Assessing the potential of deep learning for protein-ligand docking.

Area of Science:

• Computational Biology
• Biochemistry
• Drug Discovery

Background:

• Ligand binding affects protein structure and function, crucial for biomedical research and drug discovery.
• Existing deep learning (DL) docking benchmarks lack systematic evaluation for real-world scenarios like using predicted protein structures or handling multiple ligands.

Purpose of the Study:

• Introduce PoseBench, a comprehensive benchmark for evaluating protein-ligand docking methods.
• Assess DL methods in broadly applicable contexts: using predicted apo protein structures, multi-ligand binding, and unknown binding pockets.

Main Methods:

• Developed PoseBench, a benchmark dataset for protein-ligand docking and structure prediction.
• Included primary ligand and novel multi-ligand datasets for rigorous evaluation.
• Systematically evaluated various DL and conventional docking algorithms.

Main Results:

• DL cofolding methods generally outperform baseline algorithms but struggle with novel protein-ligand binding poses.
• Sensitivity to input multiple sequence alignments varies among DL cofolding methods.
• DL methods face challenges balancing structural accuracy and chemical specificity for new or multi-ligand targets.

Conclusions:

• PoseBench provides a robust platform for advancing protein-ligand docking research.
• Current DL methods require further development for broader real-world applicability in drug discovery and enzyme design.
• Future research should focus on improving DL model generalization and accuracy for complex binding scenarios.