Using Machine Learning to Classify Bioactivity for 3486 Per- and Polyfluoroalkyl Substances (PFASs) from the OECD List

A recent OECD report estimated that more than 4000 per- and polyfluorinated alkyl substances (PFASs) have been produced and used in a broad range of industrial and consumer applications. However, little is known about the potential hazards (e.g., bioactivity, bioaccumulation, and toxicity) of most PFASs. Here, we built machine-learning-based quantitative structure–activity relationship (QSAR) models to predict the bioactivity of those PFASs. By examining a number of available molecular data sets, we constructed the first PFAS-specific database that contains the bioactivity information on 1012 PFASs for 26 bioassays. On the basis of the collected PFAS data set, we trained 5 different machine learning models that cover a variety of conventional models (e.g., random forest and multitask neural network (MNN)) and advanced graph-based models (e.g., graph convolutional network). Those models were evaluated based on the validation data set. Both MNN and graph-based models demonstrated the best performance. The average of the best area-under-the-curve score for each bioassay is 0.916. For predictions on the OECD list, most of the biologically active PFASs have perfluoroalkyl chain lengths less than 12 and are categorized into fluorotelomer-related compounds and perfluoroalkyl acids and their precursors.