Coloring Molecules with Explainable Artificial Intelligence for Preclinical Relevance Assessment.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Graph neural networks (GNNs) show promise in drug discovery tasks like molecular property prediction.
• GNNs are often 'black-box' models, limiting interpretability and debugging for rational molecular design.

Purpose of the Study:

• To improve the transparency of GNN models in drug discovery.
• To apply the integrated gradients explainable artificial intelligence (XAI) approach to GNNs for rational molecular design.

Main Methods:

• Trained GNN models for predicting key drug properties: plasma protein binding, hERG channel inhibition, passive permeability, and cytochrome P450 inhibition.
• Employed the integrated gradients XAI technique to analyze and interpret the GNN models.

Main Results:

• The XAI methodology successfully highlighted critical molecular features and structural elements consistent with known pharmacophore motifs.
• Identified property cliffs and provided insights into unspecific ligand-target interactions, enhancing understanding of model predictions.
• Demonstrated the utility of XAI for rational molecular design and debugging of GNNs.

Conclusions:

• The integrated gradients XAI approach enhances the interpretability of GNNs in drug discovery.
• This open-sourced methodology facilitates rational molecular design and can be applied to other clinically relevant endpoints.
• Improves trust and debugging capabilities for AI models in pharmaceutical research.