NORMAN aquatic ecotoxicity model v0.1: QMRF, QPRF example, model card and SusDat predictions

Abstract

A multi-species model for the acute and chronic aquatic toxicity of organic chemicals, developed for the NORMAN network. It predicts toxicity values (log10 µg/L) for 121 species-and-endpoint combinations (heads), of which 82 are released for use (59 acute: LC50/EC50; 23 chronic: NOEC/EC10), covering fish, crustaceans, algae, macrophytes, insects, molluscs, amphibians and other invertebrates. Eight learners (neural networks with a shared chemical representation and taxonomic context, a fine-tuned graph network and a gradient-boosted tree model) are stacked, calibrated per head and, where a substance has measured values at other species, corrected by them (anchor stage). Every prediction carries an applicability-domain level, calibrated probabilities of lying within a factor of 10 and 3 of a measurement, a 90 % prediction interval and a lower bound. In six-fold cluster cross-validation the RMSE is 0.89 log units without and 0.70 with anchor (R² 0.67 and 0.80); on a cluster hold-out test set 0.85 and 0.66 (R² 0.68 and 0.81); on 228 comparisons at 159 substances from independent collections 1.01 (R² 0.39). Against ECOSAR 2.2 on the same test set: 1.34 for ECOSAR, 0.81 without and 0.61 with anchor. The record holds the QMRF (OECD format, with the head table and figures), an example QPRF, the model card (PDF and JSON), the README, the public part of the training and validation table with honest out-of-fold predictions, predictions with applicability-domain and reliability fields for the NORMAN Substance Database (SusDat), per-head statistics, the frozen parameters of the applicability-domain and reliability layer, validation reports, figure data and the scripts. The folders figures/, model_metadata/ and reports/ are uploaded as zip files; zenodo_paket_v0_2_md5.txt lists the checksums of the single files. On request from the contact for the NORMAN network: the fitted model files (2.6 GB), and the labels from the Pesticide Properties Database and the REACH dossier study results, whose values cannot be redistributed. Contact: Dr. Peter C. von der Ohe, author of the model and contact for the NORMAN network, peter.vonderohe@uba.de.