Natural-Language-Guided Generator-Agnostic Shortlisting for Protein Binder Design

Modern de novo design workflows generate many candidate protein binders, but wet-lab validation capacity remains limited, making shortlisting a major bottleneck. We study whether LLMs can generate interpretable multi-metric ranking policies from precomputed structural-confidence and interface-quality proxy scores. Rather than reproducing any one design pipeline's internal filtering stack, we evaluate a post-generation task in which candidates are already generated and only the final top-K shortlist is chosen. A global gpt-4o-2024-11-20 policy averaging strategy inferred from development targets slightly outperforms the strongest fixed heuristic, and target-conditioned iterative policy averaging improves further on the 9-target held-out split in target-averaged Recall@10 (0.562 vs. 0.506). On a smaller 2-target subset (Nipah and RBX1), applying the same iterative strategy with gpt-5.5 also exceeds the fixed AF2 and Boltz-2 rules in Recall@10 (0.456). These results suggest that natural-language-guided shortlisting can produce interpretable feature-weighted ranking rules for new binder-design candidate pools.