Benchmarking of Molecular Fingerprints for PFAS molecules

We present a computational approach and a refined annotated dataset for per- and polyfluoroalkyl substances (PFAS) classification that supports a comprehensive classification of all PFAS compounds from the U.S. Environmental Protection Agency (EPA) 2024 PFAS8a7v3 list. This work builds on our earlier efforts, including PFAS-Atlas. Traditional classification schemes leave many PFAS compounds unclassified due to their structural complexity and deviation from standard definitions. We provide a proximity-based network resource derived from the U.S. EPA’s 2024 PFAS8a7v3 list of 13,028 compounds. This framework captures structural connectivity using K-Nearest Neighbor Graphs (K-NNG) and molecular fingerprints. We evaluate eight fingerprint types against established classification systems using graph-theoretic metrics such as assortativity and modularity. Graph-based fingerprints, particularly MHFP and Mol2Vec, showed the highest alignment with known classes. We introduce a proximity-based classification method that integrates 288 previously unclassified PFASs, achieving 100% coverage. All classification outputs, networks, and code are publicly available for reuse, extension to new datasets, and integration into prioritization workflows for safer chemical design, regulatory screening, and consistent classification of future entries.