Ligand-Conditioned Side Chain Packing for Flexible Molecular Docking

Molecular docking is a crucial technique for elucidating protein-ligand interactions. Machine learning-based docking methods offer promising advantages over traditional approaches, with significant potential for further development. However, many current machine learning-based methods face challenges in ensuring the physical plausibility of generated docking poses. Additionally, accommodating protein flexibility remains difficult for existing methods, limiting their effectiveness in real-world scenarios. Herein, we present ApoDock, a modular docking paradigm that combines machine learning-driven conditional side chain packing based on the protein backbone and ligand information with traditional sampling methods to ensure physically realistic poses. The generated poses are finally scored by the developed mixture density network-based scoring function. With accurate side chain packing, physical-based pose sampling, and accurate pose ranking ability, ApoDock demonstrates competitive performance across diverse applications, especially when using modeled structure (AlphaFold2 and ESMFold) for docking, exhibiting a success rate of 28.5% higher than that of other state of the art (SOTA), highlighting its potential as a valuable tool for protein-ligand binding studies and related applications.