Cardiovascular disease is the leading cause of death worldwide. However, despite massive investment in drug discovery and development, the number of approved drugs has not kept pace. This is partially due to cardiovascular toxicity, as drug-induced cardiac side effects can derail promising candidates during clinical development or lead to serious adverse reactions after approval.
Although preclinical studies can help to identify compounds that disrupt cardiac function, some adverse effects only become apparent when drugs are administered to larger and more diverse patient populations. Differences in biology, drug exposure, and experimental measurements can all influence whether a compound triggers a dangerous cardiac response, making these risks difficult to predict before clinical use.
Computational models could help bridge this gap by modelling how these variables might influence cardiac safety. By simulating drug responses across different concentrations and biological conditions, researchers can explore scenarios that would be difficult, costly, or unethical to test directly in people. But how much should drug developers trust the predictions — particularly when the experimental data behind them are uncertain?
In a study published in Regulatory Toxicology and Pharmacology, researchers used AI-powered cardiac models to investigate that very question. Using loperamide, a widely available antidiarrheal medicine that can cause serious cardiac complications when taken in excessive amounts, they explored the impact of variable data, sex, and high drug exposures to assess proarrhythmic risks more accurately.
How virtual hearts account for uncertainty
Before a drug reaches clinical trials, researchers investigate how it affects the ion channels that control the heart’s electrical activity. These measurements help predict whether a compound could disrupt the heartbeat, but results often vary between experiments and laboratories.
“The quality of the data that you put in [affects] the quality of the results that come out,” Jazmin Aguado-Sierra, Lead Scientist at ELEM Biotech and co-author of the study, told DDN.
To investigate how this variability affects predictions, the researchers used AI-powered emulators trained on detailed computational models of cardiac electrophysiology. These emulators approximate the behavior of more complex simulations, enabling the team to rapidly explore a wide range of experimental measurements and drug exposure combinations.
The team found that variability in laboratory measurements widened the predicted range of QT intervals — the time it takes for the heart’s ventricles to electrically recharge between beats. This is crucial since prolonged QT intervals can indicate an increased risk of dangerous heart rhythms. Consequently, depending solely on average measurements may mask the inherent uncertainty regarding a drug's potential cardiac risk.
It can tell you, if your assays are varying this much, what you can expect is an output that could vary.
—Jazmin Aguado-Sierra, ELEM Biotech
Understanding how much that uncertainty affects the outcome could help researchers decide which data needs closer scrutiny before concluding a drug’s cardiac safety. “Our simulations cannot tell you exactly which one is correct,” Aguado-Sierra said. “It can tell you, if your assays are varying this much, what you can expect is an output that could vary.”
Testing the limits of cardiac safety
The team also used the emulators to investigate loperamide at extreme concentrations. Since researchers cannot ethically expose people to dangerously high doses, virtual hearts provide a way to explore these scenarios computationally.
The models predicted that female hearts could develop arrhythmias at substantially lower drug exposures than male hearts. Female cardiac models reached the predicted arrhythmia threshold at roughly half the drug exposure required for male models, suggesting a potentially important sex difference in cardiac susceptibility.
Although these thresholds are model predictions rather than clinical cut-offs, the findings illustrate how virtual populations could help researchers investigate differences in drug susceptibility that would be difficult to study systematically in people.
Building confidence in the predictions
Validating these models is difficult. Researchers can test whether an emulator reproduces the outputs of its underlying mechanistic simulator, but agreement between two computational models does not, by itself, establish that either accurately predicts human outcomes.
Independent clinical evidence can help assess that question, but the necessary data are not always available. Adverse drug reactions may be documented in individual case reports without detailed information about plasma drug concentrations, co-administered medications, or the characteristics that made a particular patient susceptible to the reaction.
“This is one of the biggest questions that we currently have,” said Aguado-Sierra, referring to how researchers can establish whether model predictions correspond to real-world outcomes.
She pointed to previous work modelling cardiac responses to hydroxychloroquine and azithromycin during the COVID-19 pandemic, in which the team found that predicted arrhythmia frequencies were similar to those observed clinically. However, she cautioned against expecting a single validation exercise to establish reliability across every drug and patient population. Cardiac responses depend on numerous interacting parameters, and collecting sufficiently detailed data for every possible scenario would be impractical.
Instead, researchers are working to build confidence incrementally by comparing model predictions with data from different compounds and studies. Aguado-Sierra said her team is collaborating with pharmaceutical companies to assess the models using in-house datasets, while wider community efforts are seeking to assemble additional compound data for benchmarking.
“The biggest main obstacle currently is [...] trust,” she said. “Which will only be built [...] with many papers [and] many benchmarks.”
From virtual predictions to drug-development decisions
In the near term, Aguado-Sierra sees preclinical safety assessment as a promising application. Researchers could use cardiac emulators to investigate whether experimental findings point towards a consistent prediction of human risk, or whether contradictory results warrant further investigation before advancing a drug candidate.
The longer-term challenge is to represent human biological diversity more comprehensively. The team has developed cardiac models incorporating different comorbidities and is exploring how to represent additional populations, including pregnant people and children. However, these efforts depend on obtaining sufficiently detailed data about how cardiac function differs between populations.
For now, Aguado-Sierra sees the development of reliable models as a collaborative undertaking, requiring input from researchers, pharmaceutical companies, and, eventually, regulatory bodies.
Virtual hearts could provide an effective way to explore drug responses that cannot be comprehensively tested in people. Their broader adoption will depend on continuing to demonstrate how predictions compare with experimental and clinical evidence, and on defining clearly the contexts in which those predictions can support decisions.
