Causality-integrated graph learning for multi-endpoint toxicity prediction

Toxicity prediction remains a longstanding challenge in the safety assessment of numerous artificial chemicals. While conventional single-endpoint models are effective for predicting many molecular properties, they fail to capture the complex mechanisms underlying toxicity, which involve multiple molecular targets, interconnected pathways, and diverse outcomes. We present a causality-integrated graph learning framework that embeds toxicological mechanisms extracted from large-scale literature mining as directed graphs within deep learning (DL) models. By using chemical structure as input, the framework generates compound-specific perturbation profiles within a fixed causal space, enabling system-level classification, quantitative prioritization, and mechanistic interpretation across multiple outcomes. Focusing on endocrine-disrupting chemicals, we constructed a large-scale, causally organized knowledge graph (EDKG) and implemented the framework as EDKG-DL, a predictive model that incorporates mechanism-aware graph reasoning. Through extensive external validations, EDKG-DL outperforms structure-driven state-of-the-art approaches in both stability and cross-scenario generalization. This work establishes a mechanism-constrained, causality-informed learning paradigm that is highly relevant for multi-endpoint toxicity assessment and regulatory decision-making.