Summary
A bioRxiv preprint describes MIMOSA, a framework that guides antibody-generating models using the structure of a known binder interface. The authors report experimentally validated binders against three targets, with hit rates of 20–26%.
A bioRxiv preprint introduces MIMOSA, a framework that steers generative models towards antibody designs using the structural features of a known antibody–target interface. The authors report experimentally validated binders against KEAP1, uPA and IL-8, with hit rates of 20–26% and measured affinities ranging from sub-micromolar to single-digit nanomolar.
Using a known interface to guide design
De novo antibody design aims to generate an antibody that binds a chosen target, rather than starting from an antibody already found to bind it. In the approach described in the preprint, MIMOSA uses the geometry and chemistry of a target’s known cognate binder interface as a guide. The generative model can then design the surrounding antibody interface around those features.
The framework is designed for use at inference time: the authors say it was applied without retraining the underlying generators. They tested it with two different systems, RFAntibody and BoltzGen. This makes MIMOSA a way to constrain existing design pipelines, rather than a new generator that replaces them.
The study considers two kinds of interface features. Some form contiguous motifs that fit within a single complementarity-determining region (CDR) loop, a variable part of an antibody involved in target recognition. Others are hotspots distributed across the antibody’s binding surface, or paratope. The framework was tested with both arrangements.
Reported results and significance
The authors report binders for three targets—KEAP1, uPA and IL-8—where unconstrained design methods produced zero or near-zero binders. Across the reported results, MIMOSA achieved hit rates of 20–26%; the abstract does not assign a specific rate to each target. The resulting binders had affinities from below one micromolar down to single-digit nanomolar levels.
The approach addresses a practical challenge in generative design: producing plausible protein structures is not enough if the designed molecules fail to bind experimentally. Guiding generation with an interface already associated with binding may narrow the search space. In principle, this could make some targets more approachable when a structurally characterised cognate binder is available.
These findings are an early design result reported in a preprint, with experimental binding evidence rather than clinical evidence. MIMOSA’s stated strategy also depends on having a known binder interface with structural information; its reach therefore depends on such information being available for a target.