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

arXiv:2602.07084 · q-bio.QM, cs.AI, cs.CE · Submitted 2026-02-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching".

Jane: The paper was written by Wenda Wang, Yang Zhang, Zhewei Wei and Wenbing Huang from Renmin University of China and China National Institute of Standardization.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: We're starting today with a massive paper titled AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching.

Jane: That is quite a long title, Tom.

Tom: It really is, but the work from Wenda Wang and the team at Renmin University is incredibly important.

Jane: I think we should start by explaining what a paratope actually is for our listeners.

Jane: You can imagine an antibody as a specialized tool that reaches out to grab a specific virus or toxin.

Jane: The paratope is essentially the very tip of that tool that makes the actual contact.

Lu: It reminds me of a key designed for a very complex lock.

Lu: The key has to have the exact right shape to move the tumblers inside.

Lu: If even one tiny part of that key is slightly off, the whole mechanism stays stuck.

Meng: That level of precision sounds like a nightmare to calculate.

Meng: I'm wondering how much computational power it takes to design such a specific shape from scratch.

Meng: Modeling every single atom in that binding site must require an enormous amount of memory.

Lalam: The complexity is high, but the payoff for human health is even higher.

Lalam: If we can automate this, we can respond to new diseases in days instead of months.

Lalam: This could fundamentally change our readiness for the next global health crisis.

Tom: It's a high-stakes game of molecular architecture, and we're about to look at how they actually build it.

Summary: Tom: Now that we know the goal, let's look at the actual mechanism in AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching.

Jane: They've introduced a clever component called the Surface Multi-channel Encoder, or SME.

Tom: That SME module seems to be the heart of the whole system.

Jane: It works by looking at the surface of the antigen, which is the target the antibody is attacking.

Jane: Instead of just guessing, it reads the tiny bumps and grooves on that surface.

Lu: It's like giving a sculptor a high-resolution three dee scan of the mold they are working with.

Lu: They aren't just working in the dark anymore.

Lu: They can see exactly where the clay needs to go to fit the shape perfectly.

Meng: I noticed they aren't trying to move the entire antibody through the "flow" process at once.

Meng: They only apply the flow matching to the paratope region to keep things efficient.

Meng: Does that risk making the rest of the antibody look disconnected or unstable?

Jane: That's a smart observation, Meng, but they solve that with an EGNN-based refinement.

Jane: This allows the structural information to ripple out from the paratope to the rest of the antibody.

Lalam: That ripple effect ensures the whole structure stays coherent and stable.

Lalam: It's a much more efficient way to use biological energy.

Lalam: We're moving toward designs that are optimized for real-world biological function.

Tom: It's a brilliant way to balance local precision with global stability.

Improvements: Tom: We should talk about how AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching actually beats the old methods.

Jane: Most older approaches were step-by-step, which created a lot of room for error.

Tom: They would predict a sequence and then try to dock it later.

Jane: That often led to what researchers call steric clashes.

Jane: If you look at Figure three the older dyMEAN model actually had atoms overlapping with the target.

Lu: An overlap like that is a total dealbreaker.

Lu: You can't have two atoms occupying the same space in a real cell.

Lu: It makes the entire design physically impossible.

Meng: I was checking their efficiency data in Table seven to see if this was practical.

Meng: Even with the extra steps of flow matching, they are still very competitive on time.

Meng: It's a great trade-off between speed and accuracy.

Lalam: It feels like we're finally moving from guessing to actual calculation.

Lalam: We are seeing much better results in binding affinity, which is the real test.

Lalam: They use a metric called ΔΔG to prove their designs bind much more strongly.

Tom: It's a massive leap from just making something that looks right to making something that actually works.

Conclusion: Tom: We've covered a huge amount of ground today on AbFlow: End-to-end Paratope-Centric Antibody Design by Interaction Enhanced Flow Matching.

Jane: From the SME module reading the antigen surface to the way they use flow matching, it's a game-changer.

Lu: I can see a future where we architect our own biological solutions in real-time.

Lu: We won't just be reacting to pathogens; we'll be designing the perfect response.

Meng: From an engineering standpoint, this looks ready for real-world pharmaceutical pipelines.

Meng: If it can run efficiently, it's ready to be used in a lab right now.

Lalam: This turns biology into something as programmable as the software we use every day.

Lalam: Our relationship with medicine will become much more proactive and precise.

Tom: That is a profound thought to end on, Lalam.

Jane: It really is, and we're so glad we could share this deep dive with you.

Tom: Thanks to the whole team for joining us today.

Jane: We'll see you next time for another look at a major breakthrough!

Renmin University of China · China National Institute of Standardization

q-bio.QM, cs.AI, cs.CE

Submitted: 2026-02-06

Updated: 2026-09-12

DOI: 10.1145/3770854.3780296

Code: https://github.com/WangWenda87/AbFlow

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 84/100

The gist: AbFlow is a novel generative framework for "end-to-end full-atom antibody design" that addresses the limitations of existing models in modeling structural information flow and utilizing fine-grained

Key concepts

Paratope
The paratope is the specific tip of an antibody that makes actual contact with a target, such as a virus or toxin. It must have the exact right shape to fit its target perfectly, much like a key must match the tumblers inside a complex lock to function.
Surface Multi-channel Encoder (SME)
The SME is a module that reads the tiny bumps and grooves on an antigen's surface. Instead of guessing, it uses this high-resolution information to see exactly where the antibody needs to fit the shape of the target perfectly.
Steric Clashes
Steric clashes occur when atoms in a molecular design overlap and occupy the same space. This makes a design physically impossible in a real cell, an issue found in older models like dyMEAN that AbFlow's end-to-end approach helps to avoid.

Terminology

Summary

AbFlow is a novel generative framework for end-to-end full-atom antibody design that addresses the limitations of existing models in modeling structural information flow and utilizing fine-grained antigen geometry. By integrating sequence and structure generation with interaction enhancement, it enables the creation of antibodies with high binding affinity and precise interface complementarity, which is critical for therapeutic development.

The Core Framework

AbFlow utilizes a paratope-restricted flow matching module to achieve efficient high-fidelity generation. Instead of applying continuous normalizing flows (CNFs) to the entire antibody—which would cause a calculation efficiency problem due to the large 3D structure—the framework strategically constrains CNFs to focus solely on paratope structures. The model learns a time-dependent velocity field that guides samples from a simple Gaussian distribution to a target data distribution representing true paratope structures. To avoid trajectory crossings that may occur from interpolating randomly paired samples, the authors adopt a method of pre-aligning sample pairs using:

  1. The Kabsch algorithm to compute optimal translation and rotation.

  2. The Hungarian algorithm to compute the optimal permutation.

Interaction Enhancement via SME

To address the lack of antigen-specific geometric context in existing methods, AbFlow introduces an equivariant Surface Multi-channel Encoder (SME). This module extracts fine-grained structural cues from the antigen surface and injects them into the velocity field network, ensuring that paratope generation is conducted in an antigen-aware and interaction-consistent manner. The SME facilitates directional information from the epitope surface to the antibody paratope, which is critical for improving binding affinity.

The process involves several specific steps:

  • Generating a dense set of surface vertices using MSMS within a 1.5 Å radius of the epitope.

  • Assigning attribution residues by identifying the nearest epitope residue for each vertex.

  • Randomly sampling at most M 0 surface vertices per epitope residue to maintain robustness and reduce computational cost.

Velocity Field and Structural Refinement

The architecture features an extended velocity field network that combines the SME with an Atom Multi-channel Encoder (AME). This hybrid backbone serves a dual purpose: the SME focuses on fine-grained surface-level transmission of antigen-epitope information, while the AME handles global structural propagation. This design allows for localized paratope flow modeling while maintaining full-atom generation capabilities.

The training process optimizes a total loss function that incorporates several components:

  1. Flow loss (L F) to model the temporal evolution of paratope coordinates.

  2. Sequence loss (L seq) measuring differences in amino acid types.

  3. Structure loss (L struct), comprising coordinate and chemical bond losses.

  4. Docking loss (L dock) to ensure the quality of the docking interface.

Experimental Validation

Extensive experiments demonstrate that AbFlow produces superior antigen-antibody complexes, especially at the contact interface. The model was evaluated across four critical tasks: paratope-centric antibody design, multi-CDRs and full-atom antibody design, binding affinity optimization, and complex structure prediction.

Key experimental results include:

  • In paratope-centric design, AbFlow achieves the best performance in metrics such as TMscore, lDDT, CAAR, and DockQ.

  • In multi-CDR design, it consistently achieves the lowest RMSD across all CDRs (H1–H3 and L1–L3).

  • In binding affinity optimization, AbFlow outperforms baselines by achieving a higher IMprovement Percentage (IMP) while minimizing the number of altered residues.

  • The model demonstrates high efficiency, completing full-atom generation in comparable or less time than step-by-step methods like DiffAb.

Improvements for AI systems

1. Improvement: Implementation of Paratope-Restricted Flow Matching with Global EGNN Propagation

  • The Improvement: Transition from computationally expensive full-structure diffusion/flow models to a localized flow matching framework. This involves constraining the Continuous Normalizing Flow (CNF) dynamics specifically to the paratope (the primary binding region) while utilizing an Equivariant Graph Neural Network (EGNN) as a velocity field network to propagate structural updates from the local paratope to the entire global antibody framework.

  • What the improved system can do: It can generate high-fidelity, full-atom molecular structures with significantly higher computational efficiency. By focusing generative dynamics on critical interaction residues while maintaining global coherence through message passing, the system avoids the decoupling error where local binding sites are optimized but the overall protein scaffold becomes structurally unstable or unphysical.

2. Improvement: Integration of Equivariant Surface Multi-channel Encoders (SME)

  • The Improvement: Augment atom-based or residue-based graph representations with a dense surface vertex representation. The system should utilize an SME to sample fine-grained geometric and chemical cues from the target's molecular surface (using MSMS or similar density sampling) and inject these multi-channel features into the velocity field network via equivariant message passing.

  • What the improved system can do: It can perform precise interface modeling at a sub-residue level. This allows the AI to sense the exact topography and chemical environment of an antigen's epitope, enabling it to design binders that maximize complementarity, minimize steric clashes (atomic overlaps), and significantly increase predicted binding affinity (G).

3. Improvement: Hybrid Multi-Channel Message Passing (SME + AME) for Co-Design

  • The Improvement: Replace single-scale encoders with a hybrid architecture that combines a Surface Multi-channel Encoder (for fine-grained antigen-antibody interaction) and an Atom Multi-channel Encoder (for global structural propagation). This creates a dual-stream information flow: one stream handles the high-resolution geometric requirements of the binding interface, while the other maintains the structural integrity of the protein backbone.

  • What the improved system can do: It enables true end-to-end sequence-structure co-design. The system can simultaneously output an optimized amino acid sequence and a precise 3D atomic coordinate set that are mutually consistent, producing ready-to-synthesize therapeutic candidates rather than just structural approximations.

4. Improvement: Multi-Objective Loss Function for Affinity and Docking Optimization

  • The Improvement: Shift from simple reconstruction losses to a composite objective function that integrates Flow Matching loss (L CFM), sequence cross-entropy, structural Huber loss (covering both coordinates and chemical bond lengths), and a dedicated docking interface loss (L dock).

  • What the improved system can do: It can perform automated lead optimization. The system can take an existing antibody scaffold and suggest minimal residue mutations that maximize binding affinity improvements (IMP) while ensuring the structural changes remain within physically plausible, stable limits, effectively automating the most expensive stage of antibody engineering.

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 paratope-restricted one-step flow-matching framework for designing 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 paratope-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.

Sources

Related papers