Novel machine learning approaches revolutionize protein knowledge.

Breakthrough methods in machine learning (ML), protein structure prediction, and novel ultrafast structural aligners are revolutionizing structural biology. Obtaining accurate models of proteins and annotating their functions on a large scale is no longer limited by time and resources. The most recent method to be top ranked by the Critical Assessment of Structure Prediction (CASP) assessment, AlphaFold 2 (AF2), is capable of building structural models with an accuracy comparable to that of experimental structures. Annotations of 3D models are keeping pace with the deposition of the structures due to advancements in protein language models (pLMs) and structural aligners that help validate these transferred annotations. In this review we describe how recent developments in ML for protein science are making large-scale structural bioinformatics available to the general scientific community.

CATHe: detection of remote homologues for CATH superfamilies using embeddings from protein language models.

MOTIVATION: CATH is a protein domain classification resource that exploits an automated workflow of structure and sequence comparison alongside expert manual curation to construct a hierarchical classification of evolutionary and structural relationships. The aim of this study was to develop algorithms for detecting remote homologues missed by state-of-the-art hidden Markov model (HMM)-based approaches. The method developed (CATHe) combines a neural network with sequence representations obtained from protein language models. It was assessed using a dataset of remote homologues having less than 20% sequence identity to any domain in the training set. RESULTS: The CATHe models trained on 1773 largest and 50 largest CATH superfamilies had an accuracy of 85.6 ± 0.4% and 98.2 ± 0.3%, respectively. As a further test of the power of CATHe to detect more remote homologues missed by HMMs derived from CATH domains, we used a dataset consisting of protein domains that had annotations in Pfam, but not in CATH. By using highly reliable CATHe predictions (expected error rate <0.5%), we were able to provide CATH annotations for 4.62 million Pfam domains. For a subset of these domains from Homo sapiens, we structurally validated 90.86% of the predictions by comparing their corresponding AlphaFold2 structures with structures from the CATH superfamilies to which they were assigned. AVAILABILITY AND IMPLEMENTATION: The code for the developed models is available on https://github.com/vam-sin/CATHe, and the datasets developed in this study can be accessed on https://zenodo.org/record/6327572. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Prevalent bee venom genes evolved before the aculeate stinger and eusociality.

BACKGROUND: Venoms, which have evolved numerous times in animals, are ideal models of convergent trait evolution. However, detailed genomic studies of toxin-encoding genes exist for only a few animal groups. The hyper-diverse hymenopteran insects are the most speciose venomous clade, but investigation of the origin of their venom genes has been largely neglected. RESULTS: Utilizing a combination of genomic and proteo-transcriptomic data, we investigated the origin of 11 toxin genes in 29 published and 3 new hymenopteran genomes and compiled an up-to-date list of prevalent bee venom proteins. Observed patterns indicate that bee venom genes predominantly originate through single gene co-option with gene duplication contributing to subsequent diversification. CONCLUSIONS: Most Hymenoptera venom genes are shared by all members of the clade and only melittin and the new venom protein family anthophilin1 appear unique to the bee lineage. Most venom proteins thus predate the mega-radiation of hymenopterans and the evolution of the aculeate stinger.

Mutations in transmembrane proteins: diseases, evolutionary insights, prediction and comparison with globular proteins.

Membrane proteins are unique in that they interact with lipid bilayers, making them indispensable for transporting molecules and relaying signals between and across cells. Due to the significance of the protein's functions, mutations often have profound effects on the fitness of the host. This is apparent both from experimental studies, which implicated numerous missense variants in diseases, as well as from evolutionary signals that allow elucidating the physicochemical constraints that intermembrane and aqueous environments bring. In this review, we report on the current state of knowledge acquired on missense variants (referred to as to single amino acid variants) affecting membrane proteins as well as the insights that can be extrapolated from data already available. This includes an overview of the annotations for membrane protein variants that have been collated within databases dedicated to the topic, bioinformatics approaches that leverage evolutionary information in order to shed light on previously uncharacterized membrane protein structures or interaction interfaces, tools for predicting the effects of mutations tailored specifically towards the characteristics of membrane proteins as well as two clinically relevant case studies explaining the implications of mutated membrane proteins in cancer and cardiomyopathy.

PredictProtein - Predicting Protein Structure and Function for 29 Years.

Since 1992 PredictProtein (https://predictprotein.org) is a one-stop online resource for protein sequence analysis with its main site hosted at the Luxembourg Centre for Systems Biomedicine (LCSB) and queried monthly by over 3,000 users in 2020. PredictProtein was the first Internet server for protein predictions. It pioneered combining evolutionary information and machine learning. Given a protein sequence as input, the server outputs multiple sequence alignments, predictions of protein structure in 1D and 2D (secondary structure, solvent accessibility, transmembrane segments, disordered regions, protein flexibility, and disulfide bridges) and predictions of protein function (functional effects of sequence variation or point mutations, Gene Ontology (GO) terms, subcellular localization, and protein-, RNA-, and DNA binding). PredictProtein's infrastructure has moved to the LCSB increasing throughput; the use of MMseqs2 sequence search reduced runtime five-fold (apparently without lowering performance of prediction methods); user interface elements improved usability, and new prediction methods were added. PredictProtein recently included predictions from deep learning embeddings (GO and secondary structure) and a method for the prediction of proteins and residues binding DNA, RNA, or other proteins. PredictProtein.org aspires to provide reliable predictions to computational and experimental biologists alike. All scripts and methods are freely available for offline execution in high-throughput settings.

Embeddings from protein language models predict conservation and variant effects.

The emergence of SARS-CoV-2 variants stressed the demand for tools allowing to interpret the effect of single amino acid variants (SAVs) on protein function. While Deep Mutational Scanning (DMS) sets continue to expand our understanding of the mutational landscape of single proteins, the results continue to challenge analyses. Protein Language Models (pLMs) use the latest deep learning (DL) algorithms to leverage growing databases of protein sequences. These methods learn to predict missing or masked amino acids from the context of entire sequence regions. Here, we used pLM representations (embeddings) to predict sequence conservation and SAV effects without multiple sequence alignments (MSAs). Embeddings alone predicted residue conservation almost as accurately from single sequences as ConSeq using MSAs (two-state Matthews Correlation Coefficient-MCC-for ProtT5 embeddings of 0.596 ± 0.006 vs. 0.608 ± 0.006 for ConSeq). Inputting the conservation prediction along with BLOSUM62 substitution scores and pLM mask reconstruction probabilities into a simplistic logistic regression (LR) ensemble for Variant Effect Score Prediction without Alignments (VESPA) predicted SAV effect magnitude without any optimization on DMS data. Comparing predictions for a standard set of 39 DMS experiments to other methods (incl. ESM-1v, DeepSequence, and GEMME) revealed our approach as competitive with the state-of-the-art (SOTA) methods using MSA input. No method outperformed all others, neither consistently nor statistically significantly, independently of the performance measure applied (Spearman and Pearson correlation). Finally, we investigated binary effect predictions on DMS experiments for four human proteins. Overall, embedding-based methods have become competitive with methods relying on MSAs for SAV effect prediction at a fraction of the costs in computing/energy. Our method predicted SAV effects for the entire human proteome (~ 20 k proteins) within 40 min on one Nvidia Quadro RTX 8000. All methods and data sets are freely available for local and online execution through bioembeddings.com, https://github.com/Rostlab/VESPA , and PredictProtein.

Clustering FunFams using sequence embeddings improves EC purity.

MOTIVATION: Classifying proteins into functional families can improve our understanding of protein function and can allow transferring annotations within one family. For this, functional families need to be 'pure', i.e., contain only proteins with identical function. Functional Families (FunFams) cluster proteins within CATH superfamilies into such groups of proteins sharing function. 11% of all FunFams (22 830 of 203 639) contain EC annotations and of those, 7% (1526 of 22 830) have inconsistent functional annotations. RESULTS: We propose an approach to further cluster FunFams into functionally more consistent sub-families by encoding their sequences through embeddings. These embeddings originate from language models transferring knowledge gained from predicting missing amino acids in a sequence (ProtBERT) and have been further optimized to distinguish between proteins belonging to the same or a different CATH superfamily (PB-Tucker). Using distances between embeddings and DBSCAN to cluster FunFams and identify outliers, doubled the number of pure clusters per FunFam compared to random clustering. Our approach was not limited to FunFams but also succeeded on families created using sequence similarity alone. Complementing EC annotations, we observed similar results for binding annotations. Thus, we expect an increased purity also for other aspects of function. Our results can help generating FunFams; the resulting clusters with improved functional consistency allow more reliable inference of annotations. We expect this approach to succeed equally for any other grouping of proteins by their phenotypes. AVAILABILITY AND IMPLEMENTATION: Code and embeddings are available via GitHub: https://github.com/Rostlab/FunFamsClustering. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Modeling aspects of the language of life through transfer-learning protein sequences.

BACKGROUND: Predicting protein function and structure from sequence is one important challenge for computational biology. For 26 years, most state-of-the-art approaches combined machine learning and evolutionary information. However, for some applications retrieving related proteins is becoming too time-consuming. Additionally, evolutionary information is less powerful for small families, e.g. for proteins from the Dark Proteome. Both these problems are addressed by the new methodology introduced here. RESULTS: We introduced a novel way to represent protein sequences as continuous vectors (embeddings) by using the language model ELMo taken from natural language processing. By modeling protein sequences, ELMo effectively captured the biophysical properties of the language of life from unlabeled big data (UniRef50). We refer to these new embeddings as SeqVec (Sequence-to-Vector) and demonstrate their effectiveness by training simple neural networks for two different tasks. At the per-residue level, secondary structure (Q3 = 79% ± 1, Q8 = 68% ± 1) and regions with intrinsic disorder (MCC = 0.59 ± 0.03) were predicted significantly better than through one-hot encoding or through Word2vec-like approaches. At the per-protein level, subcellular localization was predicted in ten classes (Q10 = 68% ± 1) and membrane-bound were distinguished from water-soluble proteins (Q2 = 87% ± 1). Although SeqVec embeddings generated the best predictions from single sequences, no solution improved over the best existing method using evolutionary information. Nevertheless, our approach improved over some popular methods using evolutionary information and for some proteins even did beat the best. Thus, they prove to condense the underlying principles of protein sequences. Overall, the important novelty is speed: where the lightning-fast HHblits needed on average about two minutes to generate the evolutionary information for a target protein, SeqVec created embeddings on average in 0.03 s. As this speed-up is independent of the size of growing sequence databases, SeqVec provides a highly scalable approach for the analysis of big data in proteomics, i.e. microbiome or metaproteome analysis. CONCLUSION: Transfer-learning succeeded to extract information from unlabeled sequence databases relevant for various protein prediction tasks. SeqVec modeled the language of life, namely the principles underlying protein sequences better than any features suggested by textbooks and prediction methods. The exception is evolutionary information, however, that information is not available on the level of a single sequence.

Dark Proteins Important for Cellular Function.

Despite substantial and successful projects for structural genomics, many proteins remain for which neither experimental structures nor homology-based models are known for any part of the amino acid sequence. These have been called "dark proteins," in contrast to non-dark proteins, in which at least part of the sequence has a known or inferred structure. It has been hypothesized that non-dark proteins may be more abundantly expressed than dark proteins, which are known to have much fewer sequence relatives. Surprisingly, the opposite has been observed: human dark and non-dark proteins had quite similar levels of expression, in terms of both mRNA and protein abundance. Such high levels of expression strongly indicate that dark proteins-as a group-are important for cellular function. This is remarkable, given how carefully structural biologists have focused on proteins crucial for function, and highlights the important challenge posed by dark proteins in future research.

Artificial Intelligence Learns Protein Prediction.

From AlphaGO over StableDiffusion to ChatGPT, the recent decade of exponential advances in artificial intelligence (AI) has been altering life. In parallel, advances in computational biology are beginning to decode the language of life: AlphaFold2 leaped forward in protein structure prediction, and protein language models (pLMs) replaced expertise and evolutionary information from multiple sequence alignments with information learned from reoccurring patterns in databases of billions of proteins without experimental annotations other than the amino acid sequences. None of those tools could have been developed 10 years ago; all will increase the wealth of experimental data and speed up the cycle from idea to proof. AI is affecting molecular and medical biology at giant steps, and the most important might be the leap toward more powerful protein design.

Protein embeddings predict binding residues in disordered regions.

The identification of protein binding residues helps to understand their biological processes as protein function is often defined through ligand binding, such as to other proteins, small molecules, ions, or nucleotides. Methods predicting binding residues often err for intrinsically disordered proteins or regions (IDPs/IDPRs), often also referred to as molecular recognition features (MoRFs). Here, we presented a novel machine learning (ML) model trained to specifically predict binding regions in IDPRs. The proposed model, IDBindT5, leveraged embeddings from the protein language model (pLM) ProtT5 to reach a balanced accuracy of 57.2 ± 3.6% (95% confidence interval). Assessed on the same data set, this did not differ at the 95% CI from the state-of-the-art (SOTA) methods ANCHOR2 and DeepDISOBind that rely on expert-crafted features and evolutionary information from multiple sequence alignments (MSAs). Assessed on other data, methods such as SPOT-MoRF reached higher MCCs. IDBindT5's SOTA predictions are much faster than other methods, easily enabling full-proteome analyses. Our findings emphasize the potential of pLMs as a promising approach for exploring and predicting features of disordered proteins. The model and a comprehensive manual are publicly available at https://github.com/jahnl/binding_in_disorder .

From sequence to function through structure: Deep learning for protein design.

The process of designing biomolecules, in particular proteins, is witnessing a rapid change in available tooling and approaches, moving from design through physicochemical force fields, to producing plausible, complex sequences fast via end-to-end differentiable statistical models. To achieve conditional and controllable protein design, researchers at the interface of artificial intelligence and biology leverage advances in natural language processing (NLP) and computer vision techniques, coupled with advances in computing hardware to learn patterns from growing biological databases, curated annotations thereof, or both. Once learned, these patterns can be used to provide novel insights into mechanistic biology and the design of biomolecules. However, navigating and understanding the practical applications for the many recent protein design tools is complex. To facilitate this, we 1) document recent advances in deep learning (DL) assisted protein design from the last three years, 2) present a practical pipeline that allows to go from de novo-generated sequences to their predicted properties and web-powered visualization within minutes, and 3) leverage it to suggest a generated protein sequence which might be used to engineer a biosynthetic gene cluster to produce a molecular glue-like compound. Lastly, we discuss challenges and highlight opportunities for the protein design field.

Domain loss enabled evolution of novel functions in the snake three-finger toxin gene superfamily.

Three-finger toxins (3FTXs) are a functionally diverse family of toxins, apparently unique to venoms of caenophidian snakes. Although the ancestral function of 3FTXs is antagonism of nicotinic acetylcholine receptors, redundancy conferred by the accumulation of duplicate genes has facilitated extensive neofunctionalization, such that derived members of the family interact with a range of targets. 3FTXs are members of the LY6/UPAR family, but their non-toxin ancestor remains unknown. Combining traditional phylogenetic approaches, manual synteny analysis, and machine learning techniques (including AlphaFold2 and ProtT5), we have reconstructed a detailed evolutionary history of 3FTXs. We identify their immediate ancestor as a non-secretory LY6, unique to squamate reptiles, and propose that changes in molecular ecology resulting from loss of a membrane-anchoring domain and changes in gene expression, paved the way for the evolution of one of the most important families of snake toxins.

AlphaFold2 reveals commonalities and novelties in protein structure space for 21 model organisms.

Deep-learning (DL) methods like DeepMind's AlphaFold2 (AF2) have led to substantial improvements in protein structure prediction. We analyse confident AF2 models from 21 model organisms using a new classification protocol (CATH-Assign) which exploits novel DL methods for structural comparison and classification. Of ~370,000 confident models, 92% can be assigned to 3253 superfamilies in our CATH domain superfamily classification. The remaining cluster into 2367 putative novel superfamilies. Detailed manual analysis on 618 of these, having at least one human relative, reveal extremely remote homologies and further unusual features. Only 25 novel superfamilies could be confirmed. Although most models map to existing superfamilies, AF2 domains expand CATH by 67% and increases the number of unique 'global' folds by 36% and will provide valuable insights on structure function relationships. CATH-Assign will harness the huge expansion in structural data provided by DeepMind to rationalise evolutionary changes driving functional divergence.

LambdaPP: Fast and accessible protein-specific phenotype predictions.

The availability of accurate and fast artificial intelligence (AI) solutions predicting aspects of proteins are revolutionizing experimental and computational molecular biology. The webserver LambdaPP aspires to supersede PredictProtein, the first internet server making AI protein predictions available in 1992. Given a protein sequence as input, LambdaPP provides easily accessible visualizations of protein 3D structure, along with predictions at the protein level (GeneOntology, subcellular location), and the residue level (binding to metal ions, small molecules, and nucleotides; conservation; intrinsic disorder; secondary structure; alpha-helical and beta-barrel transmembrane segments; signal-peptides; variant effect) in seconds. The structure prediction provided by LambdaPP-leveraging ColabFold and computed in minutes-is based on MMseqs2 multiple sequence alignments. All other feature prediction methods are based on the pLM ProtT5. Queried by a protein sequence, LambdaPP computes protein and residue predictions almost instantly for various phenotypes, including 3D structure and aspects of protein function. LambdaPP is freely available for everyone to use under embed.predictprotein.org, the interactive results for the case study can be found under https://embed.predictprotein.org/o/Q9NZC2. The frontend of LambdaPP can be found on GitHub (github.com/sacdallago/embed.predictprotein.org), and can be freely used and distributed under the academic free use license (AFL-2). For high-throughput applications, all methods can be executed locally via the bio-embeddings (bioembeddings.com) python package, or docker image at ghcr.io/bioembeddings/bio_embeddings, which also includes the backend of LambdaPP.

Protein language-model embeddings for fast, accurate, and alignment-free protein structure prediction.

Advanced protein structure prediction requires evolutionary information from multiple sequence alignments (MSAs) from evolutionary couplings that are not always available. Artificial intelligence (AI)-based predictions inputting only single sequences are faster but so inaccurate as to render speed irrelevant. Here, we described a competitive prediction of inter-residue distances (2D structure) exclusively inputting embeddings from pre-trained protein language models (pLMs), namely ProtT5, from single sequences into a convolutional neural network (CNN) with relatively few layers. The major advance used the ProtT5 attention heads. Our new method, EMBER2, which never requires any MSAs, performed similarly to other methods that fully rely on co-evolution. Although clearly not reaching AlphaFold2, our leaner solution came somehow close at substantially lower costs. By generating protein-specific rather than family-averaged predictions, EMBER2 might better capture some features of particular protein structures. Results from using protein engineering and deep mutational scanning (DMS) experiments provided at least a proof of principle for such a speculation.

Nearest neighbor search on embeddings rapidly identifies distant protein relations.

Since 1992, all state-of-the-art methods for fast and sensitive identification of evolutionary, structural, and functional relations between proteins (also referred to as "homology detection") use sequences and sequence-profiles (PSSMs). Protein Language Models (pLMs) generalize sequences, possibly capturing the same constraints as PSSMs, e.g., through embeddings. Here, we explored how to use such embeddings for nearest neighbor searches to identify relations between protein pairs with diverged sequences (remote homology detection for levels of <20% pairwise sequence identity, PIDE). While this approach excelled for proteins with single domains, we demonstrated the current challenges applying this to multi-domain proteins and presented some ideas how to overcome existing limitations, in principle. We observed that sufficiently challenging data set separations were crucial to provide deeply relevant insights into the behavior of nearest neighbor search when applied to the protein embedding space, and made all our methods readily available for others.

SETH predicts nuances of residue disorder from protein embeddings.

Predictions for millions of protein three-dimensional structures are only a few clicks away since the release of AlphaFold2 results for UniProt. However, many proteins have so-called intrinsically disordered regions (IDRs) that do not adopt unique structures in isolation. These IDRs are associated with several diseases, including Alzheimer's Disease. We showed that three recent disorder measures of AlphaFold2 predictions (pLDDT, "experimentally resolved" prediction and "relative solvent accessibility") correlated to some extent with IDRs. However, expert methods predict IDRs more reliably by combining complex machine learning models with expert-crafted input features and evolutionary information from multiple sequence alignments (MSAs). MSAs are not always available, especially for IDRs, and are computationally expensive to generate, limiting the scalability of the associated tools. Here, we present the novel method SETH that predicts residue disorder from embeddings generated by the protein Language Model ProtT5, which explicitly only uses single sequences as input. Thereby, our method, relying on a relatively shallow convolutional neural network, outperformed much more complex solutions while being much faster, allowing to create predictions for the human proteome in about 1 hour on a consumer-grade PC with one NVIDIA GeForce RTX 3060. Trained on a continuous disorder scale (CheZOD scores), our method captured subtle variations in disorder, thereby providing important information beyond the binary classification of most methods. High performance paired with speed revealed that SETH's nuanced disorder predictions for entire proteomes capture aspects of the evolution of organisms. Additionally, SETH could also be used to filter out regions or proteins with probable low-quality AlphaFold2 3D structures to prioritize running the compute-intensive predictions for large data sets. SETH is freely publicly available at: https://github.com/Rostlab/SETH.

Contrastive learning on protein embeddings enlightens midnight zone.

Experimental structures are leveraged through multiple sequence alignments, or more generally through homology-based inference (HBI), facilitating the transfer of information from a protein with known annotation to a query without any annotation. A recent alternative expands the concept of HBI from sequence-distance lookup to embedding-based annotation transfer (EAT). These embeddings are derived from protein Language Models (pLMs). Here, we introduce using single protein representations from pLMs for contrastive learning. This learning procedure creates a new set of embeddings that optimizes constraints captured by hierarchical classifications of protein 3D structures defined by the CATH resource. The approach, dubbed ProtTucker, has an improved ability to recognize distant homologous relationships than more traditional techniques such as threading or fold recognition. Thus, these embeddings have allowed sequence comparison to step into the 'midnight zone' of protein similarity, i.e. the region in which distantly related sequences have a seemingly random pairwise sequence similarity. The novelty of this work is in the particular combination of tools and sampling techniques that ascertained good performance comparable or better to existing state-of-the-art sequence comparison methods. Additionally, since this method does not need to generate alignments it is also orders of magnitudes faster. The code is available at https://github.com/Rostlab/EAT.

Improving protein succinylation sites prediction using embeddings from protein language model.

Protein succinylation is an important post-translational modification (PTM) responsible for many vital metabolic activities in cells, including cellular respiration, regulation, and repair. Here, we present a novel approach that combines features from supervised word embedding with embedding from a protein language model called ProtT5-XL-UniRef50 (hereafter termed, ProtT5) in a deep learning framework to predict protein succinylation sites. To our knowledge, this is one of the first attempts to employ embedding from a pre-trained protein language model to predict protein succinylation sites. The proposed model, dubbed LMSuccSite, achieves state-of-the-art results compared to existing methods, with performance scores of 0.36, 0.79, 0.79 for MCC, sensitivity, and specificity, respectively. LMSuccSite is likely to serve as a valuable resource for exploration of succinylation and its role in cellular physiology and disease.