SchNetPack Hyperparameter Optimization for a More Reliable Top Docking Scores Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Predicting top docking scores is crucial for identifying potent drug candidates.
• Machine learning models, like those using SchNetPack, often struggle with extrapolating to rare, high-scoring compounds.

Purpose of the Study:

• To enhance the extrapolation power of SchNetPack neural networks for predicting top docking scores.
• To identify key hyperparameters that improve the prediction accuracy of high-scoring compounds.

Main Methods:

• Hyperparameter tuning of the SchNetPack atomistic model representation.
• Evaluation of prediction robustness using mean square error (MSE) and loss landscape entropy.
• Comparison of cutoff distance, radial basis functions, network layers, and feature vector size.
• Analysis of oversampling and undersampling techniques for training data.

Main Results:

• Optimizing the cutoff hyperparameter (at 5 Å) significantly improves the prediction of top docking scores (MSE reduced from ~3.5 to 0.9 kcal/mol).
• Other hyperparameters showed minimal impact on predicting top-scoring compounds compared to the cutoff.
• The cutoff optimization improved accuracy for top scores more than data sampling methods (undersampling > oversampling).
• Overall prediction accuracy slightly decreased when focusing on top-scoring compounds.

Conclusions:

• The cutoff distance is the most critical hyperparameter for improving SchNetPack's prediction of top docking scores.
• Hyperparameter tuning offers a more effective strategy than data sampling for enhancing the prediction of rare, high-value compounds in machine learning models.
• Careful tuning can improve the identification of promising drug candidates through more accurate docking score predictions.