Abstract
Humans are exposed to many chemicals, yet limited toxicity data hinder effectively managing their impact on human health. High-performing machine learning models hold potential for addressing this gap, but their uncharacterized prediction performance across the wider chemical space undermines confidence in their results. We developed uncertainty-aware models to predict reproductive/developmental and general non-cancer human toxicity points of departure, aimed at supporting chemical risk and impact assessments. Our well-calibrated models provide uncertainty estimates aligned with observed prediction errors and chemical familiarity. We predict toxicity with 95% confidence intervals for >130,000 marketed chemicals and identify toxicity and uncertainty hotspots, including organothiophosphates, PFAS, steroids and dioxin-like compounds. These insights help prioritize efforts to reduce human health risks and improve model performance. By enhancing transparency in prediction uncertainty, our work supports the sound integration of ML-based predictions in chemical assessments and provides key insights for systematically building confidence in digitally derived toxicity information.

This publication has 58 references indexed in Scilit: