Summary
A bioRxiv preprint presents ForceFlowAb, a mixture-of-experts flow-matching model that designs antibody CDR sequences and structures with force-field guidance. The authors report more favourable antibody–antigen interaction-energy results than two comparison methods in CDR-H3 and six-CDR design tasks.
Researchers have presented ForceFlowAb, a generative model designed to create antibody complementarity-determining regions, or CDRs, against a specified antigen. The bioRxiv preprint describes a mixture-of-experts flow-matching system that generates CDR amino-acid sequences and structures while using differentiable force-field guidance to favour physically favourable antibody–antigen interactions.
The authors, Zonghui Li, Zexin Lv and Guijun Zhang of Zhejiang University of Technology, report better interaction-energy results than existing comparison methods for both CDR-H3 design and simultaneous design of all six antibody CDRs. The work is a preprint describing model development and computational evaluation; its supplied abstract does not include experimental binding measurements or biological testing.
How ForceFlowAb combines generation with physical guidance
Antibodies recognise target molecules through parts of their structure called complementarity-determining regions. These loops, particularly the heavy-chain CDR3 region known as CDR-H3, help form the binding interface with an antigen. Designing them requires selecting amino-acid sequences that can also adopt suitable three-dimensional shapes.
ForceFlowAb treats this as a sequence–structure co-design problem. Rather than generating only a sequence or only a predicted conformation, the framework is intended to produce both aspects together while conditioning the generation process on the target antigen.
Its mixture-of-experts design addresses the fact that antibody–antigen interfaces can have different local environments. In this type of architecture, specialised expert components handle different patterns and a routing mechanism directs parts of the generation process to the relevant experts. The authors say this adaptive routing is intended to capture diversity in possible binding modes.
The second component is physics-aware sampling. Flow matching is a generative approach that learns how to transform an initial distribution into samples resembling a target data distribution. During this process, ForceFlowAb adds guidance from a differentiable force field. Because the guidance can provide a direction for changing the generated structure, it is used to steer candidates towards more favourable calculated interaction energies and physically plausible conformations.
Reported results for CDR design
For CDR-H3 design, the preprint reports an improvement-rate, or IMP, value of 46.5% for ForceFlowAb. The corresponding reported values were 35.0% for FlowDesign and 35.5% for Diffab.
The authors also evaluated simultaneous design of all six CDRs. In that task, ForceFlowAb had a reported IMP value of 16%, compared with 9% for Diffab. The paper attributes the results to complementary contributions from its two main design choices: expert routing for representing different interface environments and energy-based guidance for applying physical constraints during sampling.
Interaction energy is a computational description of how favourable a proposed antibody–antigen interface is under a chosen scoring or force-field model. It is useful for ranking or filtering generated candidates, but it is not the same as a direct measurement of binding affinity. The reported improvements therefore describe the model's design-stage evaluation rather than a demonstrated therapeutic antibody.
Availability and evidence level
The authors have made a ForceFlowAb web server freely available and provide implementation code through a public GitHub repository. The study is posted on bioRxiv as a preprint and has been declared to have no competing interests.
The supplied abstract does not give the benchmark sample sizes, dataset composition, absolute interaction-energy values or experimental follow-up. Those details are important for judging how broadly the reported gains apply and whether generated candidates retain their predicted properties when produced and tested in the laboratory. For now, ForceFlowAb is best understood as a computational antibody-design framework that combines adaptive generative modelling with physics-based steering.