UF Health researchers have used artificial intelligence to develop models that can reliably predict specific organ systems at risk from common cancer-causing chemicals.
Researchers can now quickly screen chemicals using an online tool that flags high-risk chemicals. That could help scientists and drug developers prioritize which compounds need in-depth follow-up testing early in the drug discovery process. The findings were published recently in the journal Environmental Pollution.
“Instead of simply flagging a chemical as ‘cancer-causing in general,’ the models can reliably predict which specific organ systems are at risk,” said lead author Chi-Yun Chen, Ph.D., a postdoctoral associate in the Department of Environmental and Global Health in the UF College of Public Health and Health Professions. “A key finding is that different organs have different chemical warning signs. For example, respiratory and endocrine cancer risks are heavily driven by the presence of specific molecular fragments. In contrast, liver and urinary cancer risks are more closely tied to a chemical’s physical properties, such as its fat solubility or electrical charge.”
Traditional methods for identifying cancer-causing chemicals, known as carcinogens, rely on two-year animal studies, which are slow, expensive and can be ethically challenging. While some computer-based prediction tools have been developed to speed up this process, they generally only flag whether a chemical might cause cancer overall, not which organ it targets.
“That organ-level detail is critical for drug development, risk assessment and understanding how a chemical causes harm,” said Chen, who works in the lab of senior author and UF Health Cancer Institute member Zhoumeng Lin, Ph.D.
To build their models, the team curated a dataset of 945 chemical compounds with data on cancer risk in five major organ systems: endocrine, exocrine, hepatobiliary, respiratory and urinary.
Each chemical was encoded using methods that translate a molecule’s structure and chemistry into numerical data that a machine learning algorithm can interpret. Machine learning is a type of artificial intelligence. The team also incorporated biological activity data from the federal Tox21 program, which records how thousands of chemicals interact with biological pathways in the lab.
Finally, the team trained several machine learning algorithms to learn the complex relationships between these chemical features and organ-specific cancers. To understand not just what the models predicted but why, the researchers used feature importance analysis, which shows how individual chemical features drive each prediction.
One key finding was that biological activity data from Tox21 significantly improved cancer risk predictions for organs like the breast and pancreas. That aligns with those enzymes’ well-established roles in activating cancer-causing chemicals in the body.
All top-performing models were converted into a web dashboard. Researchers can screen chemicals using only a standard molecular code, known as a SMILES string.
“Importantly, this new web tool is not intended to completely replace traditional animal testing or expert human assessment,” said Lin, an associate professor of toxicology in the Department of Environmental and Global Health. “Instead, it serves as a powerful, early-warning screening system to indicate where follow-up testing is needed. This AI-assisted approach supports a broader movement to reduce animal use in safety testing while maintaining strong public health protection.”
Moving forward, the team plans to expand and enhance the AI system. Next steps include incorporating newly released chemical datasets to broaden the database and testing the tool against real-world data. The researchers also plan to explore more advanced molecular representations.
UF Health Cancer Institute member Christopher Vulpe, M.D., Ph.D., is a co-author on the study.
