AI to Democratize Access to Cheap, High-Throughput Cardiotoxicity Testing

By Deborah Borfitz

March 4, 2025 | Researchers have trained an artificial intelligence (AI) model to reconstruct the electrical signals inside of heart cells based on recordings taken from outside those cells, opening the possibility of cheap, high-yielding cardiotoxicity testing using commercially available microelectrode arrays. The hope is that the U.S. Food and Drug Administration (FDA) will allow the tool to be used in lieu of inserting pipettes into cells to record signaling changes with the addition of a drug, which is both costly and cumbersome to do at scale, says Zeinab Jahed, Ph.D., assistant professor of chemical and nano engineering at the University of California, San Diego. 

Demand is unquestionably high among drug developers for a more practical, noninvasive method to read the heart’s electrical signals and predict at an early stage, with fewer resources, if a drug is going to be cardiotoxic. Many drugs cost upwards of $1 billion to develop and cardiotoxicity is one of the main reasons they’re taken off the market, she says.  

An article about the novel AI-based model, trained on nanoelectrode array recordings, was published recently in Nature Communications (DOI: 10.1038/s41467-024-55571-6). Although currently used on 2D cell cultures, the method is also applicable to organoid systems, though further studies are needed for validation says Jahed. 

Among the research team’s next steps is to test the model’s ability to map the response of cells in a dish to an actual human heart, she adds. Organoids, like animal models, can be expensive and hard to maintain for drug screening purposes and may not faithfully recapitulate whole-organ response. 

The idea here is to improve the prediction of human outcomes when taking an experimental drug by screening the substances directly on patient-derived heart cells. “Many researchers are working on generating large numbers of heart cells from induced pluripotent stem cells (iPSCs) derived from specific patients, but the tests you can do on them are very low throughput,” says Jahed. The current means of recording signals involves peering through a microscope to find and puncture cells one at a time. 

Multiple companies are now in the business of reprogramming human cells by manipulating gene expression and epigenetic factors to convert one cell type into another, including cardiomyocytes, which are being used by Jahed and her colleagues. They are expanding their work on the model to commercial brain cell lines—and not just for screening drugs for treating neurological disorders, but for better understanding the function of the brain, Jahed says. The neuron-focused project is funded by the Air Force Office of Scientific Research. 

There is so much more to learn with neurons because, unlike heart cells, they have more than one job, she notes. Different types of brain cells have different types of electrical signals that are not easy to detect from outside the cell and each has different implications for health and disease. 

If the AI model can decipher the signals from neurons, the applications would be seemingly endless, says Jahed. Currently, electrodes can be implanted in the brain to stimulate and sense signals from neurons, but “they’re only able to do extracellular recording to say the neuron fired,” not what kind of signal was sent or received. 

AI-Based Signal Translation

Work on the intelligent in-cell electrophysiology method goes back five or six years to when Jahed, as a postdoctoral research fellow at Stanford University, helped engineer an array of nanoscale, needle-shaped electrodes onto which many heart muscle cells could be placed to record hundreds of electrical signals all at once. The technology could pick up both signals from inside the cell, as well as “noisy tiny signal” from outside cells, somewhat like hearing a conversation from behind a closed door without understanding every word.  

For the latest study, this nanoelectrode array recorded signals coming from both inside and outside cells and machine learning was trained to find the strong correlations between extracellular and intracellular features (e.g., amplitude and spiking velocity) to understand exactly what’s going on inside the cells, she explains. 

The researchers collected thousands of pairs of electrical signals, each linking an extracellular recording with its corresponding intracellular signal. The data included how the cells responded when exposed to various drugs.  

Simple analysis suggested correlations existed between extracellular and intracellular recordings, which motivated the team to apply deep learning to the signal translating exercise along with some “physics-informed constraints,” says Keivan Rahmani, a nano engineering Ph.D. student in Jahed’s lab. Computational models attempting to relate the two domains can be found in the literature, but they all require many parameters—in some cases, up to 100.    

The idea arose that perhaps all that information is already encoded inside the extracellular signal, based on its shape, and could be deciphered with the help of AI, he says. High-throughput electrophysiology assays using automated patch clamp technology can provide intracellular recordings, but the setup cost can approach half a million dollars. Microelectrode assays are more affordable at around $500, making them a fixture in almost every lab and biotech company doing drug toxicity screening, but they’re only able to make extracellular recordings. 

But, as the published study has demonstrated, the new AI model could be integrated with any of the microelectrode assays to do the signal translation work with extraordinary accuracy, says Rahmani. This has the potential to democratize access to high-throughput, intracellular-level electrophysiology. 

Universal Phenomenon 

The researchers have yet to test their model with all the different extracellular recording systems on the market, but the code and data supporting findings of the study are all available online, says Jahed, “so we are hoping to collaborate and start testing this on other people’s data.” The FDA Modernization Act also encourages the use of non-animal models and AI for any type of drug development work, which could mean a regulatory nod for the new cardiotoxicity testing approach. 

Since the system focuses on electrical activity, it is most relevant to tissues that show electrical activity and might be impacted by drugs, Jahed says. The focus currently is on heart and neural tissue, both of which have significant electrical activity. 

It is becoming increasingly clear that electrical signaling is a “universal process across many biological systems. Beyond heart and neural tissue, “it has been discovered in bacteria coordinating biofilms, plants responding to environmental cues, fungi transmitting impulses, and even in the gut, heart, immune cells, and cancer cells,” she adds. These findings open new possibilities for developing tools to study electrical communication across diverse tissues. 

“More complex solutions are not necessarily better,” Jahed points out. Even wearable sensors that are less complex and not as sensitive could detect signals with much higher resolution and accuracy by using AI to make sense of what the signals are saying. 

When faced with a lot of “distorted data” that includes the information being sought, “AI is definitely the way to go,” says Rahmani.