Search-Based Inference-Time Scaling for All-Atom Protein Binder Design

All-atom protein binder design with diffusion models follows a costly generate-and-filter pipeline in which candidates are sampled, redesigned, and refolded; for difficult targets, fewer than 5\% of samples survive. Rather than scaling by generating more candidates, we apply diffusion inference-time search methods---steering computation within the denoising trajectory toward structures that are more likely to be designable. We introduce lightweight rewards based on confidence-head predictions and geometric-decoding/inverse-folding self-consistency, evaluated on intermediate denoised estimates without running the full refolding pipeline. Protein-specific adaptations, including noise-level gating, adaptive quantile thresholds, and failure-scaled exploration noise, address the unique failure modes of all-atom diffusion intermediates. On six difficult binder targets, our method reaches equivalent prefold confidence with $5$-$10\times$ fewer function evaluations and increases the throughput of designable binders by $4$-$8\times$.