Discriminating physiological from non-physiological interfaces in structures of protein complexes: A community-wide study.

Reliably scoring and ranking candidate models of protein complexes and assigning their oligomeric state from the structure of the crystal lattice represent outstanding challenges. A community-wide effort was launched to tackle these challenges. The latest resources on protein complexes and interfaces were exploited to derive a benchmark dataset consisting of 1677 homodimer protein crystal structures, including a balanced mix of physiological and non-physiological complexes. The non-physiological complexes in the benchmark were selected to bury a similar or larger interface area than their physiological counterparts, making it more difficult for scoring functions to differentiate between them. Next, 252 functions for scoring protein-protein interfaces previously developed by 13 groups were collected and evaluated for their ability to discriminate between physiological and non-physiological complexes. A simple consensus score generated using the best performing score of each of the 13 groups, and a cross-validated Random Forest (RF) classifier were created. Both approaches showed excellent performance, with an area under the Receiver Operating Characteristic (ROC) curve of 0.93 and 0.94, respectively, outperforming individual scores developed by different groups. Additionally, AlphaFold2 engines recalled the physiological dimers with significantly higher accuracy than the non-physiological set, lending support to the reliability of our benchmark dataset annotations. Optimizing the combined power of interface scoring functions and evaluating it on challenging benchmark datasets appears to be a promising strategy.

EvoRator2: Predicting Site-specific Amino Acid Substitutions Based on Protein Structural Information Using Deep Learning.

Multiple sequence alignments (MSAs) are the workhorse of molecular evolution and structural biology research. From MSAs, the amino acids that are tolerated at each site during protein evolution can be inferred. However, little is known regarding the repertoire of tolerated amino acids in proteins when only a few or no sequence homologs are available, such as orphan and de novo designed proteins. Here we present EvoRator2, a deep-learning algorithm trained on over 15,000 protein structures that can predict which amino acids are tolerated at any given site, based exclusively on protein structural information mined from atomic coordinate files. We show that EvoRator2 obtained satisfying results for the prediction of position-weighted scoring matrices (PSSM). We further show that EvoRator2 obtained near state-of-the-art performance on proteins with high quality structures in predicting the effect of mutations in deep mutation scanning (DMS) experiments and that for certain DMS targets, EvoRator2 outperformed state-of-the-art methods. We also show that by combining EvoRator2's predictions with those obtained by a state-of-the-art deep-learning method that accounts for the information in the MSA, the prediction of the effect of mutation in DMS experiments was improved in terms of both accuracy and stability. EvoRator2 is designed to predict which amino-acid substitutions are tolerated in such proteins without many homologous sequences, including orphan or de novo designed proteins. We implemented our approach in the EvoRator web server (https://evorator.tau.ac.il).

Machine learning for evolutionary-based and physics-inspired protein design: Current and future synergies.

Computational protein design facilitates the discovery of novel proteins with prescribed structure and functionality. Exciting designs were recently reported using novel data-driven methodologies that can be roughly divided into two categories: evolutionary-based and physics-inspired approaches. The former infer characteristic sequence features shared by sets of evolutionary-related proteins, such as conserved or coevolving positions, and recombine them to generate candidates with similar structure and function. The latter approaches estimate key biochemical properties, such as structure free energy, conformational entropy, or binding affinities using machine learning surrogates, and optimize them to yield improved designs. Here, we review recent progress along both tracks, discuss their strengths and weaknesses, and highlight opportunities for synergistic approaches.

Funneling modulatory peptide design with generative models: Discovery and characterization of disruptors of calcineurin protein-protein interactions.

Design of peptide binders is an attractive strategy for targeting "undruggable" protein-protein interfaces. Current design protocols rely on the extraction of an initial sequence from one known protein interactor of the target protein, followed by in-silico or in-vitro mutagenesis-based optimization of its binding affinity. Wet lab protocols can explore only a minor portion of the vast sequence space and cannot efficiently screen for other desirable properties such as high specificity and low toxicity, while in-silico design requires intensive computational resources and often relies on simplified binding models. Yet, for a multivalent protein target, dozens to hundreds of natural protein partners already exist in the cellular environment. Here, we describe a peptide design protocol that harnesses this diversity via a machine learning generative model. After identifying putative natural binding fragments by literature and homology search, a compositional Restricted Boltzmann Machine is trained and sampled to yield hundreds of diverse candidate peptides. The latter are further filtered via flexible molecular docking and an in-vitro microchip-based binding assay. We validate and test our protocol on calcineurin, a calcium-dependent protein phosphatase involved in various cellular pathways in health and disease. In a single screening round, we identified multiple 16-length peptides with up to six mutations from their closest natural sequence that successfully interfere with the binding of calcineurin to its substrates. In summary, integrating protein interaction and sequence databases, generative modeling, molecular docking and interaction assays enables the discovery of novel protein-protein interaction modulators.

Neural assemblies uncovered by generative modeling explain whole-brain activity statistics and reflect structural connectivity.

Patterns of endogenous activity in the brain reflect a stochastic exploration of the neuronal state space that is constrained by the underlying assembly organization of neurons. Yet, it remains to be shown that this interplay between neurons and their assembly dynamics indeed suffices to generate whole-brain data statistics. Here, we recorded the activity from ∼40,000 neurons simultaneously in zebrafish larvae, and show that a data-driven generative model of neuron-assembly interactions can accurately reproduce the mean activity and pairwise correlation statistics of their spontaneous activity. This model, the compositional Restricted Boltzmann Machine (cRBM), unveils ∼200 neural assemblies, which compose neurophysiological circuits and whose various combinations form successive brain states. We then performed in silico perturbation experiments to determine the interregional functional connectivity, which is conserved across individual animals and correlates well with structural connectivity. Our results showcase how cRBMs can capture the coarse-grained organization of the zebrafish brain. Notably, this generative model can readily be deployed to parse neural data obtained by other large-scale recording techniques.

Reduced B cell antigenicity of Omicron lowers host serologic response.

The SARS-CoV-2 Omicron variant evades most neutralizing vaccine-induced antibodies and is associated with lower antibody titers upon breakthrough infections than previous variants. However, the mechanism remains unclear. Here, we find using a geometric deep-learning model that Omicron's extensively mutated receptor binding site (RBS) features reduced antigenicity compared with previous variants. Mice immunization experiments with different recombinant receptor binding domain (RBD) variants confirm that the serological response to Omicron is drastically attenuated and less potent. Analyses of serum cross-reactivity and competitive ELISA reveal a reduction in antibody response across both variable and conserved RBD epitopes. Computational modeling confirms that the RBS has a potential for further antigenicity reduction while retaining efficient receptor binding. Finally, we find a similar trend of antigenicity reduction over decades for hCoV229E, a common cold coronavirus. Thus, our study explains the reduced antibody titers associated with Omicron infection and reveals a possible trajectory of future viral evolution.

ScanNet: A Web Server for Structure-based Prediction of Protein Binding Sites with Geometric Deep Learning.

Predicting the various binding sites of a protein from its structure sheds light on its function and paves the way towards design of interaction inhibitors. Here, we report ScanNet, a freely available web server for prediction of protein-protein, protein - disordered protein and protein - antibody binding sites from structure. ScanNet (Spatio-Chemical Arrangement of Neighbors Network) is an end-to-end, interpretable geometric deep learning model that learns spatio-chemical patterns directly from 3D structures. ScanNet consistently outperforms Machine Learning models based on handcrafted features and comparative modeling approaches. The web server is linked to both the PDB and AlphaFoldDB, and supports user-provided structure files. Predictions can be readily visualized on the website via the Molstar web app and locally via ChimeraX. ScanNet is available at http://bioinfo3d.cs.tau.ac.il/ScanNet/.

Superimmunity by pan-sarbecovirus nanobodies.

Vaccine boosters and infection can facilitate the development of SARS-CoV-2 antibodies with improved potency and breadth. Here, we observe superimmunity in a camelid extensively immunized with the SARS-CoV-2 receptor-binding domain (RBD). We rapidly isolate a large repertoire of specific ultra-high-affinity nanobodies that bind strongly to all known sarbecovirus clades using integrative proteomics. These pan-sarbecovirus nanobodies (psNbs) are highly effective against SARS-CoV and SARS-CoV-2 variants, including Omicron, with the best median neutralization potency at single-digit nanograms per milliliter. A highly potent, inhalable, and bispecific psNb (PiN-31) is also developed. Structural determinations of 13 psNbs with the SARS-CoV-2 spike or RBD reveal five epitope classes, providing insights into the mechanisms and evolution of their broad activities. The highly evolved psNbs target small, flat, and flexible epitopes that contain over 75% of conserved RBD surface residues. Their potencies are strongly and negatively correlated with the distance of the epitopes from the receptor binding sites.

ScanNet: an interpretable geometric deep learning model for structure-based protein binding site prediction.

Predicting the functional sites of a protein from its structure, such as the binding sites of small molecules, other proteins or antibodies, sheds light on its function in vivo. Currently, two classes of methods prevail: machine learning models built on top of handcrafted features and comparative modeling. They are, respectively, limited by the expressivity of the handcrafted features and the availability of similar proteins. Here, we introduce ScanNet, an end-to-end, interpretable geometric deep learning model that learns features directly from 3D structures. ScanNet builds representations of atoms and amino acids based on the spatio-chemical arrangement of their neighbors. We train ScanNet for detecting protein-protein and protein-antibody binding sites, demonstrate its accuracy-including for unseen protein folds-and interpret the filters learned. Finally, we predict epitopes of the SARS-CoV-2 spike protein, validating known antigenic regions and predicting previously uncharacterized ones. Overall, ScanNet is a versatile, powerful and interpretable model suitable for functional site prediction tasks. A webserver for ScanNet is available from http://bioinfo3d.cs.tau.ac.il/ScanNet/ .

Reduced antigenicity of Omicron lowers host serologic response.

SARS-CoV-2 Omicron variant of concern (VOC) contains fifteen mutations on the receptor binding domain (RBD), evading most neutralizing antibodies from vaccinated sera. Emerging evidence suggests that Omicron breakthrough cases are associated with substantially lower antibody titers than other VOC cases. However, the mechanism remains unclear. Here, using a novel geometric deep-learning model, we discovered that the antigenic profile of Omicron RBD is distinct from the prior VOCs, featuring reduced antigenicity in its remodeled receptor binding sites (RBS). To substantiate our deep-learning prediction, we immunized mice with different recombinant RBD variants and found that the Omicron's extensive mutations can lead to a drastically attenuated serologic response with limited neutralizing activity in vivo , while the T cell response remains potent. Analyses of serum cross-reactivity and competitive ELISA with epitope-specific nanobodies revealed that the antibody response to Omicron was reduced across RBD epitopes, including both the variable RBS and epitopes without any known VOC mutations. Moreover, computational modeling confirmed that the RBS is highly versatile with a capacity to further decrease antigenicity while retaining efficient receptor binding. Longitudinal analysis showed that this evolutionary trend of decrease in antigenicity was also found in hCoV229E, a common cold coronavirus that has been circulating in humans for decades. Thus, our study provided unprecedented insights into the reduced antibody titers associated with Omicron infection, revealed a possible trajectory of future viral evolution and may inform the vaccine development against future outbreaks.

RBM-MHC: A Semi-Supervised Machine-Learning Method for Sample-Specific Prediction of Antigen Presentation by HLA-I Alleles.

The recent increase of immunopeptidomics data, obtained by mass spectrometry or binding assays, opens up possibilities for investigating endogenous antigen presentation by the highly polymorphic human leukocyte antigen class I (HLA-I) protein. State-of-the-art methods predict with high accuracy presentation by HLA alleles that are well represented in databases at the time of release but have a poorer performance for rarer and less characterized alleles. Here, we introduce a method based on Restricted Boltzmann Machines (RBMs) for prediction of antigens presented on the Major Histocompatibility Complex (MHC) encoded by HLA genes-RBM-MHC. RBM-MHC can be trained on custom and newly available samples with no or a small amount of HLA annotations. RBM-MHC ensures improved predictions for rare alleles and matches state-of-the-art performance for well-characterized alleles while being less data demanding. RBM-MHC is shown to be a flexible and easily interpretable method that can be used as a predictor of cancer neoantigens and viral epitopes, as a tool for feature discovery, and to reconstruct peptide motifs presented on specific HLA molecules.

Blind deconvolution for spike inference from fluorescence recordings.

The parallel developments of genetically-encoded calcium indicators and fast fluorescence imaging techniques allows one to simultaneously record neural activity of extended neuronal populations in vivo. To fully harness the potential of functional imaging, one needs to infer the sequence of action potentials from fluorescence traces. Here we build on recently proposed computational approaches to develop a blind sparse deconvolution (BSD) algorithm based on a generative model for inferring spike trains from fluorescence traces. BSD features, (1) automatic (fully unsupervised) estimation of the hyperparameters, such as spike amplitude, noise level and rise and decay time constants, (2) a novel analytical estimate of the sparsity prior, which yields enhanced robustness and computational speed with respect to existing methods, (3) automatic thresholding for binarizing spikes that maximizes the precision-recall performance, (4) super-resolution capabilities increasing the temporal resolution beyond the fluorescence signal acquisition rate. BSD also uniquely provides theoretically-grounded estimates of the expected performance of the spike reconstruction in terms of precision-recall and temporal accuracy for each recording. The performance of the algorithm is established using synthetic data and through the SpikeFinder challenge, a community-based initiative for spike-rate inference benchmarking based on a collection of joint electrophysiological and fluorescence recordings. Our method outperforms classical sparse deconvolution algorithms in terms of robustness, speed and/or accuracy and performs competitively in the SpikeFinder challenge. This algorithm is modular, easy-to-use and made freely available. Its novel features can thus be incorporated in a straightforward way into existing calcium imaging packages.

Inference of compressed Potts graphical models.

We consider the problem of inferring a graphical Potts model on a population of variables. This inverse Potts problem generally involves the inference of a large number of parameters, often larger than the number of available data, and, hence, requires the introduction of regularization. We study here a double regularization scheme, in which the number of Potts states (colors) available to each variable is reduced and interaction networks are made sparse. To achieve the color compression, only Potts states with large empirical frequency (exceeding some threshold) are explicitly modeled on each site, while the others are grouped into a single state. We benchmark the performances of this mixed regularization approach, with two inference algorithms, adaptive cluster expansion (ACE) and pseudolikelihood maximization (PLM), on synthetic data obtained by sampling disordered Potts models on Erdos-Rényi random graphs. We show in particular that color compression does not affect the quality of reconstruction of the parameters corresponding to high-frequency symbols, while drastically reducing the number of the other parameters and thus the computational time. Our procedure is also applied to multisequence alignments of protein families, with similar results.

Learning Compositional Representations of Interacting Systems with Restricted Boltzmann Machines: Comparative Study of Lattice Proteins.

A restricted Boltzmann machine (RBM) is an unsupervised machine learning bipartite graphical model that jointly learns a probability distribution over data and extracts their relevant statistical features. RBMs were recently proposed for characterizing the patterns of coevolution between amino acids in protein sequences and for designing new sequences. Here, we study how the nature of the features learned by RBM changes with its defining parameters, such as the dimensionality of the representations (size of the hidden layer) and the sparsity of the features. We show that for adequate values of these parameters, RBMs operate in a so-called compositional phase in which visible configurations sampled from the RBM are obtained by recombining these features. We then compare the performance of RBM with other standard representation learning algorithms, including principal or independent component analysis (PCA, ICA), autoencoders (AE), variational autoencoders (VAE), and their sparse variants. We show that RBMs, due to the stochastic mapping between data configurations and representations, better capture the underlying interactions in the system and are significantly more robust with respect to sample size than deterministic methods such as PCA or ICA. In addition, this stochastic mapping is not prescribed a priori as in VAE, but learned from data, which allows RBMs to show good performance even with shallow architectures. All numerical results are illustrated on synthetic lattice protein data that share similar statistical features with real protein sequences and for which ground-truth interactions are known.

Learning protein constitutive motifs from sequence data.

Statistical analysis of evolutionary-related protein sequences provides information about their structure, function, and history. We show that Restricted Boltzmann Machines (RBM), designed to learn complex high-dimensional data and their statistical features, can efficiently model protein families from sequence information. We here apply RBM to 20 protein families, and present detailed results for two short protein domains (Kunitz and WW), one long chaperone protein (Hsp70), and synthetic lattice proteins for benchmarking. The features inferred by the RBM are biologically interpretable: they are related to structure (residue-residue tertiary contacts, extended secondary motifs (α-helixes and ß-sheets) and intrinsically disordered regions), to function (activity and ligand specificity), or to phylogenetic identity. In addition, we use RBM to design new protein sequences with putative properties by composing and 'turning up' or 'turning down' the different modes at will. Our work therefore shows that RBM are versatile and practical tools that can be used to unveil and exploit the genotype-phenotype relationship for protein families.