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

Endocrine hazard assessment needs models that are accurate and mechanistically transparent. We present a multimodal cross-attentive graph framework that fuses molecular graphs with adverse-outcome-pathway (AOP)-anchored assay signals to predict organism-level outcomes in the organisation for economic co-operation and development (OECD) Hershberger and uterotrophic assays. In Tier-1, multitask graph neural networks (GNNs) learn estrogen and androgen receptor molecular-initiating and key events across 46 in vitro ToxCast/Tox21 assays. In Tier-2, a cross-attentive multimodal GNN integrates Tier-1 pathway signals with molecular graphs, yielding high predictive performance for both the in vivo Hershberger (AUROC = 0.97 ± 0.014) and uterotrophic (AUROC = 0.97 ± 0.008) assays. Retrospective analysis of literature compounds showed 88% concordance (Hershberger 15/18; uterotrophic 23/26). Bidirectional cross-attention highlights associations between molecular substructures and pathway-level assay nodes, while counterfactual perturbations rank assays and structural motifs most influential for each decision. The framework couple's high accuracy with assay-traceable explanations, supporting targeted testing within the integrated approaches.