Our pipeline consisted of four stages: target analysis, generative sampling, filtering, and ranking.

In the target analysis stage, we started from a reference binder-target complex and selected binder interface residues with any atom located within 8 A of the target. These residues defined the redesign region, allowing the method to focus sequence-structure optimization on the local binding interface while preserving the overall binding context from the reference complex.

In the generative sampling stage, we used ODesign to redesign the selected interface region and then applied ProteinMPNN for inverse folding to generate amino acid sequences compatible with the designed structures. This ODesign plus ProteinMPNN combination formed the generator component of the pipeline.

In the filtering stage, all candidates were refolded and scored with AlphaFold 3 and Rosetta. From AF3, we used binder pTM and ipae_min as the main confidence signals. From Rosetta, we used ddg, CMS, and sap_score to evaluate binding energetics, interface quality, and developability-related surface hydrophobicity. These scores were used jointly as an orthogonal verifier framework to remove low-confidence candidates before final ranking.

In the ranking stage, candidates with acceptable sap_score were prioritized by Rosetta ddg. This ranking rule emphasized favorable predicted binding energetics while retaining candidates that passed structural confidence and interface-quality filters.

The proposed generation-and-verification framework efficiently produced an in-silico hit rate of approximately 5% across the sampled candidate pool.