SUMMARY: Recently, deep learning models, initially developed in the field of natural language processing (NLP), were applied successfully to analyze protein sequences. A major drawback of these models is their size in terms of the number of parameters needed to be fitted and the amount of computational resources they require. Recently, 'distilled' models using the concept of student and teacher networks have been widely used in NLP. Here, we adapted this concept to the problem of protein sequence analysis, by developing DistilProtBert, a distilled version of the successful ProtBert model. Implementing this approach, we reduced the size of the network and the running time by 50%, and the computational resources needed for pretraining by 98% relative to ProtBert model. Using two published tasks, we showed that the performance of the distilled model approaches that of the full model. We next tested the ability of DistilProtBert to distinguish between real and random protein sequences. The task is highly challenging if the composition is maintained on the level of singlet, doublet and triplet amino acids. Indeed, traditional machine-learning algorithms have difficulties with this task. Here, we show that DistilProtBert preforms very well on singlet, doublet and even triplet-shuffled versions of the human proteome, with AUC of 0.92, 0.91 and 0.87, respectively. Finally, we suggest that by examining the small number of false-positive classifications (i.e. shuffled sequences classified as proteins by DistilProtBert), we may be able to identify de novo potential natural-like proteins based on random shuffling of amino acid sequences. AVAILABILITY AND IMPLEMENTATION: https://github.com/yarongef/DistilProtBert.

UNLABELLED: Antibody epitope mapping is a key step in understanding antibody-antigen recognition and is of particular interest for drug development, diagnostics and vaccine design. Most computational methods for epitope prediction are based on properties of the antigen sequence and/or structure, not taking into account the antibody for which the epitope is predicted. Here, we introduce PEASE, a web server predicting antibody-specific epitopes, utilizing the sequence of the antibody. The predictions are provided both at the residue level and as patches on the antigen structure. The tradeoff between recall and precision can be tuned by the user, by changing the default parameters. The results are provided as text and HTML files as well as a graph, and can be viewed on the antigen 3D structure. AVAILABILITY AND IMPLEMENTATION: PEASE is freely available on the web at www.ofranlab.org/PEASE. CONTACT: [email protected].

Non-covalent residue-residue contacts drive the folding of proteins and stabilize them. They may be local-i.e. involve residues that are close in sequence, or non-local. It has been suggested that, in most proteins, local contacts drive protein folding by providing crucial constraints of the conformational space, thus allowing proteins to fold. We compared residues that are involved in local contacts to residues that are involved in non-local contacts and found that, in most proteins, residues in non-local contacts are significantly more conserved evolutionarily than residues in local contacts. Moreover, non-local contacts are more structurally conserved: a contact between positions that are distant in sequence is more likely to exist in many structural homologues compared with a contact between positions that are close in sequence. These results provide new insights into the mechanisms of protein folding and may allow for better prediction of critical intra-chain contacts.

MOTIVATION: Assessing the false positive rate of function prediction methods is difficult, as it is hard to establish that a protein does not have a certain function. To determine to what extent proteins with similar sequences have a common function, we focused on photosynthesis-related proteins. A protein that comes from a non-photosynthetic organism is, undoubtedly, not involved in photosynthesis. RESULTS: We show that function diverges very rapidly: 70% of the close homologs of photosynthetic proteins come from non-photosynthetic organisms. Therefore, high sequence similarity, in most cases, is not tantamount to similar function. However, we found that many functionally similar proteins often share short sequence elements, which may correspond to a functional site and could reveal functional similarities more accurately than sequence similarity. CONCLUSIONS: These results shed light on the way biological function is conserved in evolution and may help improve large-scale analysis of protein function.

MOTIVATION: Biological processes are dynamic, whereas the networks that depict them are typically static. Quantitative modeling using differential equations or logic-based functions can offer quantitative predictions of the behavior of biological systems, but they require detailed experimental characterization of interaction kinetics, which is typically unavailable. To determine to what extent complex biological processes can be modeled and analyzed using only the static structure of the network (i.e. the direction and sign of the edges), we attempt to predict the phenotypic effect of perturbations in biological networks from the static network structure. RESULTS: We analyzed three networks from different sources: The EGFR/MAPK and PI3K/AKT network from a detailed experimental study, the TNF regulatory network from the STRING database and a large network of all NCI-curated pathways from the Protein Interaction Database. Altogether, we predicted the effect of 39 perturbations (e.g. by one or two drugs) on 433 target proteins/genes. In up to 82% of the cases, an algorithm that used only the static structure of the network correctly predicted whether any given protein/gene is upregulated or downregulated as a result of perturbations of other proteins/genes. CONCLUSION: While quantitative modeling requires detailed experimental data and heavy computations, which limit its scalability for large networks, a wiring-based approach can use available data from pathway and interaction databases and may be scalable. These results lay the foundations for a large-scale approach of predicting phenotypes based on the schematic structure of networks.

The Complementarity Determining Regions (CDRs) of antibodies are assumed to account for the antigen recognition and binding and thus to contain also the antigen binding site. CDRs are typically discerned by searching for regions that are most different, in sequence or in structure, between different antibodies. Here, we show that ~20% of the antibody residues that actually bind the antigen fall outside the CDRs. However, virtually all antigen binding residues lie in regions of structural consensus across antibodies. Furthermore, we show that these regions of structural consensus which cover the antigen binding site are identifiable from the sequence of the antibody. Analyzing the predicted contribution of antigen binding residues to the stability of the antibody-antigen complex, we show that residues that fall outside of the traditionally defined CDRs are at least as important to antigen binding as residues within the CDRs, and in some cases, they are even more important energetically. Furthermore, antigen binding residues that fall outside of the structural consensus regions but within traditionally defined CDRs show a marginal energetic contribution to antigen binding. These findings allow for systematic and comprehensive identification of antigen binding sites, which can improve the understanding of antigenic interactions and may be useful in antibody engineering and B-cell epitope identification.

Choosing the right function prediction tools The vast majority of known proteins have not yet been characterized experimentally, and there is very little that is known about their function. New unannotated sequences are added to the databases at a pace that far exceeds the one in which they are annotated in the lab. Computational biology offers tools that can provide insight into the function of proteins based on their sequence, their structure, their evolutionary history, and their association with other proteins. In this contribution, we attempt to provide a framework that will enable biologists and computational biologists to decide which type of computational tool is appropriate for the analysis of their protein of interest, and what kind of insights into its function these tools can provide. In particular, we describe computational methods for predicting protein function directly from sequence or structure, focusing mainly on methods for predicting molecular function. We do not discuss methods that rely on sources of information that are beyond the protein itself, such as genomic context [1], protein–protein interaction networks [2], or membership in biochemical pathways [3]. When choosing a tool for function prediction, one would typically want to identify the best performing tool. However, a quantitative comparison of different tools is a tricky task. While most developers report their own assessment of their tool, in most cases there are no standard datasets and generally agreed-upon measures and criteria for benchmarking function prediction methods. In the absence of independent benchmarks, comparing the figures reported by the developers is almost always comparing oranges and apples (for discussion of this problem see [4]). Therefore, we refrain from reporting numerical assessments of specific methods. For those cases in which independent assessment of performance is available, we refer the reader to the original publications. Finally, we discuss only methods that are either accessible as Web servers or freely available for download (relevant Web links can be found in Table S1).

MOTIVATION: Large-scale experiments reveal pairs of interacting proteins but leave the residues involved in the interactions unknown. These interface residues are essential for understanding the mechanism of interaction and are often desired drug targets. Reliable identification of residues that reside in protein-protein interface typically requires analysis of protein structure. Therefore, for the vast majority of proteins, for which there is no high-resolution structure, there is no effective way of identifying interface residues. RESULTS: Here we present a machine learning-based method that identifies interacting residues from sequence alone. Although the method is developed using transient protein-protein interfaces from complexes of experimentally known 3D structures, it never explicitly uses 3D information. Instead, we combine predicted structural features with evolutionary information. The strongest predictions of the method reached over 90% accuracy in a cross-validation experiment. Our results suggest that despite the significant diversity in the nature of protein-protein interactions, they all share common basic principles and that these principles are identifiable from sequence alone.

Protein-protein interactions, a key to almost any biological process, are mediated by molecular mechanisms that are not entirely clear. The study of these mechanisms often focuses on all residues at protein-protein interfaces. However, only a small subset of all interface residues is actually essential for recognition or binding. Commonly referred to as "hotspots," these essential residues are defined as residues that impede protein-protein interactions if mutated. While no in silico tool identifies hotspots in unbound chains, numerous prediction methods were designed to identify all the residues in a protein that are likely to be a part of protein-protein interfaces. These methods typically identify successfully only a small fraction of all interface residues. Here, we analyzed the hypothesis that the two subsets correspond (i.e., that in silico methods may predict few residues because they preferentially predict hotspots). We demonstrate that this is indeed the case and that we can therefore predict directly from the sequence of a single protein which residues are interaction hotspots (without knowledge of the interaction partner). Our results suggested that most protein complexes are stabilized by similar basic principles. The ability to accurately and efficiently identify hotspots from sequence enables the annotation and analysis of protein-protein interaction hotspots in entire organisms and thus may benefit function prediction and drug development. The server for prediction is available at http://www.rostlab.org/services/isis.