Integrating multi-biomarker in silico ventricular cell models with ranking algorithms for enhanced prediction of drug-induced cardiac risk

Purpose Drug-induced Torsades de Pointes (TdP) remains a major challenge in cardiac safety assessment, while the predictive capability of single-biomarker approaches remains limited. This study evaluated the predictive performance of the ToR-ORd single-cell ventricular model using ordinal logistic regression (OLR) and investigated whether multi-biomarker integration improves TdP risk classification. Methods in silico simulations were performed for 28 CiPA reference drugs using the ToR-ORd ventricular action potential model. A total of 2000 in silico samples were generated for each drug by sampling Hill-model parameters. Eleven electrophysiological biomarkers were extracted from the simulations and systematically combined as inputs for OLR-based TdP risk classification. Model performance was evaluated over 10,000 repeated testing iterations. Results Among the single-biomarker models, five biomarkers (qNet, APD90, APD50, dv/dtmax_repol, and Vrest) satisfied the predefined CiPA performance thresholds; however, their performance scores remained lower than previously reported dynamic qNet-based performance. For the two-biomarker scheme, the APD50-APD90 combination achieved a performance score of 0.73. The highest overall performance was obtained using a four-biomarker combination of qNet, Vmax, APD90, and CaTD50, which achieved a performance score of 0.93. However, classification performance declined when more than six biomarkers were included, suggesting increased redundancy among biomarker inputs. Conclusion The ToR-ORd single-cell model combined with a structured multi-biomarker OLR framework provides a robust and interpretable approach for TdP risk classification. These findings demonstrate that carefully selected biomarker combinations can improve CiPA-aligned cardiac safety prediction compared with single-biomarker approaches.