Assessing AlphaFold 3 for Per- and Polyfluoroalkyl Substances Docking in Protein Structures.

Per- and polyfluoroalkyl substances (PFAS) are persistent environmental pollutants that may pose health risks due to strong protein interactions. While AlphaFold 3 (AF3) was recently introduced for protein-ligand modeling with high claimed accuracy, its reliability for docking PFAS remains unclear. This study evaluates AF3's performance in predicting protein-PFAS interactions using a curated data set from the Protein Data Bank, divided into a "Before Set" (seen during AF3 training) and an "After Set" (unseen). AF3 accurately predicts protein structures and pockets but shows reduced performance in pocket-aligned ligand predictions, achieving ∼74.5% success in "Before Set" but only ∼55.8% in "After Set", indicative of possible overfitting. We further assess the effects of PFAS type on docking outcomes. Although AF3 accurately predicts binding pockets, it favors poses where the headgroup of environment-relevant PFAS interacts with polar or positively charged residues. This is different from another native binding mode in several cases, where the hydrophobic tail is inserted in the protein, but the headgroup is exposed to the solvent. Notably, a hybrid approach combining AF3 and Vina, especially considering multiple top-ranked poses, can improve prediction accuracy. These findings support the complementary use of AF3 and Vina for accurately modeling protein-PFAS interactions.