ArcRefine improves de novo binder designs while preserving binding modes

Abstract

De novo protein design now enables the creation of high-affinity binders from a target structure alone. However, current design methods face a trade-off between site control and success rate. Generative models such as RFdiffusion and BoltzGen place binders at chosen sites but often require large candidate pools. Hallucination methods such as BindCraft and Mosaic update sequences using gradients from structure predictors and achieve high success rates, but they build binders from random sequences and cannot refine a given design while keeping its binding mode. We introduce ArcRefine, a gradient-based refinement method that uses structural carryover to preserve a binder's fold and binding mode during sequence optimization. It works by reusing Boltz-2's single and pair representations to condition each new prediction on the previous complex. Despite changes to about two-thirds of the amino acid residues, fresh predictions retained the starting interface in 45/60 PD-L1 and 23/60 TNFα designs. Mean interface confidence (ipSAE) increased in all six other predictors tested, with pooled median ipSAE rising from 0.49 to 0.69. Using this approach, we obtained six BLI-confirmed binders from ten submitted designs against TNFα, a target with few reported de novo binders, at a campaign compute cost of about US$700. Carryover can be combined with additional differentiable loss terms and also extends to the design of binders to DNA and small molecules. This record contains the manuscript and supplementary data. This manuscript is a preprint and has not been peer reviewed. Code: https://github.com/ken-osumi/ArcRefineResearch overview and full text: https://foldarc.com/research/structural-carryover/