Benchmarking and Consensus Ranking of Inverse Folding Models for Protein-Ligand Interface Design

Machine learning has advanced the progress of protein design, also enabling more efficient and accurate modeling of protein-ligand interfaces. Due to the complexity of biological systems, selecting optimal candidates from the heterogeneous outputs of generative protein design tools remains a persistent challenge. In this work, we introduce a consensus ranking framework that integrates five state-of-the-art inverse folding models — ProteinMPNN, LigandMPNN, ESM-IF1, CARBonAra, and ProRefiner — applied to 25,716 curated protein-ligand complexes from the BioLip database. Our approach frames design selection as a supervised learning-to-rank problem and leverages a LightGBM-based LambdaMART model to fuse heterogeneous scoring features into a unified ranking. We pointed out that consensus-ranked sequences outperform individual model selections in stability, binding affinity, and structural fidelity, as evaluated using Schrödinger and MOE free energy difference calculations. In a case study on three enzymes (NOV1, CYP153A, and LCD), our method consistently improves design quality, suggesting that consensus ranking can significantly enhance the success rate and efficiency of AI-driven protein engineering.