ToxiSpecies: Task-Aware Meta-Learning for Cross-Species Modeling of Acute Chemical Toxicity under Distribution Shift

Area of Science:

• Computational toxicology
• Machine learning for cheminformatics
• Quantitative structure-activity relationship (QSAR) modeling

Background:

• Chemical toxicity prediction is challenged by heterogeneous data across species, endpoints, and measurement scales.
• Cross-species toxicity modeling involves complex shifts in chemical features and toxicity labels.
• Existing methods struggle with low-resource settings and data scarcity.

Purpose of the Study:

• To introduce ToxiSpecies, a task-aware meta-learning framework for robust cross-species chemical toxicity prediction.
• To address distributional mismatches in heterogeneous, low-resource toxicity modeling.
• To enable effective knowledge transfer across species and toxicity endpoints.

Main Methods:

• Developed a dual-adapter architecture to decouple species-level feature adaptation from endpoint-level toxicity modeling.
• Evaluated the framework on a large benchmark of 107 endpoint-species tasks, 42 species, and 78,820 chemicals.
• Compared performance against QSAR models, deep learning baselines, and other meta-learning approaches.

Main Results:

• ToxiSpecies achieved competitive performance across a heterogeneous benchmark of aquatic and mammalian acute toxicity.
• The feature adapter significantly reduced error in low-resource settings (up to 72.6%).
• The dual adapter improved prediction accuracy and correlation, showing benefits in mixed-transfer scenarios.

Conclusions:

• ToxiSpecies offers a promising framework for heterogeneous chemical toxicity prediction, especially in low-data conditions.
• The model demonstrates improved robustness, reliable animal-to-human extrapolation, and better differentiation of toxic profiles.
• Learned representations capture crucial structure-toxicity determinants for hazard assessment and regulatory toxicology.