Efficient and accurate large library ligand docking with KarmaDock.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in pharmacology

Background:

• Structure-based virtual screening relies heavily on ligand docking, a crucial step in identifying potential drug candidates.
• Existing docking tools face limitations in speed, pose quality, and binding affinity prediction accuracy.
• Deep learning approaches show promise but require further development for comprehensive docking solutions.

Purpose of the Study:

• To introduce KarmaDock, a deep learning framework designed to overcome the limitations of conventional and current deep learning docking methods.
• To integrate docking acceleration, precise binding pose generation/correction, and accurate binding strength estimation into a single model.
• To enhance the efficiency and reliability of virtual screening in drug discovery.

Main Methods:

• Development of a three-stage deep learning model, KarmaDock.
• Stage 1: Protein and ligand encoders to learn intramolecular interaction representations.
• Stage 2: E(n) equivariant graph neural networks with self-attention for ligand pose updates and post-processing for chemical plausibility.
• Stage 3: A mixture density network for binding strength scoring.

Main Results:

• KarmaDock demonstrated robust performance across four benchmark datasets.
• The model successfully accelerated docking, improved pose generation, and accurately estimated binding affinities.
• In a real-world screening, KarmaDock identified validated active inhibitors for leukocyte tyrosine kinase (LTK).

Conclusions:

• KarmaDock offers a significant advancement in ligand docking technology for structure-based drug discovery.
• The integrated approach enhances speed, accuracy, and reliability in virtual screening processes.
• KarmaDock's success in identifying LTK inhibitors highlights its potential for discovering novel therapeutics.