Interpretable QSPR modeling of PFAS endpoints via transformer-driven SMILES-to-molecular descriptor learning

A data-centric and interpretable Quantitative structure-property relationship (QSPR) framework was developed for four Per- and polyfluoroalkyl substances (PFAS) endpoints (i.e., LC₅₀, log BCF, log K_OA, and log K_OW) with molecular- and descriptor-level interpretation of prediction behavior. First, PFAS endpoint datasets were collected while molecular descriptors and tokenized SMILES sequences were integrated for descriptor imputation. Descriptor redundancy was then reduced through statistical feature selection, and MissTransformer reconstructed missing descriptor values from latent structural information encoded in SMILES. Finally, explainable AI (XAI) analysis was applied to interpret chemical descriptor contributions to PFAS endpoint prediction. As a result, MissTransformer increased the number of descriptor-complete compounds available for QSPR modeling by 133.71% for log BCF data. Among the endpoint prediction models, the data-centric XGB model showed the best performance for log K_OW prediction, with R² and Q²_LMO values of 0.976 and 0.775, respectively. Interpretation results showed that hydrophobicity-related descriptors and fluorine atoms strongly contributed to log K_OW prediction, whereas global structural descriptors related to molecular size and steric effects were important for log BCF prediction. The proposed framework can be used to prioritize data-poor PFAS compounds for further testing, monitoring, and risk-based environmental management.