Advancing aquatic hazard assessment: Curated fish toxicity data and benchmarking of predictive models

Maximizing the use of already generated in vivo ecotoxicological data is essential, to develop and validate alternative in vitro and in silico methods, establish relative species sensitivity, and assess the role of fish acute and chronic studies in hazard assessment. We constructed high-quality reference datasets for fish toxicity by extracting experimental studies from REACH registration dossiers and rigorously curating them using AI-assisted algorithms. Strict filtering criteria were applied to ensure study reliability, which include verification of the exposure concentrations, overall compliance with test guidelines, and accurate molecular structure assignment. The resulting dataset includes acute toxicity data for 2315 structures and chronic toxicity data for 418 structures, encompassing a broad spectrum of industrial chemistries, test designs, and regulatory guidelines. The datasets contain short-term fish toxicity LC50 values for 610 standardised structures and fish early-life stage toxicity NOEC/EC10 values for 135 standardised structures amenable to modelling, when considering only bounded effects levels derived from concentration-response modelling using analytically verified exposure concentrations. These were used to quantitatively assess the accuracy and applicability domain coverage of widely used predictive models. We benchmarked the models for their ability to predict effect concentrations and regulatory classification thresholds both globally and within chemical subspaces defined with a range of clustering methods. For acute toxicity, up to 94% of LC50 values can be predicted within 1 log unit at an applicability domain coverage of 30%; this coverage increases to nearly 80% when 85% of the predictions fall within 1 log unit. Absolute error increases with ionisation potential, and local predictive performance varies substantially even within the applicability domain. For chronic toxicity, performance was considerably lower, particularly for highly toxic substances, with much smaller applicability domain coverage due to the scarcity of experimental data, which also makes the model assessment less conclusive. Overall, our analysis demonstrates the potential of modelling for aquatic toxicity hazard assessment, highlights current limitations, and provides curated data and suggestions to support the further development and promotion of predictive methods as alternatives to in vivo testing.