Uncertainty-aware machine learning to predict non-cancer human toxicity for the global chemicals market.

Abstract

Humans are exposed to thousands of chemicals, yet limited toxicity data hinder effective management of their impacts on human health. High-performing machine learning models hold potential for addressing this gap, but their uncharacterized prediction performance across the wider range of chemicals undermines confidence in their results. We develop uncertainty-aware models to predict reproductive/developmental and general non-cancer human toxicity effect doses. Our well-calibrated models provide uncertainty estimates aligned with observed prediction errors and chemical familiarity. We predict toxicity with 95% confidence intervals for >100,000 globally marketed chemicals and identify toxicity and uncertainty hotspots. These results can be applied to inform decisions aimed at reducing potential human health impacts and guide targeted data generation and modeling efforts to reduce prediction uncertainty. Here, we show that enhancing transparency in prediction uncertainty provides key insights for building confidence in toxicity predictions, supporting the sound integration of machine learning-based predictions in chemical assessments.