Improved Prediction of Ligand–Protein Binding Affinities by Meta-modeling • This study introduces a meta-modeling framework integrating classical docking scores and sequence-based deep learning (DL) predictions to enhance ligand-protein binding affinity predictions. • The meta-models significantly outperform individual docking (SMINA, Vinardo) and DL models, achieving a 20-30% improvement in Pearson correlation and a reduction in mean squared error across key benchmarks like CASF-2016 and PDBbind CoreSet. • By leveraging diverse input features (e.g., docking scores, DL embeddings, and physicochemical properties), the meta-models combine the strengths of physics-based and data-driven approaches, providing robust predictions even in challenging datasets. • The framework employs fine-tuning of DL models on curated datasets (e.g., PDBbind), enabling transfer learning that improves accuracy for unseen protein-ligand complexes. • Principal component analysis (PCA) is applied to unify predictions from multiple models, with the first principal component (PC1) consistently showing the best correlation with experimental binding affinities. • Extensive validation includes external datasets like LIT-PCBA for virtual screening, demonstrating that the meta-models excel in ranking active ligands while maintaining generalization to diverse chemical spaces. • The study provides a reproducible pipeline for integrating hybrid modeling techniques, setting a new standard for binding affinity predictions in computational drug discovery. @MarkGerstein @GersteinLab 📜Paper: pubs.acs.org/doi/10.1021/acs… #DrugDiscovery #Bioinformatics #DeepLearning #MachineLearning #MetaModeling #ProteinLigand #OpenScience