Integrating modern computational techniques elevates binding mode identification for molecular fragments

Molecular docking is a routine and essential procedure in computational drug discovery. However, it is well known that current techniques are still far from perfect, especially when being applied to molecular fragments. In this work, we combined advantages of traditional energy-based docking, a diffusion model, intermolecular interaction filters, and ML-assisted scoring function to achieve higher accuracy. We compiled a comprehensive dataset of molecular fragments to evaluate our new protocol - HybriDock - in comparison with three other docking algorithms including DiffDock, Matcha and AutoDock Vina. HybriDock demonstrated remarkable improvements over these methods, highlighting the advantage of combining modern molecular modeling approaches. Additionally, we analyzed the influence of fragment chemical properties and protein family on the ability to predict accurate binding modes and highlighted up-to-date limitations and future directions.