Texas A&M develops artificial intelligence tools to assess chemical safety

Chemical toxicity prediction is advancing through new artificial intelligence research at Texas A&M University, where scientists are developing tools that can estimate both the potential hazards of chemicals and the reliability of those predictions.

Researchers at the Texas A&M College of Veterinary Medicine and Biomedical Sciences recently expanded on findings published in Nature Communications that examine how artificial intelligence can be used to predict chemical toxicity while also measuring uncertainty in those assessments.

The research is led by Dr. Weihsueh Chiu, a professor in the Department of Veterinary Physiology and Pharmacology. According to Texas A&M, the goal is to help address a longstanding challenge in toxicology: the large number of chemicals in commerce that lack comprehensive safety data.

Traditional chemical safety evaluations often rely on animal studies or long-term human epidemiological research, both of which require significant time and resources. As a result, many chemicals remain insufficiently studied.

To help close this gap, researchers have developed machine learning models known as quantitative structure-activity relationship models that use a chemical's structure to estimate safe exposure levels. Chiu's team has also focused on improving transparency by designing models that rely on familiar chemical properties, such as water solubility, biodegradability, and toxicity indicators, rather than relying solely on abstract molecular descriptors.

The latest advancement incorporates uncertainty-aware machine learning, enabling the models to estimate how confident they are in each prediction. The confidence level depends on the amount and quality of existing data available for similar chemicals.

According to Chiu, understanding uncertainty is critical because chemicals with similar predicted toxicity levels may present different levels of risk if one prediction is based on limited supporting data. The models generate a range of possible outcomes, allowing researchers to identify chemicals that may require additional study or expert review.

When applied to more than 126,000 chemicals, the models identified patterns in both toxicity and uncertainty. Researchers found that metals, polychlorinated compounds, and per- and polyfluoroalkyl substances, commonly known as PFAS, often showed higher uncertainty levels because of limited data availability or complex chemical behavior.

Texas A&M researchers believe these findings can help direct future testing efforts toward areas where scientific knowledge is limited. The approach supports a tiered evaluation process in which artificial intelligence is used for large-scale chemical screening, while experts focus on substances that present higher risks or greater uncertainty.

According to the researchers, continued advances in artificial intelligence could help shift chemical safety evaluation from a reactive process toward a more predictive approach, improving how scientists and regulators identify and assess potential hazards.