Towards the accurate modelling of antibody-antigen complexes from sequence using machine learning and information-driven docking

Aug 4, 2026

Upload by Hazel

Antibody-antigen complex modelling underpins computational workflows for therapeutic antibody design, yet experimental structures of both partners are often unavailable. The study evaluates whether protein-protein docking can leverage machine-learning generated input structures for information-driven docking. It shows that HADDOCK can produce accurate antibody-antigen models using an ensemble of ML-generated antibody structures together with AlphaFold2-predicted antigen structures, enabling targeted docking of complementary regions for reduced sampling and improved cost-effectiveness beyond ZDOCK.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

AccessFree to download

Introduction

Antibodies, CDRs, and modelling challenges

ML-based antibody modelling and complex prediction limits

Information-driven docking and study aim

Study design and benchmarking scenarios

bioRxiv preprint doi: [https://doi.org/10.1101/2023.11.17.567543](https://doi.org/10.1101/2023.11.17.567543); this version posted November 17, 2023. The copyright holder for this preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made available under aCC-BY-ND 4.0 International license. Towards the accurate modelling of antibody-antigen complexes from sequence using machine learning and information-driven docking Marco Giulini 1 , Constantin Schneider2 , Daniel Cutting2 , Nikita Desai2 , Charlotte M. Deane2 , Alexandre M.J.J. Bonvin 1 1 Bijvoet Centre for Biomolecular Research, Faculty of Science - Chemistry, Utrecht University, Padualaan 8, 3584 Utrecht, CH, The Netherlands 2 Exscientia plc, The Schroedinger Building, OX4 4GE, Oxford, UK Antibody-antigen complex modelling is an important step in computational workflows for therapeutic antibody design. While experimentally determined structures of both antibody and the cognate antigen are often not available, recent advances in machine learning-driven protein modelling have enabled accurate prediction of both antibody and antigen structures. Here, we analyse the ability of protein-protein docking tools to use machine learning generated input structures for information-driven docking. We find that HADDOCK can generate accurate models of antibodyantigen complexes using an ensemble of antibody structures generated by machine learning tools and AlphaFold2 predicted antigen structures. Targeted docking using knowledge of the complementary determining regions on the antibody and some information about the targeted epitope allows the generation of high quality models of the complex with reduced sampling, resulting in a computationally cheap protocol that outperforms the ZDOCK baseline. The data set used to benchmark the docking protocols in this study is available at [github.com/haddocking/ai-antibodies. The](github.com/haddocking/ai-antibodies. The) docking models will be deposited at [data.sbgrid.org/labs/32/ upon](data.sbgrid.org/labs/32/ upon) acceptance. I. INTRODUCTION Antibodies are Y-shaped proteins produced by B-cells that bind with high selectivity and affinity to invading antigens recognized as potentially dangerous by the immune system, making them useful candidates for therapeutics development: as of June 2022, 162 antibody therapeutics have been approved globally [1] . Their highly desirable binding properties are mainly due to the process of somatic hypermutation, in which the complementarity determining regions (CDRs) of the antibody are optimized in order to specifically bind the epitope, theset of amino acids on the antigen molecule engaged by the antibody. The CDRs of an antibody are six hypervariable loops, distributed over the variable regions of the light and heavy chains. The third hypervariable loop on the heavy chain (CDR H3) usually corresponds to the most important region for antigen binding [2], as well asthe most difficult region to be modelled by computational approaches [3–5], due to its high variability in both sequence composition and length. As of 30 Oct 2023, 7853 antibody structures have been collated in the Structural Antibody Database (SAbDab) [6], the majority of which, 7495, are found in complex with the cognate antigen. Following the recent achievements in computer-aided structure prediction [7–9], several studies have shown that it is possible to accurately model antibody structures from sequence information using machine learning (ML)-based approaches [5, 10] . While ML-based antibody modelling tools are already reaching high accuracy [11], the accurate prediction of antibody-antigen complex structures from sequence is a much harder problem, which still represents a challenge for state-of-the-art methods such as AlphaFold2-Multimer [12, 13] . As an example, the three antibody-antigen complexes present in the recent CA

FAQs

Why is modelling antibody-antigen complexes important for therapeutic antibody design?

It supports computational workflows that generate binding hypotheses when experimental structures are not available, enabling downstream design and optimization of therapeutic antibodies.

What is the main question this study investigates?

Whether protein-protein docking tools can use machine-learning generated antibody and antigen structures as inputs for information-driven docking to produce accurate complex models.

Which docking approach and inputs does the study identify as most effective?

HADDOCK using an ensemble of ML-generated antibody structures (e.g., from ABodyBuilder2) together with AlphaFold2-predicted antigen structures, with targeted docking guided by complementary determining regions and epitope information.

Aug 4, 2026

Upload by Hazel