AI Model To Predict Drug-Induced Cardiotoxicity
by Micaela Harris
January 22, 2025
Drug-induced cardiotoxicity (DICT) is a leading cause of drug withdrawals and clinical trial failures. There are many reasons why cardiotoxicity may occur, and it is a complex process to predict the probability or mechanism of occurrence. DICT can result in arrhythmias and cardiomyopathies, which may lead to heart failure and even sudden cardiac death. Although there are methods to assess drug-induced cardiotoxicity, such as using cardiac cells and animal models, these methods are ultimately slow and expensive and often do not correlate well with cardiotoxicity in humans.
With the rise in machine learning (ML) capabilities, models can be trained to use real-world clinical drug-induced cardiotoxicity data to help predict particular properties of treatment failure. In a recent study published in Circulation, scientists at Stanford University and Greenstone Bio Inc., led by Souhrid Mukherjee, PhD; Kyle Swanson, MSc, MEng; and Joseph C. Wu, MD, PhD, developed an ML platform titled “ADMET-AI” to help predict and interpret drug-induced cardiotoxicity. ADMET-AI is currently the fastest and most accurate publicly available web server to predict the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of therapeutic drugs. ADMET-AI uses graph neural network models to create accurate predictions and to provide human-understandable interpretations of sources of drug-induced cardiotoxicity.
The workflow for building a drug-induced cardiotoxicity (DICT) prediction model using ADMET-AI. for training, testing, validation, and interpretation of ADMET-AI for DICT, based on the FDA DICTRank data set.
The ADMET-AI algorithm predicts 41 ADMET properties of a given molecule and uses another computational model to output the likelihood of DICT. The computational model was trained using 555 drugs (262 non-toxic, 293 most cardiotoxic) from a drug-induced cardiotoxicity dataset published by the FDA (DICTRank) that includes many different and important drug classes in cardiovascular disease therapies, such as anti-arrhythmias, calcium channel blockers, and more. The algorithm ultimately categorizes drugs as having a higher or lower concern for cardiotoxicity.
The study explores the potential of using the ADMET-AI algorithm to predict DICT, with the goal of helping to prevent drug-toxicity-related failures during clinical development. This AI model, including its data and materials, are publicly available to support clinical therapies and reduce unsuccessful treatments.
Additional Stanford Cardiovascular Institute investigators include Parker Walther, BA, Rabindra Shivnaraine PhD, Jeremy Leitz, PhD, Paul Pang, PhD, and James Zou, PhD. This work was funded by the Knight-Hennessy Scholarship, the National Institutes of Health, the California Institute of Regenerative Medicine, and the Department of Defense.
Souhrid Mukherjee, PhD
Kyle Swanson, MSc, MEng