Mechanistically Interpretable Toxicity Prediction Through Multimodal Integration of Structure and Transcriptomics

Abstract

Advances in computational toxicology increasingly emphasize the need for models that deliver both scalable predictive performance and mechanistic insights. Nevertheless, most approaches fall short of capturing the underlying mechanisms that drive toxicity. Herein, we describe a scalable multimodal modeling framework that integrates chemical fingerprints with high-throughput transcriptomic dose-response profiles across three human cell lines to predict activity for 41 curated Tox21 assay endpoints. Using gradient-boosted decision trees and nested compound-aware cross-validation, 13 assays achieved robust performance (mean AUPRC > 0.75), spanning nuclear receptor signaling, stress-response pathways, and xenobiotic metabolism. SHAP-based feature attribution analysis showed that predictions depend on both structural motifs and transcriptional programs, in a manner consistent with established mechanistic relationships between chemical structure, nuclear receptor biology, and adaptive cellular responses. These findings illustrate how structure-HTTr dose-response signature integration enables models that are accurate and mechanistically grounded, shifting computational toxicology toward transparent and biologically informed chemical risk evaluation.