Speaker
Description
Torsades de Pointes (TdP) is a life-threatening ventricular arrhythmia that originates when a drug compound inhibits cardiac ion channels at the nanoscale, altering the ionic currents underlying the cardiac action potential. Translating this nanoscale inhibition into a clinically actionable risk category remains central to the Comprehensive in vitro Proarrhythmia Assay (CiPA) paradigm, yet the three TdP risk classes (low, intermediate, high) are inherently ordinal, a structure conventional classifiers typically disregard. Here, in silico simulations were performed using the optimized IKr-dynamic O'Hara-Rudy (CiPAORdv1) ventricular cell model for the 28 CiPA reference drugs at four concentrations (1-4× Cmax), parameterized using manual patch-clamp ion channel pharmacology data. Eleven biomarkers (qNet, dvdtmax, vmax, vrest, APD50, APD90, max_dv, camax, carest, CaTD50, CaTD90) were extracted from the resulting action potentials and calcium transients and used to train a COnsistent RAnk Logits (CORAL) ordinal artificial neural network with a single hidden layer. Because CORAL's rank-monotonic logits directly encode the ordinal relationship among risk classes, a Torsade Metric Score (TMS) was derived from these logits indicating higher scores intuitively correspond to higher risk. The resulting ordinal deep learning model achieved classification performance comparable to the established qNet metric, while additionally producing an interpretable, continuous TMS rather than a discrete label alone. These findings indicate that ordinal deep learning can serve as a computational bridge between nanoscale drug-ion channel inhibition and clinically meaningful, intuitively interpretable TdP risk stratification, offering a methodological foundation for extending nanobiotechnology-informed cardiac drug safety assessment beyond conventional, non-ordinal classification approaches.
