Predicting cardiotoxicity in drug development: A deep learning approach.

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Cardiotoxicity is a significant risk in drug development, often linked to the hERG potassium channel.
• Traditional cardiotoxicity testing is costly and time-consuming.
• Computational virtual screening offers a more efficient alternative.

Purpose of the Study:

• To develop accurate and efficient computational models for predicting compound cardiotoxicity.
• To improve drug safety assessment through machine learning and deep learning techniques.

Main Methods:

• Utilized molecular fingerprints and descriptors with machine learning (Gaussian NB, RF, SVM, KNN, XGBoost) and deep learning (Transformer) algorithms.
• Evaluated model performance using accuracy (ACC) and area under the curve (AUC).
• Employed SHapley Additive exPlanations (SHAP) for feature interpretability.

Main Results:

• The best machine learning model (XGBoost Morgan) achieved an ACC of 0.84.
• The best deep learning model (Transformer_Morgan) achieved an ACC of 0.85.
• The Transformer_Morgan model achieved an AUC of 0.93 on an independent validation set, outperforming existing tools.
• SHAP analysis identified key chemical features associated with cardiotoxicity, such as benzene rings and fluorine-containing groups.

Conclusions:

• Machine learning and deep learning models provide highly accurate predictions for cardiotoxicity.
• These computational approaches offer a reliable and interpretable method for drug safety evaluation.
• The study facilitates efficient drug development, reduces costs, and enhances the safety of new drug candidates.