Bioactivity Classification of 2,649 Per- and Polyfluoroalkyl Substances (PFASs) via Quantitative Structure–Activity Relationships and Molecular Docking to Health-Relevant Proteins

The structural diversity of poly- and perfluoroalkyl substances (PFASs) poses challenges for class-based chemical hazard assessments. Computational simulations of PFAS-protein binding provide a means to improve screening-level evaluations by considering molecular interactions with toxicologically relevant proteins. Many PFASs are known to interact with carrier and channel proteins that influence bioaccumulation and toxicity, yet experimental data remain limited. To address this gap, we developed multicondition quantitative structure-activity relationship (QSAR) models that incorporate PFAS-protein associations. We simulated interactions of 4,397 PFASs with six health-relevant proteins, generating molecular docking binding scores for 12 ligand-binding sites. We then constructed seven bioactivity prediction models, integrating 44 additional chemical descriptors for each target. Training used binary bioactivity classifications for 560 PFASs derived from public toxicology and bioactivity databases. Using the best performing model (random forest), we classified the bioactivity potential of 2,649 PFASs within the applicability domain. Strong binding of PFASs to human serum albumin (fatty acid binding sites 3 and 4) and peroxisome proliferator-activated receptor γ, along with molecular complexity and the number of saturated hetero rings, were associated with PFAS bioactivity. A structure-activity landscape index analysis further suggested that nonfluorinated moietiesincluding bromine substitution, benzimidazole, and dinitroaniline groupscontribute to PFAS bioactivity.