AbFlow : End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching

TL;DR

AbFlow introduces a paratope-centric flow-matching framework for end-to-end full-atom antibody design, leveraging continuous normalizing flows restricted to the paratope and optimal transport to transport probability mass from Gaussian noise to the target paratope distribution. It augments the velocity field with an equivariant Surface Multi-channel Encoder (SME) to articulate fine-grained antigen-surface geometry and propagates this information through an EGNN-based refinement to achieve globally coherent antibody structures. Across paratope-centric design, multi-CDR generation, affinity optimization, and complex structure prediction, AbFlow demonstrates superior interface quality, high docking metrics, and favorable efficiency, outperforming stepwise baselines and several end-to-end competitors. The framework holds promise for accelerated, therapeutically relevant antibody design and can be extended to broader protein–protein interactions and personalized immunotherapies.

Abstract

Antigen-antibody binding is a critical process in the immune response. Although recent progress has advanced antibody design, current methods lack a generative framework for end-to-end modeling of full-atom antibody structures and struggle to fully exploit antigen-specific geometric information for optimizing local binding interfaces and global structures. To overcome these limitations, we introduce AbFlow, a flow-matching framework that leverages optimal transport to design full-atom antibodies end-to-end. AbFlow incorporates an extended velocity field network featuring an equivariant Surface Multi-channel Encoder, which uses surface-level antigen interaction data to refine the antibody structure, particularly the CDR-H3 region. Extensive experiments in paratoep-centric antibody design, multi-CDRs and full-atom antibody design, binding affinity optimization, and complex structure prediction show that AbFlow produces superior antigen-antibody complexes, especially at the contact interface, and markedly improves the binding affinity of generated antibodies.

Figures (5)

• Figure 1: Overall illustration of antibody design algorithm. Given antibody sequences lacking CDR-H3, generate complete sequences and 3D structures. Unlike existing methods that mostly focus on a single step (e.g., CDR-H3 generation) in antibody design tasks and a few end-to-end frameworks, we propose a novel generative end-to-end architecture.
• Figure 2: Overall architecture of AbFlow. Input the antigen epitope and the antibody sequence lacking CDR-H3 as introduced in Section \ref{['sec:pro setting']}. We extract the paratope as the flow for the transport of probability distribution. In the single-step forward process, we introduce an extended velocity field network that not only predicts flow dynamics but also enhances specific-antigen binding interface information through a Surface Multi-channel Encoder (SME), propagating this information from the paratope to the entire antibody structure.
• Figure 3: Case Study of the Full-atom Antibody Design Task. The structures on the left and right show antibodies generated by dyMEAN and AbFlow, respectively. The antibody designed by dyMEAN (yellow) exhibits a distinct steric clash (red circle) with the epitope, while AbFlow (pink) effectively avoids this clash, generating a structure highly similar to the true antibody (green).
• Figure 4: $\Delta\Delta G$ Comparison at Different Length of Changed Residues.
• Figure 5: Schematic diagram of surface definition and message passing. We generate surface vertices within a cutoff of 1.5 Å around the epitope. For each vertex, we assign the nearest residue. Then, we perform message passing between the surface vertices and the antibody paratope using SME to enhance the learning ability of the interaction.

Theorems & Definitions (1)

• theorem 1