Uncertainty-Informed Deep Transfer Learning of Perfluoroalkyl and Polyfluoroalkyl Substance Toxicity

Area of Science:

• Environmental Chemistry
• Toxicology
• Computational Chemistry

Background:

• Perfluoroalkyl and polyfluoroalkyl substances (PFAS) are persistent environmental contaminants with significant human health concerns.
• Existing toxicity data for PFAS is limited, hindering comprehensive risk assessment.
• In vivo toxicity testing is resource-intensive and time-consuming.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting the acute toxicity (LD50) of PFAS compounds.
• To address the scarcity of PFAS-specific toxicity data through transfer learning from broader organic compound datasets.
• To generate reliable toxicity predictions with associated uncertainty estimates for PFAS.

Main Methods:

• Utilized multiple ML methods including random forests, deep neural networks (DNN), graph convolutional networks, and Gaussian processes.
• Aggregated publicly available oral rat LD50 data for organic compounds to train ML source models.
• Employed transfer learning by applying a DNN source model to 519 fluorinated compounds for PFAS toxicity prediction.
• Integrated a SelectiveNet architecture for uncertainty quantification and identification of high-confidence predictions.

Main Results:

• Developed ML models capable of predicting acute toxicity for PFAS with defined chemical structures.
• Successfully transferred knowledge from general organic compound toxicity data to the PFAS domain.
• The SelectiveNet architecture enabled the model to identify and abstain from predictions with high uncertainty.

Conclusions:

• Machine learning, particularly DNNs with transfer learning and uncertainty quantification, offers a viable alternative to traditional in vivo toxicity testing for PFAS.
• This approach can accelerate the assessment of numerous PFAS chemicals, informing regulatory decisions and risk management.
• Accurate toxicity prediction with uncertainty assessment is crucial for managing the risks associated with persistent environmental contaminants like PFAS.