Improving In Silico Cardiac Safety Prediction by Consensus Averaging of Transmural Ventricular Cell Models

Ann Biomed Eng. 2026 Mar 3. doi: 10.1007/s10439-026-04028-6. Online ahead of print.

ABSTRACT

PURPOSE: Established in silico frameworks for assessing Torsade de Pointes (TdP) risk primarily rely on single-cell electrophysiological biomarkers, which have demonstrated strong predictive capabilities. However, ventricular transmural electrophysiological heterogeneity is known to influence repolarization dynamics and arrhythmogenic mechanisms. Explicitly incorporating endocardium, epicardium, and mid-myocardium representations may enhance physiological interpretability, but direct integration of multi-cell features can introduce severe multicollinearity and compromise model stability. To address this challenge, we propose an ordinal logistic regression (OLR) framework that integrates multi-cell qNet information through probability averaging, preserving physiological context while maintaining robust statistical behavior.

METHODS: $I{C}_{50}$ values and Hill coefficients for 28 CiPA drugs were implemented in two ventricular cell models, the CiPAORdV1.0 and the ORd in silico models. qNet was computed independently for endocardium, epicardium, and mid-myocardium cells. Cell-specific OLR models produced class probabilities that were then averaged to generate the final prediction. Performance was compared against single-cell and direct multi-cell implementations across Manual and ChanTest datasets.

RESULTS: For CiPA-ORd v1.0 using the ChanTest dataset, qNet achieved substantial performance, with AUCs for ROC1 and ROC2 of 1.000 and 0.958, respectively, and also meeting seven “excellent” classification criteria. In the ORd model, probability averaging consistently improved performance for both the Manual and ChanTest datasets relative to single-cell and direct multi-cell approaches.

CONCLUSION: Probability-averaged integration of multi-cell qNet predictions mitigates multicollinearity while preserving physiological relevance, yielding more stable and accurate in silico TdP risk classification and supporting broader applicability to preclinical safety assessment.

PMID:41774378 | DOI:10.1007/s10439-026-04028-6