Abstract
Humans are exposed to thousands of chemicals throughout their life. Many of these chemicals are detected in blood and have been catalogued in the Blood Exposome Database. Comprehensive hazard assessment of a chemical requires time-consuming and costly lab experiments using animal or cell-lines, which cannot be easily scaled up to the chemical exposome, highlighting the urgent need for computational approaches that can prioritize chemicals based on toxicological information. In this study, we trained direct message passing neural networks (D-MPNN) models using the Chemprop framework, chemical structure, and bioactivity data from 9,581 compounds profiled in theTox21 program across 148 quantitative high-throughput screening assays corresponding to distinct biological endpoints. Additionally, we trained a complementary model using chemical structures (n = 264,601) labeled with known UN-GHS classifications for acute oral toxicity. The Tox21 bioactivity and UN-GHS models demonstrated strong predictive performances, with the 47 bioactivity models and the GHS classification model each achieving AUCs greater than 0.80. We applied these high accuracy models to 52,055 chemicals from the Blood Exposome Database to predict bioactivity and the GHS hazard classification, enabling scalable in-silico prioritization of understudied chemical exposures for further toxicological investigations. Data and code are available at https://zenodo.org/records/17560382 and https://github.com/idslme/exposome-toxicity-prediction .Scientific contributions: We have developed an integrated predictive toxicology workflow for exposomics. This unique workflow utilizes Tox21 endpoints, UN-GHS classification and ADME property prediction models, all implemented in the state-of-the-art Chemprop framework, enabling a ranking of exposome compounds for their toxicity potential. By applying it for the Blood Exposome Database, we have prioritized several chemicals that have been detected in blood samples but can be toxic to human health.