Prediction and mechanism analysis of octanol-air partition coefficient for persistent organic pollutants based on machine learning models

The octanol-air partition coefficient (K OA ) is a critical physicochemical parameter describing the migration and distribution of persistent organic pollutants (POPs). Previous methods for detecting and predicting K OA have been time-consuming, labor-intensive, and costly, with limitations in both applicability and explanatory mechanisms. This study utilized machine learning to identify the optimal model and influencing mechanisms for K OA , employing molecular descriptors and extended connectivity fingerprints (ECFP). The study revealed that 1–3D-XGBoost demonstrated superior predictive performance and generalization capabilities for K OA (R 2 = 0.98, RMSE = 0.30), exhibiting smaller prediction errors than previous models. The SHAP and feature interaction analysis identified MLFER_L (solute gas-hexadecane partition coefficient), ALOGP.2 (hydrophobicity), and temperature as the primary factors influencing K OA . The study suggests that higher K OA , which warrants greater attention due to sedimentation and polar pollution, is likely when MLFER_L > 11, temperature < 20 ℃, MoRSEE3 > 2, and VR3_Dt > 10. Conversely, under other conditions, the focus should shift toward the migration of pollutants. Additionally, the ECFP model highlighted that oxygen-containing functional groups such as ether bonds, carbonyls, and hydroxyl groups significantly affect K OA . The study proposes that pollutants with ether bonds and symmetrical molecular structures or containing carbonyl groups, exhibit higher K OA . When the molecular structure of the ether bond is asymmetric, it leads to a lower K OA . These insights enhance valuable molecular structural references for the prediction and mechanism explanation of K OA . This study improves the prediction of K OA and offers critical insights into the migration, tracing, detection, and treatment of POPs. • The XGBoost based on 1–3D descriptors has the best prediction performance for K OA . • The feature factors affecting K OA have been identified by 1–3D descriptors. • Proposed feature descriptors interaction information that affects K OA . • Oxygen functional groups have the greatest impact on K OA .