Unveiling the Structural Determinants of PFAS Toxicity: A Graph Neural Network and Interpretability Analysis

Area of Science:

• Environmental Chemistry
• Computational Toxicology
• Machine Learning

Background:

• Per- and polyfluoroalkyl substances (PFAS) are widespread, requiring efficient safety assessment tools.
• Traditional toxicology is slow and costly; current computational methods lack mechanistic detail.
• Preventing regrettable substitution of hazardous chemicals is critical.

Purpose of the Study:

• Develop a high-throughput deep learning framework for predicting acute oral rat LD50 toxicity of PFAS.
• Enhance mechanistic transparency in computational toxicology for rational chemical design.
• Assess the model's generalization potential for diverse fluorinated compounds.

Main Methods:

• Coupled graph neural networks (GNNs) with transfer learning to predict LD50 toxicity.
• Utilized a GIN architecture with stratified transfer learning (GIN-TL-Strat).
• Integrated GNNExplainer for interpretability, mapping predictions to chemical substructures.

Main Results:

• GIN-TL-Strat achieved high predictive correlation (test R^2 = 0.7426).
• Fluorinated carbon chain length was identified as the primary toxicity determinant in continuous-chain alkyl PFAS.
• Model revealed that heavy halogens (I, Br, Cl) correlated with reduced toxicity, decoupling model attention from net toxicological impact.
• Structural isomerism did not significantly affect predicted toxicity.

Conclusions:

• The deep learning framework shows promise for predicting PFAS toxicity in data-scarce scenarios.
• Mechanistic insights enable quantitative understanding of structural contributions to toxicity.
• This approach facilitates the safety-by-design of next-generation fluorinated alternatives, moving beyond passive screening.