Location:
Understanding the multi-endpoint toxicity of chemicals is an important scientific challenge in environmental health. The enormous diversity of chemical structures, together with the multi-target, multi-pathway, and multi-endpoint nature of toxic effects, makes systematic toxicity assessment particularly difficult. Although deep learning has opened new possibilities for chemical toxicity prediction, most existing models still rely largely on statistical learning between chemical structures and individual toxicity endpoints. Without explicit incorporation of toxicological mechanisms, these approaches remain limited in their ability to achieve accurate and interpretable prediction across multiple adverse outcomes.
Recently, a research team led by Professor Wei Shi from the School of the Environment at Nanjing University developed a causality-integrated graph learning framework that explicitly embeds toxicological causal knowledge into deep learning. The framework advances toxicity prediction from conventional single-endpoint statistical learning toward mechanism-informed, system-level multi-endpoint inference. The study, entitled “Causality-Integrated Graph Learning for Multi-Endpoint Toxicity Prediction”, was published online in the Proceedings of the National Academy of Sciences of the United States of America (PNAS) on October 5, 2026.
Figure. Causality-integrated framework for multi-endpoint toxicity prediction
Using endocrine-disrupting chemicals (EDCs) as a representative case, the research team integrated literature mining, toxicological databases, and expert knowledge to construct an Endocrine Disruption Knowledge Graph (EDKG) connecting molecular initiating events, key events, and adverse outcomes. Building on this knowledge graph, the team further developed EDKG-DL, a causality-integrated graph learning model. Chemical structure serves as the compound-specific input, from which the model constructs a compound-specific “virtual perturbation map”. Perturbation signals are then propagated through the toxicological causal network, enabling the model to jointly identify endocrine-disrupting activity, predict multiple adverse outcomes, estimate dose-related effects, and trace potential upstream mechanisms of toxicity.
The causality-integrated graph learning framework demonstrated clear advantages in multi-endpoint toxicity prediction. In predictive performance, EDKG-DL outperformed dozens of existing SOTA methods that rely solely on chemical structure, achieving superior results in both qualitative and quantitative multi-endpoint prediction. In generalizability, the model maintained stable performance in time-external and cross-source evaluations. In mechanistic interpretation, it distinguished isolated target perturbations from system-level disruptive effects, reducing false-positive classifications for more than 200,000 compounds. At scale, EDKG-DL identified 9,510 potential EDCs from 567,849 registered chemicals, while reducing validation costs by 99.99% and evaluation time by 93.4%. By converting toxicological knowledge into computable causal reasoning, this work provides a new AI paradigm for analyzing complex toxicity beyond conventional structure-to-endpoint statistical prediction.
Professor Wei Shi is the corresponding author of the study, and Associate Researcher Haoyue Tan is the first author. The work brought together collaborators from institutions in China, the United States, Italy, and Japan, including Nanjing University, Zhejiang University, the U.S. Food and Drug Administration, the Istituto di Ricerche Farmacologiche Mario Negri IRCCS, and the University of Tokyo. The study was supported by the National Natural Science Foundation of China, the National Key Research and Development Program of China, and the Natural Science Foundation of Jiangsu Province, among other funding programs.
Publication Information
Causality-Integrated Graph Learning for Multi-Endpoint Toxicity Prediction
Proceedings of the National Academy of Sciences of the United States of America (PNAS), 2026
DOI: 10.1073/pnas.2608919123
Paper link: https://www.pnas.org/doi/abs/10.1073/pnas.2608919123
Kudos link: https://link.growkudos.com/1f0tvlh0xs0
