Search

Summary: Self-supervised deep language modeling has shown unprecedented success across natural language tasks, and has recently been repurposed to biological sequences. However, existing models and pretraining methods are designed and optimized for text analysis. We introduce ProteinBERT, a deep language model specifically designed for proteins. Our pretraining scheme combines language modeling with a novel task of Gene Ontology (GO) annotation prediction. We introduce novel architectural elements that make the model highly efficient and flexible to long sequences. The architecture of ProteinBERT consists of both local and global representations, allowing end-to-end processing of these types of inputs and outputs. ProteinBERT obtains near state-of-the-art performance, and sometimes exceeds it, on multiple benchmarks covering diverse protein properties (including protein structure, post-translational modifications and biophysical attributes), despite using a far smaller and faster model than competing deep-learning methods. Overall, ProteinBERT provides an efficient framework for rapidly training protein predictors, even with limited labeled data., Availability and Implementation: Code and pretrained model weights are available at https://github.com/nadavbra/protein_bert., Supplementary Information: Supplementary data are available at Bioinformatics online., (© The Author(s) 2022. Published by Oxford University Press.)

Summary: Current technologies for single-cell transcriptomics allow thousands of cells to be analyzed in a single experiment. The increased scale of these methods raises the risk of cell doublets contamination. Available tools and algorithms for identifying doublets and estimating their occurrence in single-cell experimental data focus on doublets of different species, cell types or individuals. In this study, we analyze transcriptomic data from single cells having an identical genetic background. We claim that the ratio of monoallelic to biallelic expression provides a discriminating power toward doublets' identification. We present a pipeline called BIallelic Ratio for Doublets (BIRD) that relies on heterologous genetic variations, from single-cell RNA sequencing. For each dataset, doublets were artificially created from the actual data and used to train a predictive model. BIRD was applied on Smart-seq data from 163 primary fibroblast single cells. The model achieved 100% accuracy in annotating the randomly simulated doublets. Bonafide doublets were verified based on a biallelic expression signal amongst X-chromosome of female fibroblasts. Data from 10X Genomics microfluidics of human peripheral blood cells achieved in average 83% (±3.7%) accuracy, and an area under the curve of 0.88 (±0.04) for a collection of ∼13 300 single cells. BIRD addresses instances of doublets, which were formed from cell mixtures of identical genetic background and cell identity. Maximal performance is achieved for high-coverage data from Smart-seq. Success in identifying doublets is data specific which varies according to the experimental methodology, genomic diversity between haplotypes, sequence coverage and depth., Supplementary Information: Supplementary data are available at Bioinformatics online., (© The Author(s) 2020. Published by Oxford University Press.)

Motivation: The amount of sequenced genomes and proteins is growing at an unprecedented pace. Unfortunately, manual curation and functional knowledge lag behind. Homologous inference often fails at labeling proteins with diverse functions and broad classes. Thus, identifying high-level protein functionality remains challenging. We hypothesize that a universal feature engineering approach can yield classification of high-level functions and unified properties when combined with machine learning approaches, without requiring external databases or alignment., Results: In this study, we present a novel bioinformatics toolkit called ProFET (Protein Feature Engineering Toolkit). ProFET extracts hundreds of features covering the elementary biophysical and sequence derived attributes. Most features capture statistically informative patterns. In addition, different representations of sequences and the amino acids alphabet provide a compact, compressed set of features. The results from ProFET were incorporated in data analysis pipelines, implemented in python and adapted for multi-genome scale analysis. ProFET was applied on 17 established and novel protein benchmark datasets involving classification for a variety of binary and multi-class tasks. The results show state of the art performance. The extracted features' show excellent biological interpretability. The success of ProFET applies to a wide range of high-level functions such as subcellular localization, structural classes and proteins with unique functional properties (e.g. neuropeptide precursors, thermophilic and nucleic acid binding). ProFET allows easy, universal discovery of new target proteins, as well as understanding the features underlying different high-level protein functions., Availability and Implementation: ProFET source code and the datasets used are freely available at https://github.com/ddofer/ProFET., Contact: michall@cc.huji.ac.il, Supplementary Information: Supplementary data are available at Bioinformatics online., (© The Author 2015. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com.)

Unlabelled: Speed is of the essence in combating Ebola; thus, computational approaches should form a significant component of Ebola research. As for the development of any modern drug, computational biology is uniquely positioned to contribute through comparative analysis of the genome sequences of Ebola strains and three-dimensional protein modeling. Other computational approaches to Ebola may include large-scale docking studies of Ebola proteins with human proteins and with small-molecule libraries, computational modeling of the spread of the virus, computational mining of the Ebola literature and creation of a curated Ebola database. Taken together, such computational efforts could significantly accelerate traditional scientific approaches. In recognition of the need for important and immediate solutions from the field of computational biology against Ebola, the International Society for Computational Biology (ISCB) announces a prize for an important computational advance in fighting the Ebola virus. ISCB will confer the ISCB Fight against Ebola Award, along with a prize of US$2000, at its July 2016 annual meeting (ISCB Intelligent Systems for Molecular Biology 2016, Orlando, FL)., Contact: dkovats@iscb.org or rost@in.tum.de., (© The Author 2015. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com.)

Motivation: Modern protein sequencing techniques have led to the determination of >50 million protein sequences. ProtoNet is a clustering system that provides a continuous hierarchical agglomerative clustering tree for all proteins. While ProtoNet performs unsupervised classification of all included proteins, finding an optimal level of granularity for the purpose of focusing on protein functional groups remain elusive. Here, we ask whether knowledge-based annotations on protein families can support the automatic unsupervised methods for identifying high-quality protein families. We present a method that yields within the ProtoNet hierarchy an optimal partition of clusters, relative to manual annotation schemes. The method's principle is to minimize the entropy-derived distance between annotation-based partitions and all available hierarchical partitions. We describe the best front (BF) partition of 2 478 328 proteins from UniRef50. Of 4,929,553 ProtoNet tree clusters, BF based on Pfam annotations contain 26,891 clusters. The high quality of the partition is validated by the close correspondence with the set of clusters that best describe thousands of keywords of Pfam. The BF is shown to be superior to naïve cut in the ProtoNet tree that yields a similar number of clusters. Finally, we used parameters intrinsic to the clustering process to enrich a priori the BF's clusters. We present the entropy-based method's benefit in overcoming the unavoidable limitations of nested clusters in ProtoNet. We suggest that this automatic information-based cluster selection can be useful for other large-scale annotation schemes, as well as for systematically testing and comparing putative families derived from alternative clustering methods., Availability and Implementation: A catalog of BF clusters for thousands of Pfam keywords is provided at http://protonet.cs.huji.ac.il/bestFront/., (© The Author 2014. Published by Oxford University Press.)

Motivation: The evolution of multicellular organisms is associated with increasing variability of molecules governing behavioral and physiological states. This is often achieved by neuropeptides (NPs) that are produced in neurons from a longer protein, named neuropeptide precursor (NPP). The maturation of NPs occurs through a sequence of proteolytic cleavages. The difficulty in identifying NPPs is a consequence of their diversity and the lack of applicable sequence similarity among the short functionally related NPs., Results: Herein, we describe Neuropeptide Precursor Identifier (NeuroPID), a machine learning scheme that predicts metazoan NPPs. NeuroPID was trained on hundreds of identified NPPs from the UniProtKB database. Some 600 features were extracted from the primary sequences and processed using support vector machines (SVM) and ensemble decision tree classifiers. These features combined biophysical, chemical and informational-statistical properties of NPs and NPPs. Other features were guided by the defining characteristics of the dibasic cleavage sites motif. NeuroPID reached 89-94% accuracy and 90-93% precision in cross-validation blind tests against known NPPs (with an emphasis on Chordata and Arthropoda). NeuroPID also identified NPP-like proteins from extensively studied model organisms as well as from poorly annotated proteomes. We then focused on the most significant sets of features that contribute to the success of the classifiers. We propose that NPPs are attractive targets for investigating and modulating behavior, metabolism and homeostasis and that a rich repertoire of NPs remains to be identified., Availability: NeuroPID source code is freely available at http://www.protonet.cs.huji.ac.il/neuropid

Motivation: MicroRNAs (miRNAs) are short sequences that negatively regulate gene expression. The current understanding of miRNA and their corresponding mRNA targets is primarily based on prediction programs. This study addresses the potential of a coordinated action of miRNAs to manipulate the human pathways. Specifically, we investigate the effectiveness of disrupting the topology of human pathway graphs through a regulation by miRNAs., Results: From a set of miRNA candidates that is associated with a pathway, an exhaustive search for all possible doubles and triplets (coined miR-Duo, miR-Trios) is performed. The impact of each miR-combination on the connectivity of the pathway graph was quantified. About 170 human pathways were tested, and the miR-Duos and miR-Trios were scored for their ability to disrupt these pathway graphs. We show that 75% of all pathways are effectively disconnected by a small number of pathway-specific miR-Trios. Only 15% of the human pathways are resistant to fragmentation by miR-Duos or miR-Trios. Significantly, the combination of the most effective miR-Trios is unique. Thus, a specific regulation of a pathway within the cell is guaranteed. The impact of the selected miR-Duo/Trios on various diseases is discussed., Conclusions: The methodology presented shows that the synthesis of the topology of a network with a detailed understanding of the miRNAs' regulation is useful in exposing critical nodes of the network. We propose the miR-Trio approach as a basis for rationally designed perturbation experiments., Contact: michall@cc.huji.ac.il, Supplementary Information: Supplementary data are available at Bioinformatics online.

Motivation: Much of the large-scale molecular data from living cells can be represented in terms of networks. Such networks occupy a central position in cellular systems biology. In the protein-protein interaction (PPI) network, nodes represent proteins and edges represent connections between them, based on experimental evidence. As PPI networks are rich and complex, a mathematical model is sought to capture their properties and shed light on PPI evolution. The mathematical literature contains various generative models of random graphs. It is a major, still largely open question, which of these models (if any) can properly reproduce various biologically interesting networks. Here, we consider this problem where the graph at hand is the PPI network of Saccharomyces cerevisiae. We are trying to distinguishing between a model family which performs a process of copying neighbors, represented by the duplication-divergence (DD) model, and models which do not copy neighbors, with the Barabási-Albert (BA) preferential attachment model as a leading example., Results: The observed property of the network is the distribution of maximal bicliques in the graph. This is a novel criterion to distinguish between models in this area. It is particularly appropriate for this purpose, since it reflects the graph's growth pattern under either model. This test clearly favors the DD model. In particular, for the BA model, the vast majority (92.9%) of the bicliques with both sides ≥4 must be already embedded in the model's seed graph, whereas the corresponding figure for the DD model is only 5.1%. Our results, based on the biclique perspective, conclusively show that a naïve unmodified DD model can capture a key aspect of PPI networks.

Motivation: Large-scale RNA expression measurements are generating enormous quantities of data. During the last two decades, many methods were developed for extracting insights regarding the interrelationships between genes from such data. The mathematical and computational perspectives that underlie these methods are usually algebraic or probabilistic., Results: Here, we introduce an unexplored geometric view point where expression levels of genes in multiple experiments are interpreted as vectors in a high-dimensional space. Specifically, we find, for the expression profile of each particular gene, its approximation as a linear combination of profiles of a few other genes. This method is inspired by recent developments in the realm of compressed sensing in the machine learning domain. To demonstrate the power of our approach in extracting valuable information from the expression data, we independently applied it to large-scale experiments carried out on the yeast and malaria parasite whole transcriptomes. The parameters extracted from the sparse reconstruction of the expression profiles, when fed to a supervised learning platform, were used to successfully predict the relationships between genes throughout the Gene Ontology hierarchy and protein-protein interaction map. Extensive assessment of the biological results shows high accuracy in both recovering known predictions and in yielding accurate predictions missing from the current databases. We suggest that the geometrical approach presented here is suitable for a broad range of high-dimensional experimental data.

Unlabelled: SPRINT is a software package that performs computational multistate protein design using state-of-the-art inference on probabilistic graphical models. The input to SPRINT is a list of protein structures, the rotamers modeled for each structure and the pre-calculated rotamer energies. Probabilistic inference is performed using the belief propagation or A* algorithms, and dead-end elimination can be applied as pre-processing. The output can either be a list of amino acid sequences simultaneously compatible with these structures, or probabilistic amino acid profiles compatible with the structures. In addition, higher order (e.g. pairwise) amino acid probabilities can also be predicted. Finally, SPRINT also has a module for protein side-chain prediction and single-state design., Availability: The full C++ source code for SPRINT can be freely downloaded from http://www.protonet.cs.huji.ac.il/sprint.

Motivation: In nature, protein-protein interactions are constantly evolving under various selective pressures. Nonetheless, it is expected that crucial interactions are maintained through compensatory mutations between interacting proteins. Thus, many studies have used evolutionary sequence data to extract such occurrences of correlated mutation. However, this research is confounded by other evolutionary pressures that contribute to sequence covariance, such as common ancestry., Results: Here, we focus exclusively on the compensatory mutations deriving from physical protein interactions, by performing large-scale computational mutagenesis experiments for >260 protein-protein interfaces. We investigate the potential for co-adaptability present in protein pairs that are always found together in nature (obligate) and those that are occasionally in complex (transient). By modeling each complex both in bound and unbound forms, we find that naturally transient complexes possess greater relative capacity for correlated mutation than obligate complexes, even when differences in interface size are taken into account.

Motivation: Animal toxins operate by binding to receptors and ion channels. These proteins are short and vary in sequence, structure and function. Sporadic discoveries have also revealed endogenous toxin-like proteins in non-venomous organisms. Viral proteins are the largest group of quickly evolving proteomes. We tested the hypothesis that toxin-like proteins exist in viruses and that they act to modulate functions of their hosts., Results: We updated and improved a classifier for compact proteins resembling short animal toxins that is based on a machine-learning method. We applied it in a large-scale setting to identify toxin-like proteins among short viral proteins. Among the approximately 26 000 representatives of such short proteins, 510 sequences were positively identified. We focused on the 19 highest scoring proteins. Among them, we identified conotoxin-like proteins, growth factors receptor-like proteins and anti-bacterial peptides. Our predictor was shown to enhance annotation inference for many 'uncharacterized' proteins. We conclude that our protocol can expose toxin-like proteins in unexplored niches including metagenomics data and enhance the systematic discovery of novel cell modulators for drug development., Availability: ClanTox is available at http://www.clantox.cs.huji.ac.il.

Summary: The miRror application provides insights on microRNA (miRNA) regulation. It is based on the notion of a combinatorial regulation by an ensemble of miRNAs or genes. miRror integrates predictions from a dozen of miRNA resources that are based on complementary algorithms into a unified statistical framework. For miRNAs set as input, the online tool provides a ranked list of targets, based on set of resources selected by the user, according to their significance of being coordinately regulated. Symmetrically, a set of genes can be used as input to suggest a set of miRNAs. The user can restrict the analysis for the preferred tissue or cell line. miRror is suitable for analyzing results from miRNAs profiling, proteomics and gene expression arrays., Availability: http://www.proto.cs.huji.ac.il/mirror

Motivation: Mapping of remote evolutionary links is a classic computational problem of much interest. Relating protein families allows for functional and structural inference on uncharacterized families. Since sequences have diverged beyond reliable alignment, these are too remote to identify by conventional methods., Approach: We present a method to systematically identify remote evolutionary relations between protein families, leveraging a novel evolutionary-driven tree of all protein sequences and families. A global approach which considers the entire volume of similarities while clustering sequences, leads to a robust tree that allows tracing of very faint evolutionary links. The method systematically scans the tree for clusters which partition exceptionally well into extant protein families, thus suggesting an evolutionary breakpoint in a putative ancient superfamily. Our method does not require family pro.les (or HMMs), or multiple alignment., Results: Considering the entire Pfam database, we are able to suggest 710 links between protein families, 125 of which are con.rmed by existence of Pfam clans. The quality of our predictions is also validated by structural assignments. We further provide an intrinsic characterization of the validity of our results and provide examples for new biological.ndings, from our systematic scan. For example, we are able to relate several bacterial pore-forming toxin families, and then link them with a novel family of eukaryotic toxins expressed in plants.sh venom and notably also uncharacterized proteins from human pathogens., Availability: A detailed list of putative homologous superfamilies, including 210 families of unknown function, has been made available online: http://www.protonet.cs.huji.ac.il/dots

Motivation: UPGMA (average linking) is probably the most popular algorithm for hierarchical data clustering, especially in computational biology. However, UPGMA requires the entire dissimilarity matrix in memory. Due to this prohibitive requirement, UPGMA is not scalable to very large datasets., Application: We present a novel class of memory-constrained UPGMA (MC-UPGMA) algorithms. Given any practical memory size constraint, this framework guarantees the correct clustering solution without explicitly requiring all dissimilarities in memory. The algorithms are general and are applicable to any dataset. We present a data-dependent characterization of hardness and clustering efficiency. The presented concepts are applicable to any agglomerative clustering formulation., Results: We apply our algorithm to the entire collection of protein sequences, to automatically build a comprehensive evolutionary-driven hierarchy of proteins from sequence alone. The newly created tree captures protein families better than state-of-the-art large-scale methods such as CluSTr, ProtoNet4 or single-linkage clustering. We demonstrate that leveraging the entire mass embodied in all sequence similarities allows to significantly improve on current protein family clusterings which are unable to directly tackle the sheer mass of this data. Furthermore, we argue that non-metric constraints are an inherent complexity of the sequence space and should not be overlooked. The robustness of UPGMA allows significant improvement, especially for multidomain proteins, and for large or divergent families., Availability: A comprehensive tree built from all UniProt sequence similarities, together with navigation and classification tools will be made available as part of the ProtoNet service. A C++ implementation of the algorithm is available on request.

Motivation: Feature selection methods aim to reduce the complexity of data and to uncover the most relevant biological variables. In reality, information in biological datasets is often incomplete as a result of untrustworthy samples and missing values. The reliability of selection methods may therefore be questioned., Method: Information loss is incorporated into a perturbation scheme, testing which features are stable under it. This method is applied to data analysis by unsupervised feature filtering (UFF). The latter has been shown to be a very successful method in analysis of gene-expression data., Results: We find that the UFF quality degrades smoothly with information loss. It remains successful even under substantial damage. Our method allows for selection of a best imputation method on a dataset treated by UFF. More importantly, scoring features according to their stability under information loss is shown to be correlated with biological importance in cancer studies. This scoring may lead to novel biological insights.

Motivation: Many methods have been developed for selecting small informative feature subsets in large noisy data. However, unsupervised methods are scarce. Examples are using the variance of data collected for each feature, or the projection of the feature on the first principal component. We propose a novel unsupervised criterion, based on SVD-entropy, selecting a feature according to its contribution to the entropy (CE) calculated on a leave-one-out basis. This can be implemented in four ways: simple ranking according to CE values (SR); forward selection by accumulating features according to which set produces highest entropy (FS1); forward selection by accumulating features through the choice of the best CE out of the remaining ones (FS2); backward elimination (BE) of features with the lowest CE., Results: We apply our methods to different benchmarks. In each case we evaluate the success of clustering the data in the selected feature spaces, by measuring Jaccard scores with respect to known classifications. We demonstrate that feature filtering according to CE outperforms the variance method and gene-shaving. There are cases where the analysis, based on a small set of selected features, outperforms the best score reported when all information was used. Our method calls for an optimal size of the relevant feature set. This turns out to be just a few percents of the number of genes in the two Leukemia datasets that we have analyzed. Moreover, the most favored selected genes turn out to have significant GO enrichment in relevant cellular processes.

Summary: In eukaryotes, membranous proteins account for 20-30% of the proteome. Most of these proteins contain one or more transmembrane (TM) domains. These are short segments that transverse the bilayer lipid membrane. Various properties of the TM domains, such as their number, their topology and their arrangement within the membrane, are closely related to the protein's cellular functions. The properties of the TM domains also determine the cellular targeting and localization of these proteins. It is not known, however, whether the information encoded by TM domains suffices for the purpose of classifying proteins into their functional families. This is the question we address here. We introduce an algorithm that creates a profile of each functional family of membranous proteins based only on the amino acid composition of their TM domains. This is complemented by a classifier program for each such family (to determine whether a given protein belongs to it or not). We find that in most instances TM domains contain enough information to allow an accurate discrimination of approximately 80% sensitivity and approximately 90% specificity among unrelated polytopic functional families with the same number of TM domains., Supplementary Information: Available at www.protonet.cs.huji.ac.il/TM/

Motivation: Structural genomics projects aim to solve a large number of protein structures with the ultimate objective of representing the entire protein space. The computational challenge is to identify and prioritize a small set of proteins with new, currently unknown, superfamilies or folds., Results: We develop a method that assigns each protein a likelihood of it belonging to a new, yet undetermined, structural superfamily. The method relies on a variant of ProtoNet, an automatic hierarchical classification scheme of all protein sequences from SwissProt. Our results show that proteins that are remote from solved structures in the ProtoNet hierarchy are more likely to belong to new superfamilies. The results are validated against SCOP releases from recent years that account for about half of the solved structures known to date. We show that our new method and the representation of ProtoNet are superior in detecting new targets, compared to our previous method using ProtoMap classification. Furthermore, our method outperforms PSI-BLAST search in detecting potential new superfamilies.

Motivation: A major goal in structural genomics is to enrich the catalogue of proteins whose 3D structures are known. In an attempt to address this problem we mapped over 10 000 proteins with solved structures onto a graph of all Swissprot protein sequences (release 36, approximately 73 000 proteins) provided by ProtoMap, with the goal of sorting proteins according to their likelihood of belonging to new superfamilies. We hypothesized that proteins within neighbouring clusters tend to share common structural superfamilies or folds. If true, the likelihood of finding new superfamilies increases in clusters that are distal from other solved structures within the graph., Results: We defined an order relation between unsolved proteins according to their 'distance' from solved structures in the graph, and sorted approximately 48 000 proteins. Our list can be partitioned into three groups: approximately 35 000 proteins sharing a cluster with at least one known structure; approximately 6500 proteins in clusters with no solved structure but with neighbouring clusters containing known structures; and a third group contains the rest of the proteins, approximately 6100 (in 1274 clusters). We tested the quality of the order relation using thousands of recently solved structures that were not included when the order was defined. The tests show that our order is significantly better (P-value approximately 10(5)) than a random order. More interestingly, the order within the union of the second and third groups, and the order within the third group alone, perform better than random (P-values: 0.0008 and 0.15, respectively) and are better than alternative orders created using PSI-BLAST. Herein, we present a method for selecting targets to be used in structural genomics projects., Availability: List of proteins to be used for targets selection combined with a set of biological filters for narrowing down potential targets is in http://www.protarget.cs.huji.ac.il.

Motivation: A large fraction of biological research concentrates on individual proteins and on small families of proteins. One of the current major challenges in bioinformatics is to extend our knowledge to very large sets of proteins. Several major projects have tackled this problem. Such undertakings usually start with a process that clusters all known proteins or large subsets of this space. Some work in this area is carried out automatically, while other attempts incorporate expert advice and annotation., Results: We propose a novel technique that automatically clusters protein sequences. We consider all proteins in SWISSPROT, and carry out an all-against-all BLAST similarity test among them. With this similarity measure in hand we proceed to perform a continuous bottom-up clustering process by applying alternative rules for merging clusters. The outcome of this clustering process is a classification of the input proteins into a hierarchy of clusters of varying degrees of granularity. Here we compare the clusters that result from alternative merging rules, and validate the results against InterPro. Our preliminary results show that clusters that are consistent with several rather than a single merging rule tend to comply with InterPro annotation. This is an affirmation of the view that the protein space consists of families that differ markedly in their evolutionary conservation.