Enhancing missense variant pathogenicity prediction with protein language models using VariPred.

Computational approaches for predicting the pathogenicity of genetic variants have advanced in recent years. These methods enable researchers to determine the possible clinical impact of rare and novel variants. Historically these prediction methods used hand-crafted features based on structural, evolutionary, or physiochemical properties of the variant. In this study we propose a novel framework that leverages the power of pre-trained protein language models to predict variant pathogenicity. We show that our approach VariPred (Variant impact Predictor) outperforms current state-of-the-art methods by using an end-to-end model that only requires the protein sequence as input. Using one of the best-performing protein language models (ESM-1b), we establish a robust classifier that requires no calculation of structural features or multiple sequence alignments. We compare the performance of VariPred with other representative models including 3Cnet, Polyphen-2, REVEL, MetaLR, FATHMM and ESM variant. VariPred performs as well as, or in most cases better than these other predictors using six variant impact prediction benchmarks despite requiring only sequence data and no pre-processing of the data.

Do antibody CDR loops change conformation upon binding?

Antibodies have increasingly been developed as drugs with over 100 now licensed in the US or EU. During development, it is often necessary to increase or reduce the affinity of an antibody and rational attempts to do so rely on having a structure of the antibody-antigen complex often obtained by modeling. The antigen-binding site consists primarily of six loops known as complementarity-determining regions (CDRs), and an open question has been whether these loops change their conformation when they bind to an antigen. Existing surveys of antibody-antigen complex structures have only examined CDR conformational change in case studies or small-scale surveys. With an increasing number of antibodies where both free and complexed structures have been deposited in the Protein Data Bank, a large-scale survey of CDR conformational change during binding is now possible. To this end, we built a dataset, AbAgDb, that currently includes 177 antibodies with high-quality CDRs, each of which has at least one bound and one unbound structure. We analyzed the conformational change of the Cα backbone of each CDR upon binding and found that, in most cases, the CDRs (other than CDR-H3) show minimal movement, while 70.6% and 87% of CDR-H3s showed global Cα RMSD ≤ 1.0Å and ≤ 2.0Å, respectively. We also compared bound CDR conformations with the conformational space of unbound CDRs and found most of the bound conformations are included in the unbound conformational space. In future, our results will contribute to developing insights into antibodies and new methods for modeling and docking.

abYpap: improvements to the prediction of antibody VH/VL packing using gradient boosted regression.

The Fv region of the antibody (comprising VH and VL domains) is the area responsible for target binding and thus the antibody's specificity. The orientation, or packing, of these two domains relative to each other influences the topography of the Fv region, and therefore can influence the antibody's binding affinity. We present abYpap, an improved method for predicting the packing angle between the VH and VL domains. With the large data set now available, we were able to expand greatly the number of features that could be used compared with our previous work. The machine-learning model was tuned for improved performance using 37 selected residues (previously 13) and also by including the lengths of the most variable 'complementarity determining regions' (CDR-L1, CDR-L2 and CDR-H3). Our method shows large improvements from the previous version, and also against other modeling approaches, when predicting the packing angle.

Antibody markup language (AbML) - a notation language for antibody-based drug formats and software for creating and rendering AbML (abYdraw).

As interest in antibody-based drug development continues to increase, the biopharmaceutical industry has begun to focus on complex multi-specific antibodies (MsAbs) as an up-and-coming class of biologic that differ from natural monoclonal antibodies through their ability to bind to more than one type of antigen. As techniques to generate such molecules have diversified, so have their formats and the need for standard notation. Previous efforts to develop a notation language for macromolecule drugs have been insufficient, or too complex, for MsAbs. Here, we present Antibody Markup Language (AbML), a new notation language specifically for antibody formats that overcomes the limitations of existing languages and can annotate all current antibody formats, including fusions, fragments, standard antibodies and MsAbs, as well as all currently conceivable future formats. AbML V1.1 also provides explicit support for T-cell receptor domains. To assist users of this language we have also developed a tool, abYdraw, that can draw antibody schematics from AbML strings or generate an AbML string from a drawn antibody schematic. AbML has the potential to become a standardized notation for describing new MsAb formats entering clinical trials.Abbreviations: AbML: Antibody Markup Language; ADC: Antibody-drug conjugate; CAS: Chemical Abstracts Service; CH: Constant heavy; CL: Constant light; Fv: Variable fragment; HELM: Hierarchical Editing Language for Macromolecules; HSA: Human serum albumin; INN: International Nonproprietary Names; KIH: Knobs-into-holes; mAbs: Monoclonal antibodies; MsAb: Multi-specific antibody; WHO: World Health Organization; PEG: Poly-ethylene glycol; scFv: Single-chain variable fragment; SMILES: Simplified Molecular-Input Line-Entry System; VH: Variable heavy; VHH: Single-domain (Camelid) variable heavy; VL: Variable light.

International nonproprietary names for monoclonal antibodies: an evolving nomenclature system.

Appropriate nomenclature for all pharmaceutical substances is important for clinical development, licensing, prescribing, pharmacovigilance, and identification of counterfeits. Nonproprietary names that are unique and globally recognized for all pharmaceutical substances are assigned by the International Nonproprietary Names (INN) Programme of the World Health Organization (WHO). In 1991, the INN Programme implemented the first nomenclature scheme for monoclonal antibodies. To accompany biotechnological development, this nomenclature scheme has evolved over the years; however, since the scheme was introduced, all pharmacological substances that contained an immunoglobulin variable domain were coined with the stem -mab. To date, there are 879 INN with the stem -mab. Owing to this high number of names ending in -mab, devising new and distinguishable INN has become a challenge. The WHO INN Expert Group therefore decided to revise the system to ease this situation. The revised system was approved and adopted by the WHO at the 73rd INN Consultation held in October 2021, and the radical decision was made to discontinue the use of the well-known stem -mab in naming new antibody-based drugs and going forward, to replace it with four new stems: -tug, -bart, -mig, and -ment.

Current advances in biopharmaceutical informatics: guidelines, impact and challenges in the computational developability assessment of antibody therapeutics.

Therapeutic monoclonal antibodies and their derivatives are key components of clinical pipelines in the global biopharmaceutical industry. The availability of large datasets of antibody sequences, structures, and biophysical properties is increasingly enabling the development of predictive models and computational tools for the "developability assessment" of antibody drug candidates. Here, we provide an overview of the antibody informatics tools applicable to the prediction of developability issues such as stability, aggregation, immunogenicity, and chemical degradation. We further evaluate the opportunities and challenges of using biopharmaceutical informatics for drug discovery and optimization. Finally, we discuss the potential of developability guidelines based on in silico metrics that can be used for the assessment of antibody stability and manufacturability.

Zincbindpredict-Prediction of Zinc Binding Sites in Proteins.

Background: Zinc binding proteins make up a significant proportion of the proteomes of most organisms and, within those proteins, zinc performs rôles in catalysis and structure stabilisation. Identifying the ability to bind zinc in a novel protein can offer insights into its functions and the mechanism by which it carries out those functions. Computational means of doing so are faster than spectroscopic means, allowing for searching at much greater speeds and scales, and thereby guiding complimentary experimental approaches. Typically, computational models of zinc binding predict zinc binding for individual residues rather than as a single binding site, and typically do not distinguish between different classes of binding site-missing crucial properties indicative of zinc binding. Methods: Previously, we created ZincBindDB, a continuously updated database of known zinc binding sites, categorised by family (the set of liganding residues). Here, we use this dataset to create ZincBindPredict, a set of machine learning methods to predict the most common zinc binding site families for both structure and sequence. Results: The models all achieve an MCC ≥ 0.88, recall ≥ 0.93 and precision ≥ 0.91 for the structural models (mean MCC = 0.97), while the sequence models have MCC ≥ 0.64, recall ≥ 0.80 and precision ≥ 0.83 (mean MCC = 0.87), with the models for binding sites containing four liganding residues performing much better than this. Conclusions: The predictors outperform competing zinc binding site predictors and are available online via a web interface and a GraphQL API.

GraphQL for the delivery of bioinformatics web APIs and application to ZincBind.

Motivation: Many bioinformatics resources are provided as 'web services', with large databases and analysis software stored on a central server, and clients interacting with them using the hypertext transport protocol (HTTP). While some provide only a visual HTML interface, requiring a web browser to use them, many provide programmatic access using a web application programming interface (API) which returns XML, JSON or plain text that computer programs can interpret more easily. This allows access to be automated. Initially, many bioinformatics APIs used the 'simple object access protocol' (SOAP) and, more recently, representational state transfer (REST). Results: GraphQL is a novel, increasingly prevalent alternative to REST and SOAP that represents the available data in the form of a graph to which any conceivable query can be submitted, and which is seeing increasing adoption in industry. Here, we review the principles of GraphQL, outline its particular suitability to the delivery of bioinformatics resources and describe its implementation in our ZincBind resource. Availability and implementation: https://api.zincbind.net. Supplementary information: Supplementary data are available at Bioinformatics Advances online.

atomium-a Python structure parser.

SUMMARY: Structural biology relies on specific file formats to convey information about macromolecular structures. Traditionally this has been the PDB format, but increasingly newer formats, such as PDBML, mmCIF and MMTF are being used. Here we present atomium, a modern, lightweight, Python library for parsing, manipulating and saving PDB, mmCIF and MMTF file formats. In addition, we provide a web service, pdb2json, which uses atomium to give a consistent JSON representation to the entire Protein Data Bank. AVAILABILITY AND IMPLEMENTATION: atomium is implemented in Python and its performance is equivalent to the existing library BioPython. However, it has significant advantages in features and API design. atomium is available from atomium.bioinf.org.uk and pdb2json can be accessed at pdb2json.bioinf.org.uk. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

B-cell epitopes: Discontinuity and conformational analysis.

Peptide vaccines have many potential advantages over conventional ones including low cost, lack of need for cold-chain storage, safety and specificity. However, it is well known that approximately 90% of B-cell epitopes (BCEs) are discontinuous in nature making it difficult to mimic them for creating vaccines. In this study, the degree of discontinuity in B-cell epitopes and their conformational nature is examined. The discontinuity of B-cell epitopes is analyzed by defining 'regions' (consisting of at least three antibody-contacting residues each separated by ≤3 residues) and small fragments (antibody-contacting residues that do not satisfy the requirements for a region). Secondly, an algorithm has been developed that classifies each region's shape as straight, curved or folded on the basis that straight and folded regions are more likely to retain their native conformation as isolated peptides. We have investigated the structures of 488 B-cell epitopes from which 1282 regions and 1018 fragments have been identified. 90% of epitopes have five or fewer regions and five or fewer fragments with 14% containing only one region and 4% being truly linear (i.e. having one region and no fragments). Of the 1282 regions, 508 are straight in shape, 626 are curved and 148 are folded.

ZincBind-the database of zinc binding sites.

Zinc is one of the most important biologically active metals. Ten per cent of the human genome is thought to encode a zinc binding protein and its uses encompass catalysis, structural stability, gene expression and immunity. At present, there is no specific resource devoted to identifying and presenting all currently known zinc binding sites. Here we present ZincBind, a database of zinc binding sites and its web front-end. Using the structural data in the Protein Data Bank, ZincBind identifies every instance of zinc binding to a protein, identifies its binding site and clusters sites based on 90% sequence identity. There are currently 24 992 binding sites, clustered into 7489 unique sites. The data are available over the web where they can be browsed and downloaded, and via a REST API. ZincBind is regularly updated and will continue to be updated with new data and features.

AbDb: antibody structure database-a database of PDB-derived antibody structures.

Database URL: www.bioinf.org.uk/abs/abdb/.

IntPred: a structure-based predictor of protein-protein interaction sites.

MOTIVATION: Protein-protein interactions are vital for protein function with the average protein having between three and ten interacting partners. Knowledge of precise protein-protein interfaces comes from crystal structures deposited in the Protein Data Bank (PDB), but only 50% of structures in the PDB are complexes. There is therefore a need to predict protein-protein interfaces in silico and various methods for this purpose. Here we explore the use of a predictor based on structural features and which exploits random forest machine learning, comparing its performance with a number of popular established methods. RESULTS: On an independent test set of obligate and transient complexes, our IntPred predictor performs well (MCC = 0.370, ACC = 0.811, SPEC = 0.916, SENS = 0.411) and compares favourably with other methods. Overall, IntPred ranks second of six methods tested with SPPIDER having slightly better overall performance (MCC = 0.410, ACC = 0.759, SPEC = 0.783, SENS = 0.676), but considerably worse specificity than IntPred. As with SPPIDER, using an independent test set of obligate complexes enhanced performance (MCC = 0.381) while performance is somewhat reduced on a dataset of transient complexes (MCC = 0.303). The trade-off between sensitivity and specificity compared with SPPIDER suggests that the choice of the appropriate tool is application-dependent. AVAILABILITY AND IMPLEMENTATION: IntPred is implemented in Perl and may be downloaded for local use or run via a web server at www.bioinf.org.uk/intpred/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

abYsis: Integrated Antibody Sequence and Structure-Management, Analysis, and Prediction.

abYsis is a web-based antibody research system that includes an integrated database of antibody sequence and structure data. The system can be interrogated in numerous ways-from simple text and sequence searches to sophisticated queries that apply 3D structural constraints. The publicly available version includes pre-analyzed sequence data from the European Molecular Biology Laboratory European Nucleotide Archive (EMBL-ENA) and Kabat as well as structure data from the Protein Data Bank. A researcher's own sequences can also be analyzed through the web interface. A defining characteristic of abYsis is that the sequences are automatically numbered with a series of popular schemes such as Kabat and Chothia and then annotated with key information such as complementarity-determining regions and potential post-translational modifications. A unique aspect of abYsis is a set of residue frequency tables for each position in an antibody, allowing "unusual residues" (those rarely seen at a particular position) to be highlighted and decisions to be made on which mutations may be acceptable. This is especially useful when comparing antibodies from different species. abYsis is useful for any researcher specializing in antibody engineering, especially those developing antibodies as drugs. abYsis is available at www.abysis.org.

The structural effects of mutations can aid in differential phenotype prediction of beta-myosin heavy chain (Myosin-7) missense variants.

MOTIVATION: High-throughput sequencing platforms are increasingly used to screen patients with genetic disease for pathogenic mutations, but prediction of the effects of mutations remains challenging. Previously we developed SAAPdap (Single Amino Acid Polymorphism Data Analysis Pipeline) and SAAPpred (Single Amino Acid Polymorphism Predictor) that use a combination of rule-based structural measures to predict whether a missense genetic variant is pathogenic. Here we investigate whether the same methodology can be used to develop a differential phenotype predictor, which, once a mutation has been predicted as pathogenic, is able to distinguish between phenotypes-in this case the two major clinical phenotypes (hypertrophic cardiomyopathy, HCM and dilated cardiomyopathy, DCM) associated with mutations in the beta-myosin heavy chain (MYH7) gene product (Myosin-7). RESULTS: A random forest predictor trained on rule-based structural analyses together with structural clustering data gave a Matthews' correlation coefficient (MCC) of 0.53 (accuracy, 75%). A post hoc removal of machine learning models that performed particularly badly, increased the performance (MCC = 0.61, Acc = 79%). This proof of concept suggests that methods used for pathogenicity prediction can be extended for use in differential phenotype prediction. AVAILABILITY AND IMPLEMENTATION: Analyses were implemented in Perl and C and used the Java-based Weka machine learning environment. Please contact the authors for availability. CONTACTS: andrew@bioinf.org.uk or andrew.martin@ucl.ac.uk SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Extracting human antibody sequences from public databases for antibody humanization: high frequency of species assignment errors.

In antibody humanization, complementarity determining regions from a 'donor' antibody are often grafted onto a human framework selected by high sequence identity with the donor. In our own humanization experiments, we have found that species information is often incorrect. Here we take three mouse antibodies and perform BLAST searches against sequences annotated as being human. We find that the first genuine human hits for the six chains appear at Positions 30, 4, 11, 24, 18 and 29 in the hit lists. This illustrates both the need for caution in performing humanization and for improvements in annotation.

BiopLib and BiopTools--a C programming library and toolset for manipulating protein structure.

UNLABELLED: We describe BiopLib, a mature C programming library for manipulating protein structure, and BiopTools, a set of command-line tools which exploit BiopLib. The library also provides a small number of functions for handling protein sequence and general purpose programming and mathematics. BiopLib transparently handles PDBML (XML) format and standard PDB files. BiopTools provides facilities ranging from renumbering atoms and residues to calculation of solvent accessibility. AVAILABILITY AND IMPLEMENTATION: BiopLib and BiopTools are implemented in standard ANSI C. The core of the BiopLib library is a reliable PDB parser that handles alternate occupancies and deals with compressed PDB files and PDBML files automatically. The library is designed to be as flexible as possible, allowing users to handle PDB data as a simple list of atoms, or in a structured form using chains, residues and atoms. Many of the BiopTools command-line tools act as filters, taking a PDB (or PDBML) file as input and producing a PDB (or PDBML) file as output. All code is open source and documented using Doxygen. It is provided under the GNU Public Licence and is available from the authors' web site or from GitHub.

Protein moonlighting: a new factor in biology and medicine.

The phenomenon of protein moonlighting was discovered in the 1980s and 1990s, and the current definition of what constitutes a moonlighting protein was provided at the end of the 1990s. Since this time, several hundred moonlighting proteins have been identified in all three domains of life, and the rate of discovery is accelerating as the importance of protein moonlighting in biology and medicine becomes apparent. The recent re-evaluation of the number of protein-coding genes in the human genome (approximately 19000) is one reason for believing that protein moonlighting may be a more general phenomenon than the current number of moonlighting proteins would suggest, and preliminary studies of the proportion of proteins that moonlight would concur with this hypothesis. Protein moonlighting could be one way of explaining the seemingly small number of proteins that are encoded in the human genome. It is emerging that moonlighting proteins can exhibit novel biological functions, thus extending the range of the human functional proteome. The several hundred moonlighting proteins so far discovered play important roles in many aspects of biology. For example, glyceraldehyde-3-phosphate dehydrogenase (GAPDH), heat-shock protein 60 (Hsp60) and tRNA synthetases play a wide range of biological roles in eukaryotic cells, and a growing number of eukaryotic moonlighting proteins are recognized to play important roles in physiological processes such as sperm capacitation, implantation, immune regulation in pregnancy, blood coagulation, vascular regeneration and control of inflammation. The dark side of protein moonlighting finds a range of moonlighting proteins playing roles in various human diseases including cancer, cardiovascular disease, HIV and cystic fibrosis. However, some moonlighting proteins are being tested for their therapeutic potential, including immunoglobulin heavy-chain-binding protein (BiP), for rheumatoid arthritis, and Hsp90 for wound healing. In addition, it has emerged over the last 20 years that a large number of bacterial moonlighting proteins play important roles in bacteria-host interactions as virulence factors and are therefore potential therapeutic targets in bacterial infections. So as we progress in the 21st Century, it is likely that moonlighting proteins will be seen to play an increasingly important role in biology and medicine. It is hoped that some of the major unanswered questions, such as the mechanism of evolution of protein moonlighting, the structural biology of moonlighting proteins and their role in the systems biology of cellular systems can be addressed during this period.

Structural biology of moonlighting: lessons from antibodies.

Protein moonlighting is the property of a number of proteins to have more than one function. However, the definition of moonlighting is somewhat imprecise with different interpretations of the phenomenon. True moonlighting occurs when an individual evolutionary protein domain has one well-accepted role and a secondary unrelated function. The 'function' of a protein domain can be defined at different levels. For example, although the function of an antibody variable fragment (Fv) could be described as 'binding', a more detailed definition would also specify the molecule to which the Fv region binds. Using this detailed definition, antibodies as a family are consummate moonlighters. However, individual antibodies do not moonlight; the multiple functions they exhibit (first binding a molecule and second triggering the immune response) are encoded in different domains and, in any case, are related in the sense that they are a part of what an antibody needs to do. Nonetheless, antibodies provide interesting lessons on the ability of proteins to evolve binding functions. Remarkably similar antibody sequences can bind completely different antigens, suggesting that evolving the ability to bind a protein can result from very subtle sequence changes.

Viewing multiple sequence alignments with the JavaScript Sequence Alignment Viewer (JSAV).

The JavaScript Sequence Alignment Viewer (JSAV) is designed as a simple-to-use JavaScript component for displaying sequence alignments on web pages. The display of sequences is highly configurable with options to allow alternative coloring schemes, sorting of sequences and 'dotifying' repeated amino acids. An option is also available to submit selected sequences to another web site, or to other JavaScript code. JSAV is implemented purely in JavaScript making use of the JQuery and JQuery-UI libraries. It does not use any HTML5-specific options to help with browser compatibility. The code is documented using JSDOC and is available from http://www.bioinf.org.uk/software/jsav/.