Machine Learning for Predicting Environmental Mobility Based on Retention Behavior

Very persistent and very mobile (vPvM) substances threaten the environment and human health. These chemicals can persist in aquatic systems and move rapidly due to their affinity for water over soil or other adsorbents. Chemical mobility is usually classified using the organic carbon-water partition coefficient (<i>K</i>_oc), but experimental log <i>K</i>_oc data are unavailable for most substances. With thousands of new chemicals entering the market annually, there is a growing need for advanced cheminformatics tools to prioritize substances of concern. Because reversed-phase liquid chromatography (RPLC) data are more widely available, they were used here as a proxy for environmental mobility. The organic modifier fraction at elution was applied to assign mobility labels to 146,902 chemicals from an RPLC data set. For each chemical, 881 PubChem fingerprints were computed to capture structural information. A random forest classifier was then trained to predict mobility from retention behavior and fingerprints. The model achieved F1 scores of 0.87, 0.81, and 0.96 for very mobile, mobile, and nonmobile classes, respectively, in the test set. Applied to all REACH-registered chemicals (<i>n</i> = 64,492), the model classified 20% as very mobile, 26% as mobile, and 53% as nonmobile, providing a scalable tool for early identification of vPvM substances.