AtomNet PoseRanker: Enriching Ligand Pose Quality for Dynamic Proteins in Virtual High-Throughput Screens

Area of Science:

• Computational chemistry and structural biology
• Artificial intelligence in drug discovery

Background:

• Virtual High-Throughput Screening (vHTS) is crucial for identifying drug candidates against novel targets.
• Protein dynamics and conformational flexibility challenge traditional docking methods, impacting pose accuracy.
• Accurate prediction of ligand binding poses and affinities is essential for effective vHTS.

Purpose of the Study:

• To introduce AtomNet PoseRanker (ANPR), a deep learning model for reranking ligand poses.
• To address the limitations of docking to single protein conformations by accounting for flexibility.
• To improve the accuracy of vHTS by enhancing pose quality and hit rates.

Main Methods:

• Developed ANPR, a graph convolutional network, to identify and rerank crystal-like ligand poses.
• Trained ANPR on ensembles of protein conformations and ligand poses.
• Evaluated ANPR's performance on the PDBbind v2019 dataset, including cognate and noncognate interactions.

Main Results:

• ANPR significantly improved pose quality in docking experiments.
• The model effectively reranked poses, better representing experimentally observed binding modes.
• ANPR demonstrated the ability to infer and account for receptor flexibility even with single conformations.

Conclusions:

• ANPR enhances computational drug discovery by improving ligand pose prediction accuracy.
• Improved pose rankings lead to better hit rates in vHTS campaigns.
• ANPR is particularly valuable for discovering drugs against novel therapeutic targets or binding sites.