A Comparative Study of Deep Learning and Classical Modeling Approaches for Protein–Ligand Binding Pose and Affinity Prediction in Coronavirus Main Proteases

The accurate prediction of protein-ligand binding poses and affinities is central to structure-based drug design. In this study, we first benchmarked three distinct pose generation strategies for data sets from the ASAP Antiviral Challenge 2025: molecular docking (Glide and AutoDock Vina), ligand-based superposition (FlexS), and deep learning-based modeling (AlphaFold3, Boltz-2, DiffDock and Gnina). We evaluated their performance on binding pose prediction for ligands targeting SARS-CoV-2 and MERS-CoV main protease (Mpro). For binding affinity estimation, we implemented a machine learning-based scoring approach called ligand-residue interaction profile scoring function (LRIP-SF), which integrates molecular mechanics generalized Born surface area (MM-GBSA) energy decomposition with machine learning algorithms. Our results showed that deep learning-based modeling with AlphaFold3 achieved the highest pose prediction accuracy with a success rate of 88.1% and an average ligand root-mean-square deviation (LRMSD) of 1.12 Å. Moreover, binding poses predicted by AlphaFold3 enabled the most accurate potency predictions by LRIP-SF, with the lowest mean absolute error (MAE) and root-mean-square error (RMSE) in pIC₅₀ units across both targets: the MAE and RMSE are 0.606 and 0.813, respectively, for MERS-CoV Mpro and 0.724 and 0.894 respectively for SARS-CoV-2 Mpro. Although ligand-based superposition method (FlexS) was less accurate in pose prediction, it offered competitive potency prediction performance with significantly lower computational cost. To interpret model predictions by LRIP-SF and identify critical binding determinants, we performed global sensitivity analysis (GSA), revealing key residues that contributed most significantly to ligand binding. These findings highlight the importance of pose quality and interaction profiling in affinity prediction and demonstrate the great potential of deep learning-based methods for drug discovery, especially in the absence of cocrystal structures.