Convolutional neural network scoring and minimization in the D3R 2017 community challenge

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in pharmacology

Background:

• Accurate prediction of ligand-receptor interactions is crucial for drug discovery.
• Conventional scoring functions in molecular docking have limitations in pose prediction and virtual screening.
• Deep learning approaches offer potential for improving scoring function performance.

Purpose of the Study:

• To evaluate the efficacy of a convolutional neural network (CNN)-based scoring function for drug discovery tasks.
• To compare the CNN's performance against conventional scoring functions like Autodock Vina.
• To assess the CNN's utility in pose refinement and virtual screening.

Main Methods:

• Development and application of a CNN-based scoring function.
• Re-scoring and refinement of ligand poses generated by Autodock Vina.
• Benchmarking against the D3R 2017 community challenge, including the Cathepsin S target.
• Evaluation of different strategies for selecting reference receptors.

Main Results:

• The CNN scoring function outperformed Autodock Vina in most evaluated tasks.
• The CNN achieved best-in-class performance in several virtual screening scenarios.
• The Cathepsin S target proved challenging for de novo docking.
• CNN performance was robust without requiring manual operator intervention.

Conclusions:

• CNN-based scoring functions demonstrate significant potential to advance drug discovery.
• Deep learning offers a powerful tool for improving molecular docking and virtual screening accuracy.
• The developed CNN shows promise for enhancing the efficiency of identifying potential drug candidates.