Multiscale-Information-Embedded Universal Toxicity Prediction Framework.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Accurate chemical hazard identification is vital for environmental and health safety.
• Current deep learning models struggle with generalizability, especially for rare toxicities and 3D chemical properties.
• Existing methods often fail to capture interdependencies between different toxicity endpoints.

Purpose of the Study:

• To develop a universal toxicity prediction framework, ToxScan, addressing limitations of current models.
• To incorporate 3D structural and stereochemical information for enhanced toxicity profiling.
• To improve generalizability and accuracy across diverse toxicities and environmental pollutants.

Main Methods:

• Proposed ToxScan, an SE(3)-equivariant multiscale model incorporating 3D geometry.
• Utilized a two-level molecular and atomic representation learning protocol.
• Implemented parallel multiscale modeling and multitask learning for universal toxicological characteristics.

Main Results:

• ToxScan demonstrated 7.8-37.6% improvement over state-of-the-art models for various toxicity endpoints.
• The model successfully differentiated structural analogues with contrasting toxicities.
• Maintained generalizability for environmental pollutants and provided interpretable atomic-level insights.

Conclusions:

• ToxScan offers a robust framework for accurate and generalizable chemical toxicity prediction.
• The model's interpretability aids in identifying structural alerts and elucidating pollutant mechanisms.
• An accessible web platform is available for rapid toxicity predictions of new compounds.