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

Reliable modeling of chemical toxicity across species poses a fundamental challenge for molecular property prediction, as toxicity data sets are inherently heterogeneous in species coverage, endpoints, measurement scales, and data availability. From a learning perspective, cross-species acute toxicity prediction constitutes a heterogeneous multitask problem characterized by simultaneous covariate shifts in chemical feature distributions and conditional shifts in toxicity labels across endpoints such as LC50, LD50, LDLo, and TDLo. Here, we present ToxiSpecies, a task-aware meta-learning framework designed to explicitly address these distributional mismatches in low-resource toxicity modeling. The framework introduces a dual-adapter architecture that decouples species-level feature adaptation from endpoint-level modeling of toxicity semantics, enabling robust knowledge transfer across species and toxicity endpoints under low-resource settings. ToxiSpecies was evaluated on a large-scale benchmark comprising 107 endpoint-species tasks, 42 species, and 78,820 chemicals, spanning aquatic and mammalian acute toxicity measurements. Across this heterogeneous benchmark, ToxiSpecies achieved strong and broadly competitive performance relative to classical quantitative structure-activity relationship models, modern deep learning baselines, and representative meta-learning approaches. Under low-resource intra-aquatic transfer settings, the feature adapter substantially reduced root mean squared error (up to 72.6% in low-resource intra-aquatic transfer), whereas the dual adapter provided a more balanced overall trade-off between prediction accuracy and correlation performance and showed clear benefits in the 8:1:1 mixed-transfer setting. Beyond aggregate accuracy gains, ToxiSpecies demonstrated improved robustness under low-data conditions, reliable animal-to-human extrapolation, and enhanced separation of structurally related compounds exhibiting substantially divergent toxicity profiles. Embedding analyses further revealed that the learned representations capture fine-grained structure-toxicity determinants relevant to hazard differentiation and endpoint-specific sensitivity. Together, these results suggest that ToxiSpecies represents a promising framework for heterogeneous chemical toxicity prediction, potentially supporting data-driven hazard prioritization in regulatory toxicology.