Assessing the potential of deep learning for protein-ligand docking

Area of Science:

• Computational Biology
• Structural Biology
• Biotechnology

Background:

• Ligand binding influences protein structure and function, crucial for drug discovery and biotechnology.
• Existing deep learning (DL) docking methods lack systematic evaluation for broad applicability, including using predicted protein structures, multi-ligand binding, and unknown pockets.

Purpose of the Study:

• Introduce PoseBench, the first comprehensive benchmark for broadly applicable protein-ligand docking.
• Enable rigorous evaluation of DL methods for apo-to-holo docking and protein-ligand structure prediction.
• Facilitate systematic assessment using primary and novel multi-ligand datasets.

Main Methods:

• Developed PoseBench, a benchmark for evaluating protein-ligand docking and structure prediction.
• Included datasets for single and multiple ligand binding scenarios.
• Assessed DL co-folding methods against conventional and DL docking baselines.

Main Results:

• DL co-folding methods generally outperform baseline docking methods.
• Popular DL methods like AlphaFold 3 face challenges with novel protein sequences.
• DL methods exhibit sensitivity to multiple sequence alignments and struggle with balancing accuracy and specificity in multi-ligand predictions.

Conclusions:

• PoseBench provides a robust framework for evaluating protein-ligand docking methods.
• Current DL methods require further development for real-world applications, especially for novel targets and multi-ligand systems.
• The benchmark facilitates advancements in drug discovery and protein design through improved computational tools.