Simplified, interpretable graph convolutional neural networks for small molecule activity prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Predicting small molecule activity is crucial for drug discovery.
• Graph convolutional neural networks (gCNNs) show promise but can be complex.
• Standard gCNNs and Random Forest (RF) methods have limitations in performance and interpretability.

Purpose of the Study:

• To develop a streamlined and explainable gCNN architecture for small molecule activity prediction.
• To optimize gCNN hyperparameters for improved performance across diverse protein targets.
• To enhance model interpretability by visualizing relevant molecular substructures.

Main Methods:

• Conducted hyperparameter optimization for a simplified gCNN Quantitative Structure-Activity Relationship (QSAR) architecture across nearly 800 protein targets.
• Augmented the simplified gCNN with saliency map technology to identify important molecular substructures.
• Performed saliency analysis on nearly 100 data-rich protein targets and analyzed substructure-activity relationships.

Main Results:

• The simplified gCNN architecture achieved performance improvements over standard gCNN and RF methods on challenging datasets.
• Reductions in convolutional layer dimensions were found to be beneficial, potentially relating to molecular substructure representation.
• Saliency analysis revealed useful substructural clusters for understanding activity relationships, with connections to medicinal chemistry literature.

Conclusions:

• A streamlined, explainable gCNN architecture offers enhanced performance and interpretability for small molecule activity prediction.
• The model's ability to highlight key molecular substructures aids in understanding structure-activity relationships.
• This approach provides valuable insights for lead finding and optimization in medicinal chemistry campaigns.