Multiscale-Information-Embedded Universal Toxicity Prediction Framework

Abstract

Hazard identification and labeling of industrial chemicals and their released environmental pollutants are crucial for mitigating ecological and health risks. Comprehensive evaluation of multiple toxicity end points is essential to fully characterize chemical hazards. Existing deep-learning-based toxicity prediction models often exhibit poor generalizability, especially for rare toxicities with sparse data. Most studies fail to capture the three-dimensional (3D) spatial arrangement and stereochemical properties of chemicals, as well as the interrelated nature among end points, hindering accurate toxicity profiling. Here, we propose ToxScan, an SE(3)-equivariant multiscale model, as a universal toxicity prediction framework to address these issues. It incorporates 3D geometry information through a two-level molecular and atomic representation learning protocol. A parallel multiscale modeling and a multitask learning scheme are applied to learn universal toxicological characteristics. Results show that ToxScan achieves 7.8–37.6% improvements over state-of-the-art models for medium-/small-scale end points, demonstrates differentiation of structural analogues with contrasting toxicities, and maintains generalizability to environmental pollutants. Interpretability analysis at the atomic and molecular levels reveals identifiable atomic interaction patterns and potential structural alerts. Case studies reveal its capacity to detect subtle structural determinants while elucidating the mechanisms of pollutants. To facilitate user accessibility, we provide an intuitive web platform (https://funmg.dp.tech/Toxscan) for the rapid prediction of multiple toxicity end points of new compounds.