Sequence-Structure Co-Generation for Antibody Variable Domains: Modeling Conformational Ensembles and Antigen-Conditional Design

Due to their high specificity and integration into the host immune system, antibodies are effective therapeutics for diseases such as cancer and autoimmunity, but their development into viable drugs currently proceeds through extensive experimental screening and optimisation campaigns. Computational de novo design promises a faster alternative whereby generative deep learning models provide protein structure backbones likely to bind to a target epitope. Inverse folding can then infer amino acid sequences from backbone coordinates, which can be tested for efficacy and developability. While this two-step design process has shown early promise, the generative steps are disconnected and do not reflect that sequence and structure are intrinsically linked. For this reason, state-of-the-art de novo generation efforts increasingly focus on simultaneous sequence-structure cogeneration (co-gen). This approach promises several advantages over the two-step process, such as improved overall generative performance, as the two modalities inform each other, and a simplified, single-step workflow. It also enables the application of a single model to several critical tasks, including structure prediction by conditioning on sequence, inverse-folding by conditioning on structure, and conditional design with partial sequence and structure information (such as in-painting a missing CDRH3 or sequence humanisation). Furthermore, the CDRs of some antibodies may exhibit extensive flexibility and/or multiple stable conformations, which can impact antigen binding, stability or aggregation. Recent work on general proteins suggests that generative models can excel at rapid and accurate equilibrium conformational sampling, even without ever being trained on dynamics datasets, and could plausibly displace computationally intensive molecular dynamics simulations for this purpose. However, this tremendous potential remains almost entirely unexplored in the antibody space. This project aims to develop the first sequence-structure co-gen model specific to human antibody variable domains. Through antigen-conditional generation, it will tackle a grand challenge in discovering novel therapeutic candidates and further enable rapid conformational ensemble modelling to accelerate their development. If successful, it will thus address a number of critical antibody design applications with a single tool. This project falls within the EPSRC research areas of Synthetic Biology, Biological Informatics, Mathematical Biology, and Artificial Intelligence Technologies and is pursued in collaboration with AstraZeneca. All research outputs will be made publicly available.