dyAb: Flow Matching for Flexible Antibody Design with AlphaFold-driven Pre-binding Antigen

TL;DR

dyAb addresses the antibody design challenge under dynamic antigen conformations by leveraging AlphaFold2-predicted pre-binding antigen structures within a two-stage design pipeline. It couples coarse-grained interface alignment with fine-grained flow matching to model the evolution of the antigen–antibody complex and to design antibody sequences alongside structures. The approach optimizes a total loss $L_{total} = L_{Seq} + L_{Str} + L_{ITF}$ and demonstrates superior performance over existing end-to-end and multi-stage baselines on tasks including CDR-H3 generation, affinity optimization, and complex structure prediction. This framework offers a more realistic and efficient route to therapeutic antibody design by explicitly modeling antigen flexibility and binding dynamics.

Abstract

The development of therapeutic antibodies heavily relies on accurate predictions of how antigens will interact with antibodies. Existing computational methods in antibody design often overlook crucial conformational changes that antigens undergo during the binding process, significantly impacting the reliability of the resulting antibodies. To bridge this gap, we introduce dyAb, a flexible framework that incorporates AlphaFold2-driven predictions to model pre-binding antigen structures and specifically addresses the dynamic nature of antigen conformation changes. Our dyAb model leverages a unique combination of coarse-grained interface alignment and fine-grained flow matching techniques to simulate the interaction dynamics and structural evolution of the antigen-antibody complex, providing a realistic representation of the binding process. Extensive experiments show that dyAb significantly outperforms existing models in antibody design involving changing antigen conformations. These results highlight dyAb's potential to streamline the design process for therapeutic antibodies, promising more efficient development cycles and improved outcomes in clinical applications.

Figures (6)

• Figure 1: Alignment of predicted and experimental antigen structures before and after binding. The pre-binding antigen structures, predicted by AlphaFold2, are depicted in yellow. The post-binding antigen structures, derived from experimental data of antigen-antibody complexes, are shown in red. The antibodies are colored in green and blue. The epitopes are highlighted with a red box.
• Figure 2: The overview framework of dyAb. The pre-binding antigen structures are predicted by AlphaFold2 and used as input to the model. dyAb consists of two main components: coarse-grained interface alignment and fine-grained flow matching. The model is trained end-to-end to predict the post-binding antibody-antigen structures and the designed antibody sequences.
• Figure 3: The key steps of the coarse-grained interface alignment process. (a) The pre-binding antigen structure is used to initialize the antibody structure. (b) The antibody structure and the binding interface are predicted by individual models. (c) The antibody structure is aligned to the predicted interface to generate the coarse-grained antibody-antigen complex.
• Figure 4: Fine-grained iterative refinement process of the antibody-antigen complex. The interfaces of both the antigen and the antibody are iteratively refined.
• Figure 5: The overall model architecture of dyAb.