CardiOT: Towards Interpretable Drug Cardiotoxicity Prediction Using Optimal Transport and Kolmogorov--Arnold Networks

Area of Science:

• Computational chemistry
• Drug discovery
• Cardiovascular pharmacology

Background:

• Assessing cardiac drug safety requires investigating compound effects on cardiac ion channels.
• Computational models for combined cardiotoxicity (CCT) prediction face challenges due to experimental data scarcity and uneven distribution.
• Existing graph neural network (GNN) models overlook chemical bond information, limiting structural recognition.

Purpose of the Study:

• To develop an advanced GNN-based model, CardiOT, for accurate CCT prediction.
• To address data limitations and enhance the recognition of crucial chemical structures in cardiotoxicity assessment.
• To improve the interpretability of cardiotoxicity prediction models.

Main Methods:

• Integration of optimal transport (OT), structure remapping (SR), and Kolmogorov-Arnold networks (KANs) into a GNN framework.
• Utilizing OT pooling with expectation maximization to optimize sample-feature distributions and identify key pairs.
• Employing SR technology to emphasize chemical bond information during message propagation within the GNN.

Main Results:

• The CardiOT model effectively mitigates issues of uneven data distribution and scarcity.
• Demonstrated robust performance across multiple public datasets.
• Enhanced model interpretability through the integration of KAN technology.

Conclusions:

• CardiOT offers a significant advancement in predicting compound inhibition on cardiac ion channels.
• The model is expected to reduce toxicity risks in drug development.
• Provides deeper insights into the mechanisms of compound-cardiac ion channel interactions.