Modern de novo protein binder design pipelines can generate thousands of candidates, yet laboratory validation capacity is limited, making the shortlisting step a major bottleneck. This work investigates whether large language models (LLMs) can synthesize multi‑metric ranking policies from pre‑computed structural‑confidence and interface‑quality proxy scores, thereby serving as a post‑generation decision layer. Rather than proposing a new design workflow, we focus on the post‑generation shortlisting problem: selecting the final top‑K binders from existing pools using a shared panel of proxy scores.
On a held‑out split of ten targets, averaging five sampled global iterative gpt‑4o policies yields a Recall@10 of 0.589, modestly surpassing the strongest single‑feature fixed baseline, Protenix binder ipTM (Recall@10 = 0.571). On a three‑target subset (Nipah, RBX1, TREM2), target‑conditioned iterative gpt‑5.4 policies achieve the best LLM performance, with Recall@10 of 0.519 and NDCG@10 of 0.583.
These results suggest that LLM‑generated ranking policies can act as an interpretable post‑generation layer, combining heterogeneous proxy metrics to prioritize binders from large candidate pools.
Blogger's Review: The paper convincingly demonstrates the practical utility of LLMs in the downstream filtering stage of protein design, showing that natural‑language‑driven policy synthesis can flexibly integrate diverse scores while remaining transparent and scalable.