Evaluating data partitioning strategies for accurate prediction of protein-ligand binding free energy changes in mutated proteins

Accurate prediction of the relative free energy of protein-ligand binding, especially regarding protein mutations, is vital for drug design and interpreting drug resistance. However, machine learning (ML) / deep learning (DL) methods often struggle with generalization due to dataset partitioning strategy. Random data partitioning potentially produces spuriously high correlations that inflate performance estimates. UniProt-based splitting preserves data independence but lacks high prediction accuracy. In this study, we first evaluate six distinct ML/DL models on the MdrDB database using two data partitioning methods. Protein sequences are embedded using the ESM-2 protein large language model, integrating wild-type and mutant features. Although all models show high predictive correlations (Pearson coefficients up to 0.70) under random partitioning, their performance declines with UniProt-based partitioning. To address this issue, we propose a query-anchor pairwise learning framework, utilizing known states as anchor points for predicting unknown query states. The proposed method is validated across three systems, revealing that even a small amount of reference data can significantly enhance prediction accuracy. This enhancement suggests that leveraging known states as anchor points allows for more precise predicting of unknown query states.