Abstract Body (Do not enter title and authors here): Background
Drug-induced cardiotoxicity (DICT) is a severe adverse drug reaction that contributes to clinical trial failures and drug withdrawals. Forecasting DICT is challenging due to the property’s complex nature, and experimental assays are cumbersome and poorly correlate with human outcomes. Machine learning trained on clinical DICT data can quickly and accurately predict cardiotoxicity, pre-empting clinical failure and providing insight.

Methods
We used our previously developed ADMET-AI machine learning platform to predict DICT and its sources. ADMET-AI is a freely available web tool (admet.ai.greenstonebio.com) that predicts 41 absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of a given molecule. We coupled ADMET-AI’s 41 ADMET property predictions with an Extreme Gradient Boosting (XGB) model to output a probability of cardiotoxicity. This XGB model was trained on 555 drugs from the published DICTrank dataset that have been labeled as having either no cardiotoxicity (262 drugs) or severe cardiotoxicity (293 drugs). Notably, over 67% of the DICTrank drugs are not in any of the 41 ADMET-AI training datasets.

Results
The ADMET-AI-based cardiotoxicity XGB model, evaluated with 10-fold cross-validation, achieved an average area under the precision-recall curve (PR-AUC) of 0.75. Our model outperformed an XGB model trained on SwissADME predictions (PR-AUC=0.69). Shapley values identified CYP2D6 Substrate and Nrf2-Antioxidant Responsive Element as key predictors of severe cardiotoxicity, while CYP2D6 inhibition and Aromatase were linked to non-cardiotoxicity. t-SNE analysis of the DICTrank chemical space using the five most predictive ADMET-AI properties showed clear clustering of toxicity. Case studies of cisapride and lactulose illustrated the model’s ability to distinguish between cardiotoxic and non-cardiotoxic drugs based on predictions. Compared to publicly available models, ADMET-AI is currently the fastest and most accurate tool for ADMET and DICT prediction.

Conclusions
ADMET-AI effectively predicts DICT with insight into potential causes, surpassing the current standard of public tools. The ADMET-AI-based cardiotoxicity prediction models are freely available to enable DICT prediction, with the potential to prevent late-phase clinical failures. While this study focuses on DICT, a similar workflow can adapt ADMET-AI to make predictions for other key properties in drug discovery.