Pharmacophore an International Research Journal
Cardiac toxicity remains a major safety challenge in drug discovery, as pro-arrhythmic risk can emerge from complex interactions between chemical structure, hERG inhibition, and broader ion-channel pharmacology, with certain molecular substructures creating toxicophoric patterns that are not always captured by single-assay interpretation. While many computational cardiotoxicity models provide useful risk predictions, their limited interpretability reduces practical value for toxicologists and medicinal chemists, who need to understand why a compound is flagged and how it could be redesigned. To address this, an explainable neural network for cardiotoxicity prediction can integrate hERG information, multi-ion-channel profiles, and molecular substructure features, offering atom-level, substructure-level, and channel-level explanations for each prediction. Using a neural architecture such as graph attention or message passing to encode molecular graphs while incorporating ion-channel inhibition data, and applying post-hoc explanation methods like SHAP or integrated gradients, the model can decompose predictions into structural alerts and channel-specific contributions. Conceptually, it could identify a compound as pro-arrhythmic and clarify that the predicted risk is driven by hERG inhibition, sodium-channel activity, and highlighted substructures consistent with known cardiotoxic motifs. By linking predictive modeling to interpretable molecular and electrophysiological drivers, such an approach could enhance transparency, support safer drug design, and facilitate more informed decision-making in cardiac safety assessment.
