Machine Learning for Predicting Environmental Mobility Based on Retention Behavior

Area of Science:

• Environmental Chemistry
• Cheminformatics
• Toxicology

Background:

• Very persistent and very mobile (vPvM) substances pose risks to ecosystems and human health.
• Assessing chemical mobility is vital, but experimental data like the organic carbon-water partition coefficient (Koc) are scarce for most chemicals.
• Thousands of new chemicals necessitate efficient prioritization tools.

Purpose of the Study:

• To develop and validate a predictive model for chemical environmental mobility using readily available chromatography data.
• To establish a scalable cheminformatics approach for identifying vPvM substances.

Main Methods:

• Utilized reversed-phase liquid chromatography (RPLC) data from 146,902 chemicals to assign mobility labels.
• Computed 881 PubChem fingerprints for each chemical to represent structural features.
• Trained a random forest classifier to predict mobility based on RPLC retention behavior and chemical fingerprints.

Main Results:

• The random forest model achieved high F1 scores: 0.87 (very mobile), 0.81 (mobile), and 0.96 (nonmobile) on the test set.
• Applied to 64,492 REACH-registered chemicals, the model classified 20% as very mobile, 26% as mobile, and 53% as nonmobile.
• Demonstrated the model's scalability for early identification of potentially harmful vPvM substances.

Conclusions:

• A robust cheminformatics model effectively predicts chemical environmental mobility using RPLC data and structural fingerprints.
• This approach offers a scalable solution for prioritizing chemicals, aiding in the early identification of vPvM substances.
• The findings support proactive environmental risk assessment and management of chemical substances.