Interpretable and causal machine learning unveil novel insights of PFAS effects on aquatic microalgal activity: key driving features, multidimensional interactions, nonlinear microalgal response patterns and species-dependent structural effects.

Widespread distribution of per- and polyfluoroalkyl substances (PFAS) arouses ongoing concern on their eco-risks. As crucial aquatic primary producers, diverse microalgae inevitably coexist with PFAS contamination, understanding PFAS impact on microalgae is vital for eco-risks assessment/control. However, multiple features including PFAS structural heterogeneity and microalgal characteristics severely impact eco-risk assessment. Disentangling non-linear microalgal responses to key features and feature interaction, and microalgal species-dependent effects of complex molecular microstructures remains challenging. This study integrated interpretable and causal machine-learning approaches to optimize a modeling framework for interpreting PFAS toxicity to microalgae. Optuna-optimized CatBoost model with 7-fold cross-validation achieved robust performance in predicting microalgal activity at PFAS exposure. PFAS concentration, exposure time duration, chain length, microalgal species and initial cell density were identified as key features governing microalgal activity, with a model-derived aggregated response transition around the 10⁴ µg/L concentration range and an intensified inhibitory pattern beyond C7 chain length. The stronger contribution of exposure features, especially PFAS concentration, indicated that the model is effective in capturing microalgal response pattern across heterogeneous exposure scenarios, while PFAS-related structural features still provided crucial information for interpreting PFAS structural effects. Feature interaction analysis verified that PFAS concentration and chain length exhibited synergistic inhibitory effects, whereas initial cell density provided antagonistic buffering capacity. Moreover, X-Learner-based causal inference revealed divergent effects of PFAS chain length, sulfonate group, ether bond and saturated fluorotelomer on different microalgal species, and proposed a conceptual interpretation termed "bio-interface structure-molecular conformation matching": rigid long-chain PFAS may interact more strongly with EPS-rich microalgal species, sulfonate-containing PFAS may encounter algaenan-related resistance in Chlorophyceae-belonging green algae, while saturated fluorotelomers may show greater conformational adaptability toward loose-cell-wall diatom via "flexible hinge" effect, whereas ether bond-containing PFAS is more likely to experience interfacial retention within irregularly porous rough surface of Scenedesmus obliquus. This study facilitates PFAS eco-risk assessment and provides novel interpretive insights into PFAS-microalgae interactions across heterogeneous exposure scenarios.