Teaching AI to "Know" All About Drugs, Not Just Their Shape: A New Way to Predict Drug Toxicity

For decades, computer programs designed to predict whether a drug is safe have relied on one main idea: a drug's chemical shape determines how it behaves in the body. These "shape-based" models have been useful, but there's a catch. Sometimes two drugs that look almost identical can have very different safety profiles. A well-known example is ibuprofen and a related drug called ibufenac. They differ by just one small chemical piece, yet ibuprofen is a safe, common pain reliever while ibufenac had to be pulled from the market for causing serious liver damage. This is a big problem for predicting two of the most common reasons drugs fail or get pulled from the market: liver damage and heart-related side effects.

To solve this issue, a research team began to develop a computer model that could also "know" everything else we understand about a drug beyond its shape, including how it works in the body, how it's broken down, and what side effects it's caused in the past. The researchers used artificial intelligence to write detailed summaries of each drug's known effects and safety history, then converted those summaries into a format a computer model could learn from. They tested this approach on liver injury and heart toxicity, training their models on older drugs and testing them on newer ones to mimic real-world use.

This knowledge-based approach consistently outperformed the traditional shape-based approach. The difference was especially clear when researchers tested both models on eight pairs of similar-looking drugs with very different liver safety records. The knowledge-based model correctly identified the more dangerous drug in all eight pairs, while the shape-based model only got five right.

This work directly supports the FDA's published goal of reducing animal testing in drug safety studies. Because this approach relies on existing knowledge rather than new physical experiments, it offers a promising, animal-free way to catch toxic drugs earlier, before they ever need to be tested in animals or people.

By teaching computer models to use everything we already know about a drug, not just its chemical shape, researchers built a more accurate, animal-free tool for catching dangerous drug side effects, even in cases where similar-looking drugs turn out to have very different safety records. This study is an example of how generative AI and large language models are being integrated into new approach methods (NAMs) for toxicology.  The Physicians Committee hosts webinars and training courses in nonanimal toxicology test methods like this through our NAM Use for Regulatory Application program.