Developing therapeutic antibodies is a challenging endeavor, often requiring large-scale screening to produce initial binders, that still often require optimization for developability. We present a computational pipeline for the discovery and design of therapeutic antibody candidates, which incorporates physics- and AI-based methods for the generation, assessment, and validation of candidate antibodies with improved developability against diverse epitopes, via efficient few-shot experimental screens. We demonstrate that these orthogonal methods can lead to promising designs. We evaluated our approach by experimentally testing a small number of candidates against multiple SARS-CoV-2 variants in three different tasks: (i) traversing sequence landscapes of binders, we identify highly sequence dissimilar antibodies that retain binding to the Wuhan strain, (ii) rescuing binding from escape mutations, we show up to 54% of designs gain binding affinity to a new subvariant and (iii) improving developability characteristics of antibodies while retaining binding properties. These results together demonstrate an end-to-end antibody design pipeline with applicability across a wide range of antibody design tasks. We experimentally characterized binding against different antigen targets, developability profiles, and cryo-EM structures of designed antibodies. Our work demonstrates how combined AI and physics computational methods improve productivity and viability of antibody designs.

Developing therapeutic antibodies is a challenging endeavour, often requiring large-scale screening to produce initial binders, that still often require optimisation for developability. We present a computational pipeline for the discovery and design of therapeutic antibody candidates, which incorporates physics- and AI-based methods for the generation, assessment, and validation of developable candidate antibodies against diverse epitopes, via efficient few-shot experimental screens. We demonstrate that these orthogonal methods can lead to promising designs. We evaluated our approach by experimentally testing a small number of candidates against multiple SARS-CoV-2 variants in three different tasks: (i) traversing sequence landscapes of binders, we identify highly sequence dissimilar antibodies that retain binding to the Wuhan strain, (ii) rescuing binding from escape mutations, we show up to 54% of designs gain binding affinity to a new subvariant and (iii) improving developability characteristics of antibodies while retaining binding properties. These results together demonstrate an end-to-end antibody design pipeline with applicability across a wide range of antibody design tasks. We experimentally characterised binding against different antigen targets, developability profiles, and cryo-EM structures of designed antibodies. Our work demonstrates how combined AI and physics computational methods improve productivity and viability of antibody designs. ### Competing Interest Statement F.D., C.S., A.K., D.C., M.B., D.N., N.W., H.K., C.M., D.E., R.G., D.D., P.T., W.B., J.D., D.P., S.S. and C.D. are or have previously been employees of Exscientia. I.D. is an employee of Thermo Fisher Scientific.

Monoclonal antibody therapeutics are often produced from non-human sources (typically murine), and can therefore generate immunogenic responses in humans. Humanization procedures aim to produce antibody therapeutics that do not elicit an immune response and are safe for human use, without impacting efficacy. Humanization is normally carried out in a largely trial-and-error experimental process. We have built machine learning classifiers that can discriminate between human and non-human antibody variable domain sequences using the large amount of repertoire data now available. Our classifiers consistently outperform existing best-in-class models, and our output scores exhibit a negative relationship with the experimental immunogenicity of existing antibody therapeutics. We used our classifiers to develop a novel, computational humanization tool, Hu-mAb, that suggests mutations to an input sequence to reduce its immunogenicity. For a set of existing therapeutics with known precursor sequences, the mutations suggested by Hu-mAb show significant overlap with those deduced experimentally. Hu-mAb is therefore an effective replacement for trial- and-error humanization experiments, producing similar results in a fraction of the time. Hu-mAb is freely available to use at opig.stats.ox.ac.uk/webapps/humab.

The naïve antibody/B-cell receptor (BCR) repertoires of different individuals ought to exhibit significant functional commonality, given that most pathogens trigger an effective antibody response to immunodominant epitopes. Sequence-based repertoire analysis has so far offered little evidence for this phenomenon. For example, a recent study estimated the number of shared (‘public’) antibody clonotypes in circulating baseline repertoires to be around 0.02% across ten unrelated individuals. However, to engage the same epitope, antibodies only require a similar binding site structure and the presence of key paratope interactions, which can occur even when their sequences are dissimilar. Here, we search for evidence of geometric similarity/convergence across human antibody repertoires. We first structurally profile naïve (‘baseline’) antibody diversity using snapshots from 41 unrelated individuals, predicting all modellable distinct structures within each repertoire. This analysis uncovers a high (much greater than random) degree of structural commonality. For instance, around 3% of distinct structures are common to the ten most diverse individual samples (‘Public Baseline’ structures). Our approach is the first computational method to find levels of BCR commonality commensurate with epitope immunodominance and could therefore be harnessed to find more genetically distant antibodies with same-epitope complementarity. We then apply the same structural profiling approach to repertoire snapshots from three individuals before and after flu vaccination, detecting a convergent structural drift indicative of recognising similar epitopes (‘Public Response’ structures). We show that Antibody Model Libraries derived from Public Baseline and Public Response structures represent a powerful geometric basis set of low-immunogenicity candidates exploitable for general or target-focused therapeutic antibody screening.

MOTIVATION:T-cell receptors (TCRs) are immune proteins that primarily target peptide antigens presented by the major histocompatibility complex. They tend to have lower specificity and affinity than their antibody counterparts, and their binding sites have been shown to adopt multiple conformations, which is potentially an important factor for their polyspecificity. None of the current TCR-modelling tools predict this variability which limits our ability to accurately predict TCR binding.RESULTS:We present TCRBuilder, a multi-state TCR structure prediction tool. Given a paired αβTCR sequence, TCRBuilder returns a model or an ensemble of models covering the potential conformations of the binding site. This enables the analysis of structurally driven polyspecificity in TCRs, which is not possible with existing tools.AVAILABILITY AND IMPLEMENTATION:http://opig.stats.ox.ac.uk/resources.CONTACT:deane@stats.ox.ac.uk.SUPPLEMENTARY INFORMATION:Supplementary data are available at Bioinformatics online.

Antibodies are vital proteins of the immune system that recognize potentially harmful molecules and initiate their removal. Mammals can efficiently create vast numbers of antibodies with different sequences capable of binding to any antigen with high affinity and specificity. Because they can be developed to bind to many disease agents, antibodies can be used as therapeutics. In an organism, after antigen exposure, antibodies specific to that antigen are enriched through clonal selection, expansion, and somatic hypermutation. The antibodies present in an organism therefore report on its immune status, describe its innate ability to deal with harmful substances, and reveal how it has previously responded. Next-generation sequencing technologies are being increasingly used to query the antibody, or B-cell receptor (BCR), sequence repertoire, and the amount of BCR data in public repositories is growing. The Observed Antibody Space database, for example, currently contains over a billion sequences from 68 different studies. Repertoires are available that represent both the naive state (i.e.antigen-inexperienced) and that after immunization. This wealth of data has created opportunities to learn more about our immune system. In this review, we discuss the many ways in which BCR repertoire data have been or could be exploited. We highlight its utility for providing insights into how the naive immune repertoire is generated and how it responds to antigens. We also consider how structural information can be used to enhance these data and may lead to more accurate depictions of the sequence space and to applications in the discovery of new therapeutics.

The antibody repertoires of different individuals ought to exhibit significant functional commonality, given that most pathogens trigger a successful immune response in most people. Sequence-based approaches have so far offered little evidence for this phenomenon. For example, a recent study estimated the number of shared (‘public’) antibody clonotypes in circulating baseline repertoires to be around 0.02% across ten unrelated individuals. However, to engage the same epitope, antibodies only require a similar binding site structure and the presence of key paratope interactions, which can occur even when their sequences are dissimilar. Here, we investigate functional convergence in human antibody repertoires by comparing the anti-body structures they contain. We first structurally profile base-line antibody diversity (using snapshots from 41 unrelated individuals), predicting all modellable distinct structures within each repertoire. This analysis uncovers a high (much greater than random) degree of structural commonality. For instance, around 3% of distinct structures are common to the ten most diverse individual samples (‘Public Baseline’ structures). Our approach is the first computational method to provide support for the long-assumed levels of baseline repertoire functional commonality. We then apply the same structural profiling approach to repertoire snapshots from three individuals before and after flu vaccination, detecting a convergent structural drift indicative of recognising similar epitopes (‘Public Response’ structures). Antibody Model Libraries derived from Public Baseline and Public Response structures represent a powerful geometric basis set of low-immunogenicity candidates exploitable for general or target-focused therapeutic antibody screening.

Most current analysis tools for antibody next-generation sequencing data work with primary sequence descriptors, leaving accompanying structural information unharnessed. We have used novel rapid methods to structurally characterize the complementary-determining regions (CDRs) of more than 180 million human and mouse B-cell receptor (BCR) repertoire sequences. These structurally annotated CDRs provide unprecedented insights into both the structural predetermination and dynamics of the adaptive immune response. We show that B-cell types can be distinguished based solely on these structural properties. Antigen-unexperienced BCR repertoires use the highest number and diversity of CDR structures and these patterns of naïve repertoire paratope usage are highly conserved across subjects. In contrast, more differentiated B-cells are more personalized in terms of CDR structure usage. Our results establish the CDR structure differences in BCR repertoires and have applications for many fields including immunodiagnostics, phage display library generation, and "humanness" assessment of BCR repertoires from transgenic animals. The software tool for structural annotation of BCR repertoires, SAAB+, is available at https://github.com/oxpig/saab_plus.

Therapeutic mAbs must not only bind to their target but must also be free from "developability issues" such as poor stability or high levels of aggregation. While small-molecule drug discovery benefits from Lipinski's rule of five to guide the selection of molecules with appropriate biophysical properties, there is currently no in silico analog for antibody design. Here, we model the variable domain structures of a large set of post-phase-I clinical-stage antibody therapeutics (CSTs) and calculate in silico metrics to estimate their typical properties. In each case, we contextualize the CST distribution against a snapshot of the human antibody gene repertoire. We describe guideline values for five metrics thought to be implicated in poor developability: the total length of the complementarity-determining regions (CDRs), the extent and magnitude of surface hydrophobicity, positive charge and negative charge in the CDRs, and asymmetry in the net heavy- and light-chain surface charges. The guideline cutoffs for each property were derived from the values seen in CSTs, and a flagging system is proposed to identify nonconforming candidates. On two mAb drug discovery sets, we were able to selectively highlight sequences with developability issues. We make available the Therapeutic Antibody Profiler (TAP), a computational tool that builds downloadable homology models of variable domain sequences, tests them against our five developability guidelines, and reports potential sequence liabilities and canonical forms. TAP is freely available at opig.stats.ox.ac.uk/webapps/sabdab-sabpred/TAP.php.

The Therapeutic Structural Antibody Database (Thera-SAbDab; http://opig.stats.ox.ac.uk/webapps/therasabdab) tracks all antibody- and nanobody-related therapeutics recognized by the World Health Organisation (WHO), and identifies any corresponding structures in the Structural Antibody Database (SAbDab) with near-exact or exact variable domain sequence matches. Thera-SAbDab is synchronized with SAbDab to update weekly, reflecting new Protein Data Bank entries and the availability of new sequence data published by the WHO. Each therapeutic summary page lists structural coverage (with links to the appropriate SAbDab entries), alignments showing where any near-matches deviate in sequence, and accompanying metadata, such as intended target and investigated conditions. Thera-SAbDab can be queried by therapeutic name, by a combination of metadata, or by variable domain sequence - returning all therapeutics that are within a specified sequence identity over a specified region of the query. The sequences of all therapeutics listed in Thera-SAbDab (461 unique molecules, as of 5 August 2019) are downloadable as a single file with accompanying metadata.

Motivation Accurate prediction of loop structures remains challenging. This is especially true for long loops where the large conformational space and limited coverage of experimentally determined structures often leads to low accuracy. Co-evolutionary contact predictors, which provide information about the proximity of pairs of residues, have been used to improve whole-protein models generated through de novo techniques. Here we investigate whether these evolutionary constraints can enhance the prediction of long loop structures. Results As a first stage, we assess the accuracy of predicted contacts that involve loop regions. We find that these are less accurate than contacts in general. We also observe that some incorrectly predicted contacts can be identified as they are never satisfied in any of our generated loop conformations. We examined two different strategies for incorporating contacts, and on a test set of long loops (10 residues or more), both approaches improve the accuracy of prediction. For a set of 135 loops, contacts were predicted and hence our methods were applicable in 97 cases. Both strategies result in an increase in the proportion of near-native decoys in the ensemble, leading to more accurate predictions and in some cases improving the root-mean-square deviation of the final model by more than 3 angstrom. Supplementary information Supplementary data are available at Bioinformatics online.

Motivation:Protein function is often facilitated by the existence of multiple stable conformations. Structure prediction algorithms need to be able to model these different conformations accurately and produce an ensemble of structures that represent a target's conformational diversity rather than just a single state. Here, we investigate whether current loop prediction algorithms are capable of this. We use the algorithms to predict the structures of loops with multiple experimentally determined conformations, and the structures of loops with only one conformation, and assess their ability to generate and select decoys that are close to any, or all, of the observed structures. Results:We find that while loops with only one known conformation are predicted well, conformationally diverse loops are modelled poorly, and in most cases the predictions returned by the methods do not resemble any of the known conformers. Our results contradict the often-held assumption that multiple native conformations will be present in the decoy set, making the production of accurate conformational ensembles impossible, and hence indicating that current methodologies are not well suited to prediction of conformationally diverse, often functionally important protein regions. Contact:marks@stats.ox.ac.uk. Supplementary information:Supplementary data are available at Bioinformatics online.

AbstractMotivationLoops are often vital for protein function, however, their irregular structures make them difficult to model accurately. Current loop modelling algorithms can mostly be divided into two categories: knowledge-based, where databases of fragments are searched to find suitable conformations and ab initio, where conformations are generated computationally. Existing knowledge-based methods only use fragments that are the same length as the target, even though loops of slightly different lengths may adopt similar conformations. Here, we present a novel method, Sphinx, which combines ab initio techniques with the potential extra structural information contained within loops of a different length to improve structure prediction.ResultsWe show that Sphinx is able to generate high-accuracy predictions and decoy sets enriched with near-native loop conformations, performing better than the ab initio algorithm on which it is based. In addition, it is able to provide predictions for every target, unlike some knowledge-based methods. Sphinx can be used successfully for the difficult problem of antibody H3 prediction, outperforming RosettaAntibody, one of the leading H3-specific ab initio methods, both in accuracy and speed.Availability and ImplementationSphinx is available at http://opig.stats.ox.ac.uk/webapps/sphinx.Supplementary informationSupplementary data are available at Bioinformatics online.

Antibodies are proteins of the immune system that are able to bind to a huge variety of different substances, making them attractive candidates for therapeutic applications. Antibody structures have the potential to be useful during drug development, allowing the implementation of rational design procedures. The most challenging part of the antibody structure to experimentally determine or model is the H3 loop, which in addition is often the most important region in an antibody's binding site. This review summarises the approaches used so far in the pursuit of accurate computational H3 structure prediction.

SAbPred is a server that makes predictions of the properties of antibodies focusing on their structures. Antibody informatics tools can help improve our understanding of immune responses to disease and aid in the design and engineering of therapeutic molecules. SAbPred is a single platform containing multiple applications which can: number and align sequences; automatically generate antibody variable fragment homology models; annotate such models with estimated accuracy alongside sequence and structural properties including potential developability issues; predict paratope residues; and predict epitope patches on protein antigens. The server is available at http://opig.stats.ox.ac.uk/webapps/sabpred.