Investigation of a Deep Learning-Assisted Workflow for the Analysis of Per-And Polyfluoroalkyl Substances in Environmental Matrices

Significance of the topic

Per- and polyfluoroalkyl substances (PFAS) are persistent organic pollutants detected at trace levels across water, soil, air, and food matrices. Their environmental persistence and potential health risks demand robust, high-throughput analytical workflows to ensure reliable monitoring and regulatory compliance.

Objectives and study overview

This study evaluates a deep learning-assisted workflow to streamline PFAS quantitation by LC-MS/MS in MRM mode. Two internally generated datasets following EPA Method 1633 and AOAC SMPR 2023.003 protocols were used to train convolutional and transformer-based models. The goal was to reduce manual peak integration steps while maintaining or improving accuracy.

Methodology and Instrumentation

• Sample preparation and acquisition following EPA 1633 and AOAC SMPR guidelines
• Instrumentation: Agilent Infinity III 1290 LC coupled to 6495D triple quadrupole MS
• Data preprocessing: retention time alignment, quantifier-qualifier correlation checks
• Conventional analysis: MassHunter Quantitative Analysis software with manual adjustment workflow
• Deep learning pipeline: custom preprocessing feeding into CNN and transformer architectures
• Model evaluation metrics: F1 score, positive predictive value (PPV), negative predictive value (NPV)

Main Results and Discussion

• AI integration flags identify peaks processed by built-in integrator, AI model, or manual intervention; confidence scores assess compound-level reliability
• Both CNN and transformer models stabilized above 0.95 for F1, PPV, and NPV after ~20 training epochs
• Deep learning-driven peak calls matched or outperformed manual integration for early-eluting PFAS, isomer separation, and matrix interferences
• Prediction time averaged 5–6 seconds per sample versus 60–120 seconds for manual review, enabling a >90% reduction in analysis time

Benefits and Practical Applications

• Significant reduction in manual data review and labor costs
• Enhanced consistency and reproducibility across varied instrument sensitivities and matrices
• Scalable implementation in routine environmental, food, and regulatory testing laboratories

Future Trends and Potential Applications

• Extension of deep learning integration to other analyte classes and separation techniques
• Real-time deployment in laboratory information management systems and cloud-based analytics platforms
• Adaptive model updates incorporating new matrices, instruments, and emerging contaminants

Conclusion

The developed deep learning-assisted workflow for PFAS LC-MS/MS analysis delivers rapid, accurate peak integration with minimal manual intervention. This approach enhances laboratory throughput, ensures high confidence in quantitation, and supports robust environmental and food safety monitoring.

References

1. United States Environmental Protection Agency. Method 1633: Analysis of PFAS in Aqueous, Solid, Biosolids, and Tissue Samples by LC-MS/MS, 4th draft, December 2024.
2. AOAC. Standard Method Performance Requirements (SMPR 2023.003) for PFAS in Produce, Beverages, Dairy, Eggs, Seafood, Meat, and Feed, 2023.
3. arXiv:1505.04597.
4. arXiv:1910.11162.
5. Perslev M. et al. npj Digit. Med. 4, 72 (2021).
6. arXiv:1803.01271.
7. arXiv:2105.15203.