Pose Classification Using Three-Dimensional Atomic Structure-Based Neural Networks Applied to Ion Channel-Ligand...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate identification of lead compounds is crucial for early-stage drug discovery.
• Virtual high-throughput screening (vHTS) using molecular docking aids in identifying potential drug candidates.
• Suboptimal docking poses from standard methods lead to inaccurate screening and property calculations.

Purpose of the Study:

• To develop and evaluate machine learning models for classifying correct molecular docking poses.
• To improve the accuracy and efficiency of virtual high-throughput screening.
• To reduce false positives in drug discovery pipelines.

Main Methods:

• Utilized two convolutional neural network (CNN) approaches: a 3D-CNN and an attention-based point cloud network (PCN).
• Trained models on the PDBbind refined set.
• Evaluated classifiers on the CASF-2016 benchmark, ion channel datasets, and an in-house KCa3.1 inhibitor dataset.

Main Results:

• Proposed CNN classifiers effectively identify correct docking poses.
• Excluding false positive poses significantly improved virtual screening performance.
• Enhanced identification of novel molecules against target proteins compared to initial docking score-based screening.

Conclusions:

• Machine learning-based pose classification is a powerful tool to enhance virtual screening accuracy.
• The developed 3D-CNN and PCN models offer a robust solution for filtering suboptimal docking poses.
• This methodology can accelerate the identification of effective drug leads in early-stage drug discovery.