Computationally Evidence‐Grounded Sequence‐First Design of Peptide Binders

Peptide binders provide a versatile modality for modulating protein targets that are poorly addressed by small molecules, but their discovery is constrained by sample-intensive screening or reliance on structural templates. Sequence-first generation offers a scalable alternative for targets lacking stable or representative structures, yet existing approaches often sacrifice target-specific control for diversity and are further limited by the imperfect transfer of protein language-model priors to short peptides. Here, we report BOND-PEP, an evidence-grounded framework for sequence-only peptide binder generation. BOND-PEP retrieves target-relevant peptide exemplars, aligns them with the query protein through bipartite message passing, and uses the resulting protein-centric representation to guide conditional decoding. In a matched AlphaFold-Multimer evaluation on a non-homologous held-out benchmark, BOND-PEP improved reference-beating ipTM success over RFdiffusion, PepPrCLIP and PepMLM. It further transferred to a compact external panel of targets with previously reported experimentally supported peptide binders. These results establish retrieval-augmented, topology-conditioned decoding as a practical route to controllable peptide binder design.