-state binding process. Subsequent molecular dynamics simulations and Markov state modeling reveal that the ubiquitin conformation selected for binding exhibits a characteristically extended C-terminus. Our framework is robust and expandable for implementation in other binding scenarios with the potential to show that conformational selection might be the design principle of the hubs in protein interaction networks.
Assuntos
Proteínas de Transporte , Domínios de Homologia de src , Proteínas de Transporte/metabolismo , Ligação Proteica , Conformação Proteica , Ubiquitina/metabolismoRESUMO
Biomolecular spin relaxation processes, such as the NOE, are commonly modeled by rotational τc-tumbling combined with fast motions on the sub-τc timescale. Motions on the supra-τc timescale, in contrast, are considered to be completely decorrelated to the molecular tumbling and therefore invisible. Here, we show how supra-τc dynamics can nonetheless influence the NOE build-up between methyl groups. This effect arises because supra-τc motions can cluster the fast-motion ensembles into discrete states, affecting distance averaging as well as the fast-motion order parameter and hence the cross-relaxation rate. We present a computational approach to estimate methyl-methyl cross-relaxation rates from extensive (>100×τc) all-atom molecular dynamics (MD) trajectories on the example of the 723-residue protein Malate Synthase G. The approach uses Markov state models (MSMs) to resolve transitions between metastable states and thus to discriminate between sub-τc and supra-τc conformational exchange. We find that supra-τc exchange typically increases NOESY cross-peak intensities. The methods described in this work extend the theory of modeling sub-µs dynamics in spin relaxation and thus contribute to a quantitative estimation of NOE cross-relaxation rates from MD simulations, eventually leading to increased precision in structural and functional studies of large proteins.
Assuntos
Simulação de Dinâmica Molecular , Proteínas , Análise por Conglomerados , Espectroscopia de Ressonância Magnética/métodos , Movimento (Física) , Ressonância Magnética Nuclear Biomolecular , Proteínas/químicaRESUMO
We present machine learning models for predicting the chemical context for Buchwald-Hartwig coupling reactions, i. e., what chemicals to add to the reactants to give a productive reaction. Using reaction data from in-house electronic lab notebooks, we train two models: one based on single-label data and one based on multi-label data. Both models show excellent top-3 accuracy of approximately 90 %, which suggests strong predictivity. Furthermore, there seems to be an advantage of including multi-label data because the multi-label model shows higher accuracy and better sensitivity for the individual contexts than the single-label model. Although the models are performant, we also show that such models need to be re-trained periodically as there is a strong temporal characteristic to the usage of different contexts. Therefore, a model trained on historical data will decrease in usefulness with time as newer and better contexts emerge and replace older ones. We hypothesize that such significant transitions in the context-usage will likely affect any model predicting chemical contexts trained on historical data. Consequently, training context prediction models warrants careful planning of what data is used for training and how often the model needs to be re-trained.
Assuntos
Aprendizado de MáquinaRESUMO
BACKGROUND: Prevalence for knee osteoarthritis is rising in both Sweden and globally due to increased age and obesity in the population. This has subsequently led to an increasing demand for knee arthroplasties. Correct diagnosis and classification of a knee osteoarthritis (OA) are therefore of a great interest in following-up and planning for either conservative or operative management. Most orthopedic surgeons rely on standard weight bearing radiographs of the knee. Improving the reliability and reproducibility of these interpretations could thus be hugely beneficial. Recently, deep learning which is a form of artificial intelligence (AI), has been showing promising results in interpreting radiographic images. In this study, we aim to evaluate how well an AI can classify the severity of knee OA, using entire image series and not excluding common visual disturbances such as an implant, cast and non-degenerative pathologies. METHODS: We selected 6103 radiographic exams of the knee taken at Danderyd University Hospital between the years 2002-2016 and manually categorized them according to the Kellgren & Lawrence grading scale (KL). We then trained a convolutional neural network (CNN) of ResNet architecture using PyTorch. We evaluated the results against a test set of 300 exams that had been reviewed independently by two senior orthopedic surgeons who settled eventual interobserver disagreements through consensus sessions. RESULTS: The CNN yielded an overall AUC of more than 0.87 for all KL grades except KL grade 2, which yielded an AUC of 0.8 and a mean AUC of 0.92. When merging adjacent KL grades, all but one group showed near perfect results with AUC > 0.95 indicating excellent performance. CONCLUSION: We have found that we could teach a CNN to correctly diagnose and classify the severity of knee OA using the KL grading system without cleaning the input data from major visual disturbances such as implants and other pathologies.
Assuntos
Aprendizado Profundo , Osteoartrite do Joelho , Adulto , Inteligência Artificial , Humanos , Articulação do Joelho , Osteoartrite do Joelho/diagnóstico por imagem , Osteoartrite do Joelho/epidemiologia , Osteoartrite do Joelho/cirurgia , Reprodutibilidade dos TestesRESUMO
The use of coarse-grained (CG) models is a popular approach to study complex biomolecular systems. By reducing the number of degrees of freedom, a CG model can explore long time- and length-scales inaccessible to computational models at higher resolution. If a CG model is designed by formally integrating out some of the system's degrees of freedom, one expects multi-body interactions to emerge in the effective CG model's energy function. In practice, it has been shown that the inclusion of multi-body terms indeed improves the accuracy of a CG model. However, no general approach has been proposed to systematically construct a CG effective energy that includes arbitrary orders of multi-body terms. In this work, we propose a neural network based approach to address this point and construct a CG model as a multi-body expansion. By applying this approach to a small protein, we evaluate the relative importance of the different multi-body terms in the definition of an accurate model. We observe a slow convergence in the multi-body expansion, where up to five-body interactions are needed to reproduce the free energy of an atomistic model.
Assuntos
Oligopeptídeos/química , Simulação de Dinâmica Molecular , Redes Neurais de Computação , TermodinâmicaRESUMO
BACKGROUND: Fractures around the knee joint are inherently complex in terms of treatment; complication rates are high, and they are difficult to diagnose on a plain radiograph. An automated way of classifying radiographic images could improve diagnostic accuracy and would enable production of uniformly classified records of fractures to be used in researching treatment strategies for different fracture types. Recently deep learning, a form of artificial intelligence (AI), has shown promising results for interpreting radiographs. In this study, we aim to evaluate how well an AI can classify knee fractures according to the detailed 2018 AO-OTA fracture classification system. METHODS: We selected 6003 radiograph exams taken at Danderyd University Hospital between the years 2002-2016, and manually categorized them according to the AO/OTA classification system and by custom classifiers. We then trained a ResNet-based neural network on this data. We evaluated the performance against a test set of 600 exams. Two senior orthopedic surgeons had reviewed these exams independently where we settled exams with disagreement through a consensus session. RESULTS: We captured a total of 49 nested fracture classes. Weighted mean AUC was 0.87 for proximal tibia fractures, 0.89 for patella fractures and 0.89 for distal femur fractures. Almost ¾ of AUC estimates were above 0.8, out of which more than half reached an AUC of 0.9 or above indicating excellent performance. CONCLUSION: Our study shows that neural networks can be used not only for fracture identification but also for more detailed classification of fractures around the knee joint.
Assuntos
Inteligência Artificial , Fraturas do Fêmur/diagnóstico por imagem , Processamento de Imagem Assistida por Computador/métodos , Fraturas da Tíbia/diagnóstico por imagem , HumanosRESUMO
BACKGROUND: Antivirals are needed to combat the COVID-19 pandemic, which is caused by SARS-CoV-2. The clinically-proven protease inhibitor Camostat mesylate inhibits SARS-CoV-2 infection by blocking the virus-activating host cell protease TMPRSS2. However, antiviral activity of Camostat mesylate metabolites and potential viral resistance have not been analyzed. Moreover, antiviral activity of Camostat mesylate in human lung tissue remains to be demonstrated. METHODS: We used recombinant TMPRSS2, reporter particles bearing the spike protein of SARS-CoV-2 or authentic SARS-CoV-2 to assess inhibition of TMPRSS2 and viral entry, respectively, by Camostat mesylate and its metabolite GBPA. FINDINGS: We show that several TMPRSS2-related proteases activate SARS-CoV-2 and that two, TMPRSS11D and TMPRSS13, are robustly expressed in the upper respiratory tract. However, entry mediated by these proteases was blocked by Camostat mesylate. The Camostat metabolite GBPA inhibited recombinant TMPRSS2 with reduced efficiency as compared to Camostat mesylate. In contrast, both inhibitors exhibited similar antiviral activity and this correlated with the rapid conversion of Camostat mesylate into GBPA in the presence of serum. Finally, Camostat mesylate and GBPA blocked SARS-CoV-2 spread in human lung tissue ex vivo and the related protease inhibitor Nafamostat mesylate exerted augmented antiviral activity. INTERPRETATION: Our results suggest that SARS-CoV-2 can use TMPRSS2 and closely related proteases for spread in the upper respiratory tract and that spread in the human lung can be blocked by Camostat mesylate and its metabolite GBPA. FUNDING: NIH, Damon Runyon Foundation, ACS, NYCT, DFG, EU, Berlin Mathematics center MATH+, BMBF, Lower Saxony, Lundbeck Foundation, Novo Nordisk Foundation.
Assuntos
Antivirais/farmacologia , Tratamento Farmacológico da COVID-19 , Ésteres/farmacologia , Guanidinas/farmacologia , Inibidores de Proteases/farmacologia , SARS-CoV-2/efeitos dos fármacos , Serina Endopeptidases/metabolismo , Animais , Linhagem Celular , Chlorocebus aethiops , Cricetinae , Células HEK293 , Humanos , Pulmão/patologia , Pulmão/virologia , Proteínas de Membrana/biossíntese , Simulação de Dinâmica Molecular , Serina Endopeptidases/biossíntese , Serina Proteases/biossíntese , Células Vero , Ativação Viral/efeitos dos fármacos , Internalização do Vírus/efeitos dos fármacosRESUMO
Bromodomains (BDs) are small protein modules that interact with acetylated marks in histones. These posttranslational modifications are pivotal to regulate gene expression, making BDs promising targets to treat several diseases. While the general structure of BDs is well known, their dynamical features and their interplay with other macromolecules are poorly understood, hampering the rational design of potent and selective inhibitors. Here, we combine extensive molecular dynamics simulations, Markov state modeling, and available structural data to reveal a transiently formed state that is conserved across all BD families. It involves the breaking of two backbone hydrogen bonds that anchor the ZA-loop with the αA helix, opening a cryptic pocket that partially occludes the one associated to histone binding. By analyzing more than 1,900 experimental structures, we unveil just two adopting the hidden state, explaining why it has been previously unnoticed and providing direct structural evidence for its existence. Our results suggest that this state is an allosteric regulatory switch for BDs, potentially related to a recently unveiled BD-DNA-binding mode.
Assuntos
Proteínas de Ciclo Celular/química , Proteínas Correpressoras/química , Proteínas de Ligação a DNA/química , Histona Acetiltransferases/química , Peptídeos e Proteínas de Sinalização Intracelular/química , Fatores Genéricos de Transcrição/química , Fatores de Transcrição/química , Proteína 28 com Motivo Tripartido/química , Sequência de Aminoácidos , Sítios de Ligação , Proteínas de Ciclo Celular/genética , Proteínas de Ciclo Celular/metabolismo , Proteínas Correpressoras/genética , Proteínas Correpressoras/metabolismo , Cristalografia por Raios X , DNA/química , DNA/genética , DNA/metabolismo , Proteínas de Ligação a DNA/genética , Proteínas de Ligação a DNA/metabolismo , Regulação da Expressão Gênica , Histona Acetiltransferases/genética , Histona Acetiltransferases/metabolismo , Humanos , Peptídeos e Proteínas de Sinalização Intracelular/genética , Peptídeos e Proteínas de Sinalização Intracelular/metabolismo , Cadeias de Markov , Simulação de Dinâmica Molecular , Ligação Proteica , Conformação Proteica em alfa-Hélice , Conformação Proteica em Folha beta , Domínios e Motivos de Interação entre Proteínas , Alinhamento de Sequência , Homologia de Sequência de Aminoácidos , Termodinâmica , Fatores de Transcrição/genética , Fatores de Transcrição/metabolismo , Fatores Genéricos de Transcrição/genética , Fatores Genéricos de Transcrição/metabolismo , Proteína 28 com Motivo Tripartido/genética , Proteína 28 com Motivo Tripartido/metabolismoRESUMO
The entry of the coronavirus SARS-CoV-2 into human lung cells can be inhibited by the approved drugs camostat and nafamostat. Here we elucidate the molecular mechanism of these drugs by combining experiments and simulations. In vitro assays confirm that both drugs inhibit the human protein TMPRSS2, a SARS-Cov-2 spike protein activator. As no experimental structure is available, we provide a model of the TMPRSS2 equilibrium structure and its fluctuations by relaxing an initial homology structure with extensive 330 microseconds of all-atom molecular dynamics (MD) and Markov modeling. Through Markov modeling, we describe the binding process of both drugs and a metabolic product of camostat (GBPA) to TMPRSS2, reaching a Michaelis complex (MC) state, which precedes the formation of a long-lived covalent inhibitory state. We find that nafamostat has a higher MC population than camostat and GBPA, suggesting that nafamostat is more readily available to form the stable covalent enzyme-substrate intermediate, effectively explaining its high potency. This model is backed by our in vitro experiments and consistent with previous virus cell entry assays. Our TMPRSS2-drug structures are made public to guide the design of more potent and specific inhibitors.
RESUMO
Coarse graining enables the investigation of molecular dynamics for larger systems and at longer timescales than is possible at an atomic resolution. However, a coarse graining model must be formulated such that the conclusions we draw from it are consistent with the conclusions we would draw from a model at a finer level of detail. It has been proved that a force matching scheme defines a thermodynamically consistent coarse-grained model for an atomistic system in the variational limit. Wang et al. [ACS Cent. Sci. 5, 755 (2019)] demonstrated that the existence of such a variational limit enables the use of a supervised machine learning framework to generate a coarse-grained force field, which can then be used for simulation in the coarse-grained space. Their framework, however, requires the manual input of molecular features to machine learn the force field. In the present contribution, we build upon the advance of Wang et al. and introduce a hybrid architecture for the machine learning of coarse-grained force fields that learn their own features via a subnetwork that leverages continuous filter convolutions on a graph neural network architecture. We demonstrate that this framework succeeds at reproducing the thermodynamics for small biomolecular systems. Since the learned molecular representations are inherently transferable, the architecture presented here sets the stage for the development of machine-learned, coarse-grained force fields that are transferable across molecular systems.
RESUMO
Protein allostery is a phenomenon involving the long range coupling between two distal sites in a protein. In order to elucidate allostery at atomic resoluion on the ligand-binding WW domain of the enzyme Pin1, multistate structures were calculated from exact nuclear Overhauser effect (eNOE). In its free form, the protein undergoes a microsecond exchange between two states, one of which is predisposed to interact with its parent catalytic domain. In presence of the positive allosteric ligand, the equilibrium between the two states is shifted towards domain-domain interaction, suggesting a population shift model. In contrast, the allostery-suppressing ligand decouples the side-chain arrangement at the inter-domain interface thereby reducing the inter-domain interaction. As such, this mechanism is an example of dynamic allostery. The presented distinct modes of action highlight the power of the interplay between dynamics and function in the biological activity of proteins.
Assuntos
Peptidilprolil Isomerase de Interação com NIMA/metabolismo , Regulação Alostérica , Humanos , Modelos Moleculares , Peptidilprolil Isomerase de Interação com NIMA/químicaRESUMO
Antiviral therapy is urgently needed to combat the coronavirus disease 2019 (COVID-19) pandemic, which is caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The protease inhibitor camostat mesylate inhibits SARS-CoV-2 infection of lung cells by blocking the virus-activating host cell protease TMPRSS2. Camostat mesylate has been approved for treatment of pancreatitis in Japan and is currently being repurposed for COVID-19 treatment. However, potential mechanisms of viral resistance as well as camostat mesylate metabolization and antiviral activity of metabolites are unclear. Here, we show that SARS-CoV-2 can employ TMPRSS2-related host cell proteases for activation and that several of them are expressed in viral target cells. However, entry mediated by these proteases was blocked by camostat mesylate. The camostat metabolite GBPA inhibited the activity of recombinant TMPRSS2 with reduced efficiency as compared to camostat mesylate and was rapidly generated in the presence of serum. Importantly, the infection experiments in which camostat mesylate was identified as a SARS-CoV-2 inhibitor involved preincubation of target cells with camostat mesylate in the presence of serum for 2 h and thus allowed conversion of camostat mesylate into GBPA. Indeed, when the antiviral activities of GBPA and camostat mesylate were compared in this setting, no major differences were identified. Our results indicate that use of TMPRSS2-related proteases for entry into target cells will not render SARS-CoV-2 camostat mesylate resistant. Moreover, the present and previous findings suggest that the peak concentrations of GBPA established after the clinically approved camostat mesylate dose (600 mg/day) will result in antiviral activity.
RESUMO
Computing equilibrium states in condensed-matter many-body systems, such as solvated proteins, is a long-standing challenge. Lacking methods for generating statistically independent equilibrium samples in "one shot," vast computational effort is invested for simulating these systems in small steps, e.g., using molecular dynamics. Combining deep learning and statistical mechanics, we developed Boltzmann generators, which are shown to generate unbiased one-shot equilibrium samples of representative condensed-matter systems and proteins. Boltzmann generators use neural networks to learn a coordinate transformation of the complex configurational equilibrium distribution to a distribution that can be easily sampled. Accurate computation of free-energy differences and discovery of new configurations are demonstrated, providing a statistical mechanics tool that can avoid rare events during sampling without prior knowledge of reaction coordinates.
RESUMO
Most current molecular dynamics simulation and analysis methods rely on the idea that the molecular system can be represented by a single global state (e.g., a Markov state in a Markov state model [MSM]). In this approach, molecules can be extensively sampled and analyzed when they only possess a few metastable states, such as small- to medium-sized proteins. However, this approach breaks down in frustrated systems and in large protein assemblies, where the number of global metastable states may grow exponentially with the system size. To address this problem, we here introduce dynamic graphical models (DGMs) that describe molecules as assemblies of coupled subsystems, akin to how spins interact in the Ising model. The change of each subsystem state is only governed by the states of itself and its neighbors. DGMs require fewer parameters than MSMs or other global state models; in particular, we do not need to observe all global system configurations to characterize them. Therefore, DGMs can predict previously unobserved molecular configurations. As a proof of concept, we demonstrate that DGMs can faithfully describe molecular thermodynamics and kinetics and predict previously unobserved metastable states for Ising models and protein simulations.
Assuntos
Simulação de Dinâmica Molecular , Proteínas/química , Cinética , Cadeias de Markov , Conformação Proteica , TermodinâmicaRESUMO
Atomistic or ab initio molecular dynamics simulations are widely used to predict thermodynamics and kinetics and relate them to molecular structure. A common approach to go beyond the time- and length-scales accessible with such computationally expensive simulations is the definition of coarse-grained molecular models. Existing coarse-graining approaches define an effective interaction potential to match defined properties of high-resolution models or experimental data. In this paper, we reformulate coarse-graining as a supervised machine learning problem. We use statistical learning theory to decompose the coarse-graining error and cross-validation to select and compare the performance of different models. We introduce CGnets, a deep learning approach, that learns coarse-grained free energy functions and can be trained by a force-matching scheme. CGnets maintain all physically relevant invariances and allow one to incorporate prior physics knowledge to avoid sampling of unphysical structures. We show that CGnets can capture all-atom explicit-solvent free energy surfaces with models using only a few coarse-grained beads and no solvent, while classical coarse-graining methods fail to capture crucial features of the free energy surface. Thus, CGnets are able to capture multibody terms that emerge from the dimensionality reduction.