Only 3.5% of chemicals currently in trade in the EU have enough data to derive a species sensitivity distribution, the standard statistical tool used to assess how sensitive different species are to a given chemical pollutant and a key input for setting environmental quality standards. Rather than modelling each chemical in isolation, as conventional approaches typically do, the researchers treated the problem as a matrix-completion task, drawing on data across all compounds and species simultaneously to predict lethal concentration values for combinations of chemicals and species that have never actually been tested together.

The resulting outputs include hazard heatmaps, individual and grouped species sensitivity distributions, and chemical hazard distributions, which the authors say could support Safe and Sustainable by Design assessments during early-stage product development, helping companies identify chemicals of least concern before committing to full-scale production. The same data could also inform both prospective market approval decisions under REACH and retrospective pollution assessments carried out under the Water Framework Directive, potentially closing gaps on both ends of a chemical's regulatory life cycle.

The researchers also see potential to extend the approach to ecotoxicological effect factors used in Life Cycle Assessments and Product Environmental Footprinting, as well as to region-specific assessments tailored to particular types of water bodies, which could make the method relevant well beyond its initial chemical safety application. As with other AI-based alternatives to animal testing under development across the EU's research ecosystem, the key remaining hurdle is less about scientific validation than about winning acceptance from regulators such as ECHA, who would need to be satisfied that predicted values are reliable enough to substitute for measured data in formal risk assessments.

#ChemicalSafety #SafeAndSustainableByDesign