Graph-Based Classification with GNN-Explainer for Predicting Cardiac Toxicity Associated with Multi-Ion Channel...

Area of Science:

• Computational chemistry and toxicology
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Cardiotoxicity is a major cause of drug attrition, traditionally assessed via Kv11.1 channel inhibition.
• The Comprehensive in Vitro Proarrhythmic Assay (CiPA) highlights the need to evaluate Cav1.2 and Nav1.5 channels.
• Existing machine learning models often lack comprehensive ion channel data.

Purpose of the Study:

• To develop a deep learning framework integrating inhibition data for Kv11.1, Cav1.2, and Nav1.5 channels.
• To improve early-stage cardiotoxicity risk assessment in drug discovery.
• To provide interpretable insights into molecular contributions to cardiotoxicity.

Main Methods:

• Curated a large dataset (Cardio-Tox) from multiple repositories (PubChem, CUPID, CToxPred2).
• Developed and trained Graph Neural Network (GNN) models for individual channel inhibition prediction.
• Utilized GNNExplainer for interpretable visualization of atom- and bond-level contributions.

Main Results:

• The integrated CardiotoxPred method achieved 86.7% average prediction accuracy on a test dataset.
• GNN models demonstrated capability for individual channel prediction.
• Interpretable visualizations provided insights into molecular drivers of cardiotoxicity.

Conclusions:

• The developed deep learning framework offers robust and interpretable cardiotoxicity prediction.
• Freely accessible models in Docker containers facilitate early safety profiling.
• This tool aids in optimizing drug candidates and reducing late-stage failures.