Machine learning prediction of 3CL^pro SARS-CoV-2 docking scores

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Molecular docking is crucial for predicting drug-target interactions.
• Machine learning (ML) offers a powerful approach to accelerate docking predictions.

Purpose of the Study:

• To evaluate the performance of different ML models for predicting molecular docking scores.
• To assess the utility of ML models in identifying compounds for drug repurposing.

Main Methods:

• Trained three ML models: Keras/TensorFlow neural networks, XGBoost, and SchNetPack neural networks.
• Utilized Smooth Overlap of Atomic Positions (SOAP) molecular descriptors.
• Validated models on diverse compound sets, including the ZINC in vivo set.

Main Results:

• Demonstrated consistent prediction capabilities across ML approaches.
• Observed slight overestimation for compounds exceeding 60 atoms, but still valid for screening.
• ML models proved effective for evaluating drug repurposing suitability.

Conclusions:

• Machine learning models reliably predict molecular docking scores.
• These models are valuable tools for accelerating drug discovery and repurposing efforts.