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

Area of Science:

• Environmental Chemistry
• Toxicology
• Computational Chemistry

Background:

• Over 4000 per- and polyfluorinated alkyl substances (PFASs) are in industrial and consumer use.
• Limited data exists on the bioactivity, bioaccumulation, and toxicity of most PFASs.
• Understanding PFAS hazards is crucial for environmental and human health protection.

Purpose of the Study:

• To develop predictive models for PFAS bioactivity using machine learning.
• To create the first comprehensive PFAS-specific database for bioactivity.
• To identify structural features associated with PFAS bioactivity.

Main Methods:

• Compiled a database of 1012 PFASs across 26 bioassays.
• Trained five machine learning models, including random forest, multitask neural network (MNN), and graph convolutional networks.
• Evaluated model performance using a validation dataset, focusing on area-under-the-curve (AUC) scores.

Main Results:

• Multitask neural network and graph-based models showed superior performance.
• Achieved an average best AUC score of 0.916 across bioassays.
• Identified that biologically active PFASs often have perfluoroalkyl chain lengths under 12 and include fluorotelomer-based compounds and perfluoroalkyl acids.

Conclusions:

• Machine learning, particularly MNN and graph-based models, effectively predicts PFAS bioactivity.
• The developed PFAS database and models provide a valuable tool for hazard assessment.
• Findings aid in prioritizing PFASs for further toxicological investigation and regulatory action.