PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Deep learning methods show promise for protein-ligand docking speed and accuracy.
• Current evaluations often rely solely on root-mean-square deviation (RMSD), potentially overlooking structural plausibility.
• Physically implausible molecular structures can hinder the reliability of docking predictions.

Purpose of the Study:

• To introduce PoseBusters, a Python package for assessing the physical plausibility of protein-ligand docking predictions.
• To evaluate the performance of deep learning-based docking methods against classical tools using rigorous quality checks.
• To highlight the importance of steric and energetic criteria in addition to RMSD for method evaluation.

Main Methods:

• Developed PoseBusters, a Python package utilizing RDKit for chemical and geometric validation.
• Implemented a test suite for stereochemistry, planarity, bond lengths, and protein-ligand clashes.
• Compared five deep learning methods (DeepDock, DiffDock, EquiBind, TankBind, Uni-Mol) and two classical methods (AutoDock Vina, CCDC Gold).
• Assessed methods with and without post-prediction energy minimization using molecular mechanics force fields.

Main Results:

• Deep learning methods frequently produced physically implausible structures, failing PoseBusters checks.
• No deep learning method outperformed classical docking tools in physical plausibility or generalization.
• Classical docking tools, especially with energy minimization, showed superior performance.
• Molecular mechanics force fields capture essential physics absent in current deep learning models.

Conclusions:

• Evaluating protein-ligand docking methods requires assessing physical plausibility (steric, energetic) alongside RMSD.
• PoseBusters provides a crucial tool for validating docking and molecular generation techniques.
• Classical docking methods remain state-of-the-art for reliable predictions.
• Further development of deep learning methods needs to incorporate physical realism and inductive biases.