Abstract Body (Do not enter title and authors here): Background: Drug-induced cardiotoxicity is a major cause of clinical trial failures and post-market drug withdrawals. Current screening methods rely primarily on cell assays and transcriptomic profiling, but metabolic perturbations may provide additional predictive signals for cardiotoxic liability in drugs not yet tested in humans.

Hypothesis: We hypothesize that predicted metabolic flux changes derived from gene expression data would outperform transcriptomic features for predicting drug cardiotoxicity in a machine learning framework.

Methods: We developed a novel computational pipeline integrating a modified iCardio genome-scale metabolic model with transcriptomic data from 5 genetically distinct hiPSC-derived cardiomyocyte cell lines treated with a variety of 31 antineoplastic and immunomodulating drugs (accessed from DToxS Center). Our mathematical framework converts gene expression changes to enzyme activity, then to relative metabolic reaction flux change (4122 reactions) using a novel constrained quadratic approach. Ensemble classifiers were trained to predict cardiotoxicity using FDA Adverse Event Reporting System Reporting Odds Ratio (ROR) as a reference, with drugs above median ROR classified as cardiotoxic. Predictive models were generated using 5-fold cross validation for hyperparameter optimization with 25% hold out for quality metric calculation (reported as mean ± SEM, P values calculated from t-test) over 100 independent iterations.

Results: The metabolic flux-based classifiers demonstrated fair predictive performance of drug cardiotoxicity with an AUROC of 0.70±0.02. The flux approach produced equivalent accuracy (+0.02, P = 0.56), and significantly higher F1 (+0.10, P = 0.01), AUROC (+0.08, P = 0.01), and AUPRC (+0.07, P = 0.03) than the gene expression approach. Further analysis revealed that perturbations to fatty acid metabolism were most predictive of cardiotoxic liability, with 42 out of top 100 predictive reactions belonging to fatty acid related subsystems (P < 1e-5 by binomial test).

Conclusions: Metabolic flux prediction from transcriptomic data provides enhanced discrimination of drug cardiotoxicity compared to gene expression analysis alone. This approach enables more informative pre-clinical screening of drug candidates before human exposure, potentially reducing late-stage clinical failures and improving drug safety assessment protocols.