Machine Learning-Based Prediction of NOECs for over 10,000 PFAS across Multiple Fish Species Enables Large-Scale...

Area of Science:

• Environmental Chemistry
• Toxicology
• Ecotoxicology

Background:

• Per- and polyfluoroalkyl substances (PFASs) are widespread, persistent chemicals in aquatic environments, posing ecological risks.
• No-observed-effect concentrations (NOECs) are crucial for assessing fish safety but are undefined for most PFASs.
• Existing data on PFAS toxicity in fish is limited, hindering comprehensive risk assessment.

Purpose of the Study:

• To develop a machine learning model for predicting mortality-based NOECs for a large number of PFASs in fish.
• To identify structure-toxicity relationships and species-specific sensitivities to PFASs.
• To establish a framework for high-throughput toxicity prediction and ecological risk screening.

Main Methods:

• Compiled existing toxicity data for 10,863 PFASs in *Danio rerio*.
• Developed a machine learning model using molecular structural descriptors to predict NOECs.
• Extrapolated toxicity profiles to 12 additional fish species using trait-based inference.

Main Results:

• The machine learning model demonstrated robust predictive performance (R2test = 0.7096).
• Predicted NOECs varied widely, with higher toxicity observed for C8-C12 PFASs and those with polar functional groups (sulfonates, carboxylates).
• Warm water and omnivorous fish species were found to be more sensitive to PFASs.

Conclusions:

• The study provides mechanistic insights into PFAS toxicity in fish, linking chemical structure to ecological characteristics.
• The developed model enables high-throughput toxicity predictions and supports ecological risk screening for numerous PFASs.
• Structure-species interactions are critical in determining overall PFAS toxicity patterns across diverse fish populations.