Molecular connectivity indices and soil properties to predict the sorption of per- and poly-fluoroalkyl substances

Abstract

This study presents a modeling approach to predict the soil-water partitioning coefficient (K_d, L kg⁻¹) for per- and poly-fluoroalkyl substances (PFASs) as a function of their molecular connectivity indices (MCIs) and soil properties (soil organic carbon, SOC, %, and cation exchange capacity, CEC, cmol kg⁻¹). The modeling framework involved compiling data, developing models, and evaluating model performance via interpretation, external validation, and scenario analyses. Two datasets consisting of simple and valence MCIs per PFAS were used: (i) carboxylic-PFCA dataset (N = 327) had only carboxylic compounds (C4–C12) and (ii) PFAS-full dataset (N = 699) entailed carboxylic acids (C4–C12), sulfonic acids (C4–C10) and fluorotelomers (C4–C8). Our multi-criteria approach revealed that the seventh-order valence path (VP-7) related to polarizability and molecular size and the third-order simple path (SP-3) related to molecular size and chain structure emerged as key predictors for the carboxylic-PFAS and PFAS-full datasets, respectively. Elastic net-regularized linear regression (MLR_EN) and artificial neural networks (ANNs) demonstrated that MCIs improved the predictive accuracy. For the PFAS-full dataset, six-predictor models (MCIs + soil properties) yielded a high predictive accuracy (R_pred² = 83.7–84.9%); however, a three-predictor MLR_EN model (SP-3, SOC, and CEC; R_pred² = 77.9%) achieved the highest external generalization (R_ext² = 52.4%). SP-3 accounted for the largest share of predictive power (68–95%), dominating the model performance (94–97%). Scenario analyses revealed that while deterministic predictions remained stable, probabilistic modeling is crucial for capturing the rare but impactful extremes. Overall, our study highlights the practical advantage of MCIs as versatile and scalable tools for predicting the adsorption of diverse PFAS, including short-chain, partially fluorinated, and less commonly studied PFASs. In the long term, this tool can provide data for preliminary, rapid, site-specific risk assessment for PFAS-impacted sites.