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 multi-metric ranking policies from precomputed structural-confidence and interface-quality proxy scores. Rather than proposing a new protein binder design pipeline, we focus on post-generation binder shortlisting: selecting the final top-K candidates from already generated binder pools using a shared panel of precomputed proxy scores. On the 10-target held-out split, averaging performance over five sampled global iterative gpt-4o policies reaches 0.589 Recall@10, modestly improving over the strongest single-feature fixed baseline, Protenix binder ipTM, which reaches 0.571 Recall@10. On the 3-target held-out subset comprising Nipah, RBX1, and TREM2, target-conditioned iterative gpt-5.4 policies reach the strongest LLM performance, with 0.519 Recall@10 and 0.583 NDCG@10. These results suggest that LLM-generated ranking policies can act as an interpretable post-generation decision layer for combining heterogeneous proxy metrics to prioritize binders from large candidate pools.
Gyubok Lee, Kiwoong Yoo, Jimin Seo et al.· 0 citations
Abstract Motivation Binding affinity governs how proteins interact and underlies essential biological processes. Computational approaches have been developed to simulate and predict protein binding, but the scarcity of high-quality data has imposed significant constraints. One consequence is that most methods focus on predicting mutational changes in binding affinity (ΔΔG), rather than binding affinity (ΔG) itself. This practice risks overfitting to skewed data distributions, limiting the generalizability of predictions. Recent advances in protein structure prediction have enabled computational modeling of protein conformations in mass, providing rich structural information from which binding interactions can be largely explained. However, leveraging these advances for effective prediction of binding affinity has yet to translate into reliable predictions. Results We present PreFold-dG, a model that estimates binding affinities of protein complexes utilizing intermediate embeddings from Boltz-2, an open-source foundation model for protein structure prediction. Our approach aggregates residue-level information weighted by interresidue distance, and predicts ΔG directly rather than its derivative, ΔΔG. PreFold-dG achieved state-of-the-art performance on well-established binding affinity prediction benchmarks and demonstrated robustness on independent test sets. Ablation studies suggest that all intermediate embeddings are utilized in the prediction, whereas their contributions to modeling ΔΔG and ΔG vary. We further validated our model through case studies on real-world broadly neutralizing antibody data with evolutionary relevance. Availability https://github.com/LGAI-Research/PreFold-dG.
Sungjoon Park, Soorin Yim, Dongyun Kim et al.· Bioinformatics· 0 citations