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

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.

💡 Research Summary

AbFlow introduces a novel end‑to‑end framework for antibody design that focuses on the paratope, particularly the highly variable CDR‑H3 loop, while simultaneously generating full‑atom structures for the entire antibody. The method leverages optimal‑transport‑based flow matching, a continuous normalizing flow (CNF) technique, to learn a time‑dependent velocity field that transports samples from a simple Gaussian distribution to realistic paratope conformations. To keep computation tractable, the flow is restricted to the paratope region; the learned velocity field is implemented with an E(3)‑equivariant graph neural network (EGNN) that also propagates structural information from the generated paratope to the rest of the antibody framework.

A key innovation is the Surface Multi‑channel Encoder (SME), which extracts fine‑grained geometric and chemical features from the antigen surface using multiple channels (distance, normal vectors, electrostatics, hydrophobicity, etc.) and processes them with an equivariant graph convolutional network. These antigen‑derived embeddings are concatenated with the paratope features inside the velocity‑field network, guiding the generation toward interfaces that are geometrically and chemically consistent with the target epitope.

Training minimizes a conditional flow‑matching loss that regresses the predicted velocity field against the true velocity derived from the optimal‑transport interpolation between noisy and target paratope coordinates. Sample pairs are pre‑aligned using the Kabsch algorithm (rotation and translation) and the Hungarian algorithm (atom‑wise permutation) to avoid trajectory crossing. The loss also includes regularization terms to enforce smoothness and equivariance. During inference, the ODE defined by the learned velocity field is integrated forward, yielding high‑quality paratope structures; the EGNN’s message‑passing layers then spread these updates throughout the antibody, producing a coherent full‑atom model.

Extensive experiments cover five major tasks: (1) paratope‑centric design (CDR‑H3 generation), (2) multi‑CDR design, (3) full‑atom antibody generation, (4) binding‑affinity optimization, and (5) antigen‑antibody complex structure prediction. Across all benchmarks, AbFlow outperforms state‑of‑the‑art step‑by‑step pipelines (e.g., MEAN, HERN, Di∂Ab) and end‑to‑end models (e.g., dyMEAN, IgFlow). Notable gains include lower RMSD and higher DockQ scores for the interface, a 15‑25 % improvement in contact‑level accuracy, and an average ΔΔG reduction of 1.8 kcal/mol in Rosetta affinity calculations. Ablation studies demonstrate that removing the SME degrades interface quality (≈0.4 Å increase in RMSD) and that relaxing the paratope‑only flow constraint harms global structural coherence.

Limitations are acknowledged: the current implementation treats only CDR‑H3 as the paratope, leaving other CDRs and framework mutations for future extensions; the EGNN‑based velocity field can be memory‑intensive on very large datasets; and experimental validation beyond in‑silico metrics remains to be performed. Future work will explore multi‑paratope flow matching, richer physicochemical surface modeling, and integration with wet‑lab assays.

The authors release code and data via Zenodo and GitHub, ensuring reproducibility. AbFlow thus establishes a new paradigm that jointly optimizes local antigen‑binding geometry and global antibody architecture, advancing computational antibody engineering toward more reliable, affinity‑optimized designs.