What is it about?
Endocrine-disrupting chemicals can interfere with the body’s natural hormone systems and may contribute to harmful health effects. However, testing thousands of chemicals using traditional experiments is slow, costly, and often provides limited insight into how these effects occur. In this study, we developed an AI approach that combines chemical structure with established biological knowledge about how hormone disruption can lead to adverse effects. Rather than simply predicting whether a chemical may be harmful, the model can also identify the biological pathways through which these effects may occur, predict specific adverse outcomes, and estimate dose levels below which harmful effects were not observed. This approach makes chemical safety prediction more interpretable and could help researchers identify potentially harmful chemicals more efficiently and prioritize them for further testing.
Featured Image
Photo by Growtika on Unsplash
Why is it important?
Most AI models for chemical safety learn statistical patterns from existing data and provide predictions without explaining how a chemical may cause harm. Our work takes a different approach by building established biological cause-and-effect relationships directly into the AI model. This allows the model to move beyond a simple “harmful or not” prediction. Starting from chemical structure alone, it can predict endocrine-disrupting activity, identify specific adverse outcomes, estimate safe dose levels, and trace the biological pathways that may connect chemical exposure to those effects. As the number of chemicals in commerce continues to grow far faster than they can be tested experimentally, this combination of prediction and biological explanation could help make chemical safety assessment faster, more transparent, and more useful for prioritizing further testing.
Perspectives
What excites me most about this work is the possibility of making AI not only more accurate, but also more scientifically meaningful. In chemical safety research, a prediction is much more useful when we can understand why the model made it and how it relates to what we already know about biology. For me, EDKG-DL represents a step toward AI that works with scientific knowledge rather than treating it simply as another source of data. I hope this work encourages more efforts to connect AI with causal biological understanding, so that computational models can become not just prediction tools, but tools that help scientists reason, generate hypotheses, and make better-informed decisions about chemical safety.
Wei Shi
Nanjing University
Read the Original
This page is a summary of: Causality-integrated graph learning for multi-endpoint toxicity prediction, Proceedings of the National Academy of Sciences, October 2026, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2608919123.
You can read the full text:
Contributors
The following have contributed to this page







