Multimodal Cross-Attentive Graph-Based Framework for Predicting In Vivo Endocrine Disruptors.

Area of Science:

• Computational toxicology and cheminformatics
• Endocrine disruption and hazard assessment
• Adverse Outcome Pathway (AOP) framework

Background:

• Accurate and mechanistically transparent models are crucial for endocrine hazard assessment.
• Existing methods often lack transparency or require extensive in vivo testing.
• Integrating molecular data with pathway information can improve predictive accuracy.

Purpose of the Study:

• To develop a multimodal, cross-attentive graph framework for predicting organism-level endocrine disruption outcomes.
• To fuse molecular graphs with Adverse Outcome Pathway (AOP)-anchored assay signals.
• To enhance mechanistic transparency in endocrine hazard assessment.

Main Methods:

• Utilized multitask graph neural networks (GNNs) in Tier-1 to learn key events from 46 in vitro ToxCast/Tox21 assays.
• Employed a cross-attentive multimodal GNN in Tier-2 to integrate pathway signals with molecular graphs.
• Applied bidirectional cross-attention and counterfactual perturbations for interpretability.

Main Results:

• Achieved high predictive performance for the in vivo Hershberger (AUROC = 0.97 ± 0.014) and uterotrophic (AUROC = 0.97 ± 0.008) assays.
• Demonstrated 88% concordance with literature data for tested compounds.
• Identified key molecular substructures and assays influencing predictions.

Conclusions:

• The developed framework accurately predicts endocrine disruption outcomes with mechanistic interpretability.
• This approach supports targeted testing strategies within integrated approaches to chemical safety assessment.
• The model enhances transparency and accuracy in endocrine hazard assessment, aligning with regulatory needs.