A Deep-Learning-Based Scoring Framework for Large-Scale Multi-donor Cardiotoxicity Screening

Abstract

Cardiotoxicity remains a major cause of drug attrition and postmarket withdrawal, yet the vast majority of environmental chemicals to which humans may be exposed remain uncharacterized for cardiotoxicity risk. Human induced pluripotent stem cell (hiPSC)-based testing has been proposed to address this gap. Here, we present an unsupervised deep learning framework for multi-donor cardiotoxicity screening using high-throughput calcium transient recordings from hiPSC-derived cardiomyocytes (hiPSC-CMs). We analyzed data from a library of 1,029 compounds tested in hiPSC-CMs from 5 donors across a concentration range. An autoencoder trained exclusively on baseline signals quantified chemically induced functional perturbations through reconstruction error, bypassing the need for labeled training data while capturing the full spectrum of calcium-handling disruptions. Aggregation of donor-specific scores revealed substantial inter-individual variability in potential cardiotoxicity, underscoring the value of this approach for multi-donor risk prediction. We identified microbiocides, dyes, and pesticides as chemical classes of potential concern, characterized by high toxicity scores and low interdonor variability. This framework establishes a scalable, human-relevant, and genetically diverse platform for cardiotoxicity surveillance across both pharmacological and environmental chemical spaces.