Mapping toxicity pathways of per- and polyfluoroalkyl substances using interpretable classification-based machine...

Area of Science:

• Environmental Chemistry
• Toxicology
• Computational Chemistry

Background:

• Per- and polyfluoroalkyl substances (PFAS) are widespread environmental contaminants with known adverse health effects.
• Limited experimental toxicity data exists for most PFAS, hindering comprehensive risk assessment.
• High-throughput screening (HTS) provides valuable toxicity data but requires robust predictive models for large chemical spaces.

Purpose of the Study:

• To develop and validate machine learning (ML)-based Quantitative Structure Toxicity Relationship (QSTR) models for predicting PFAS toxicity.
• To identify robust QSTR models for three key high-throughput screening endpoints: ALDH1A1 inhibition (AID-1030), Nrf2 pathway inhibition (AID-504444), and TGF-β/Smad3 signalling inhibition (AID-588855).
• To provide mechanistic insights into PFAS toxicity through structure-activity relationship analysis and develop a user-friendly screening tool.

Main Methods:

• Development of ML-based QSTR classification models using various data balancing techniques (ADASYN, SMOTE, etc.) to address class imbalance.
• Training and evaluation of 14 ML classifiers across multiple balanced datasets for each endpoint, resulting in 70 models per endpoint.
• Utilizing Sum-of-Ranking-Differences (SRD), SHAP, and substructure analyses to identify optimal models and interpret structure-toxicity relationships.

Main Results:

• Gradient Boosting, Random Forest, and Support Vector Classifier models were identified as optimal for AID-1030, AID-504444, and AID-588855, respectively.
• SHAP and substructure analyses revealed key PFAS structural features influencing toxicity and provided mechanistic interpretability.
• The developed QSTR models demonstrated good performance when applied to an independent external dataset of 2,361 PFAS.

Conclusions:

• ML-based QSTR models can effectively predict PFAS toxicity across multiple endpoints, overcoming limitations of experimental data scarcity.
• The developed models offer valuable tools for prioritizing PFAS for further testing and informing regulatory risk assessments.
• A Python-based screening tool, PERSIST, was created to facilitate the application of these predictive models for environmental safety assessments.