Improving the Ecotoxicological Hazard Assessment of Chemicals by Pairwise Learning

This study demonstrates how machine learning techniques can bridge data gaps in the ecotoxicological hazard assessment of chemical pollutants and illustrates how the results can be used in practice. The innovation herein consists of the prediction of the sensitivity of all species that were tested for at least one chemical for all chemicals based on all available data. As proof of concept, pairwise learning was applied to 3295 × 1267 (chemical,species) pairs of Observed LC50 data, where only 0.5% of the pairs have experimental data. This yielded more than four million Predicted LC50s for separate exposure durations. These were used to create (1) a novel Hazard Heatmap of Predicted LC50s, (2) Species Sensitivity Distributions (SSD) for all chemicals based on 1267 species each, as well as (3) for taxonomic groups separately, and (4) newly defined Chemical Hazard Distributions (CHD) for all species based on 3295 chemicals each. Validation results and graphical examples illustrate the utility of the results and highlight species and compound selection biases in the input data. The results are broadly applicable, ranging from Safe and Sustainable by Design (SSbD) assessments and setting protective standards to Life Cycle Assessment of products and assessing and mitigating impacts of chemical pollution on biodiversity.