NORMAN Kow model v0.1: QMRF, model card and SusDat predictions

Abstract

Regression model for the n-octanol/water partition coefficient (log Kow) of the neutral form of organic chemicals, developed for the NORMAN network: the mean of gradient-boosted trees (XGBoost and LightGBM) and a graph neural network (Chemprop D-MPNN). It was trained on 14 836 structures with measured values, 13 530 from the LogP_QR data set of OPERA and 1 306 from curated REACH study results. On a locked external set of REACH structures that neither OPERA 2.9 nor the EPI Suite data hold, the model reached RMSE 0.64 log units against 0.95 for KOWWIN and 0.92 for OPERA 2.9 on the 287 structures with a value from all three models, after an outlier rule fixed in advance (all 292 labels: 0.96, 1.33 and 1.03). Every prediction comes with an applicability-domain level, a calibrated reliability (R10, R3) and a 90 % prediction interval. The model serves the screening of hazard properties, in particular bioaccumulation, fills data gaps in the NORMAN Substance Database (SusDat) and supplies log Kow as an input for other NORMAN models. This record contains the QMRF (OECD QMRF v2.1), the model card, the public part of the training and validation set, the predictions for 95 792 SusDat structures, also as an upload file for the NORMAN Database System, the figures of the QMRF with their data, the rule files fixed before the runs, the validation statistics and the scripts. Start with the QMRF and README.md. Not included: the 15 model files and the measured values from the REACH study results. Both are available on request from Dr. Peter C. von der Ohe, peter.vonderohe@uba.de, under the terms of use of the ECHA REACH Study Results.