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

Abstract

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 OECD Hershberger and uterotrophic assays. In Tier-1, multitask GNNs learn ER/AR molecular-initiating and key events across 46 ToxCast/Tox21 assays. In Tier-2, a cross-attentive multimodal GNN integrates Tier-1 pathway signals with molecular graphs, achieving AUROC{square}={square}0.90 (Hershberger) and 0.96 (uterotrophic). External validation on literature compounds showed 84% concordance (Hershberger 15/18; uterotrophic 22/26). Bidirectional cross-attention links molecular substructures to pathway assays and vice-versa, while counterfactual perturbations rank assays and structural motifs most responsible for each decision. The framework couples high accuracy with assay-traceable explanations, supporting targeted testing within the Integrated Approaches.