A multiscale computational model of cardiac electrophysiology for drug-induced pro-arrhythmic risk stratification

The Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative positions in silico simulations as essential tools for cardiac safety assessment. While single-cell simulations reveal ionic perturbations, they under-represent tissue-scale conduction, electrotonic coupling, and spatial heterogeneity that shape organ-level arrhythmogenesis. Investigate whether a multiscale classifier that combines a single-cell biomarker (qNet) with an organ-level metric (simulated QT) improves Torsades de Pointes (TdP) risk stratification over either biomarker alone. Twenty-eight CiPA drugs were simulated at 1-4×Cmax. We derived Avg. qNet from single-cell simulations (2,000 IC50-h samples × 4 concentrations) and Avg. QT from 3D tissue simulations (median parameters). Ordinal Logistic Regression (OLR) models were evaluated under split-sample (12/16) and full-set (28) analyses. Avg. qNet outperformed Avg. QT. Adding Avg. QT to Avg. qNet provided no material gain across AUC, ordinal calibration, likelihood ratios (LR±), and error rates, with only a small improvement for identifying high-risk drugs in the full-set analysis. Within this framework and dataset, ECG-derived QT is insufficient as a standalone predictor of tissue-level arrhythmogenicity; Avg. qNet is a robust primary biomarker, and the multiscale (Avg. qNet + Avg. QT) model offers at most incremental benefit. Multiscale gains will likely require ECG features that capture conduction/dispersion and larger, more diverse cohorts.

The online version contains supplementary material available at 10.1007/s13534-025-00510-7.