FLAME: Fluorinated-motif Learning and Attribution for Molecular Explainability Applied to PFAS

Machine-learning (ML) architectures such as graph neural networks (GNNs) are increasingly used in structure-property modeling and molecular property prediction in fields of emerging contaminants such as per-and polyfluoroalkyl substances (PFAS). However, their black-box nature makes these models difficult to use in scientific interpretation, decision making and in providing identifiable molecular substructures that contributed to the predictions. In attempts to open up the black-box, fragment-based explainability techniques have been created to capture molecular motifs that contributed to the predicted properties. However, these fragment-based explanations are dependent on how the molecular structures are partitioned and general molecular fragmentation methods may not capture highly fluorinated motifs, functional headgroups, and linker environments relevant for PFAS and related fluorinated compounds. The resulting fragment explanations may be less chemically meaningful unless fragmentation approaches account explicitly for the distinctive chemistry of fluorinated molecular motifs. We present Fluorinated-motif Learning and Attribution for Molecular Explainability (FLAME), a PFAS-tailored framework that extends rBRICS fragmentation to improve fragment-based explainability methods. The additional cleavage environments improved the separation of PFASrelevant structures, enabling the separation of 568 additional molecules relative to the original rBRICS in a dataset of 4,738 molecules. The substructure masking explanation method was then applied to a graph neural network that predicted electron affinity, ionization potential, and HOMO–LUMO gap. The chemically tailored fragments provided identifiable units for substructure masking allowing more interpretable predictions based on motif-level attributions. Our results demonstrate a constrained explainability framework built upon chemically relevant structural motifs. FLAME is extensible to other classes of emerging contaminants whose characteristic structural motifs require tailored fragmentation for scientifically meaningful model explanations.