Generative AI-Empowered Screening Strategy for Chemical Pollutants: A Case on Per- and Polyfluoroalkyl Substances

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• Identifying unknown chemical pollutants is crucial for environmental risk management.
• Existing analytical methods are limited by databases, missing many potential pollutants.
• Per- and polyfluoroalkyl substances (PFAS) represent a vast and incompletely characterized group of environmental contaminants.

Purpose of the Study:

• To develop and apply a generative artificial intelligence (AI) strategy for identifying unknown chemical pollutants.
• To expand the known chemical space of per- and polyfluoroalkyl substances (PFAS).
• To improve the high-throughput annotation of environmental pollutants.

Main Methods:

• Utilized chemical language models (CLMs) to generate novel PFAS structures.
• Created a suspect list of over 1.4 million generated PFAS structures.
• Applied the CLM-assisted screening strategy to environmental samples (wastewater influent and effluent).
• Integrated generated suspect lists with existing computational tools for pollutant analysis.

Main Results:

• Generated over 1.4 million new PFAS structures, increasing the known PFAS chemical space by 21.6%.
• Achieved 87% Top-1 annotation accuracy on spiked samples.
• Identified 88 previously overlooked PFAS features in a fluorochemical wastewater influent sample.
• Annotated 100 new PFAS features in an effluent sample by integrating the generated suspect list with existing tools.

Conclusions:

• Generative AI, specifically CLMs, offers a powerful approach for exploring vast chemical spaces and identifying unknown pollutants.
• The proposed CLM-assisted screening strategy significantly enhances the discovery of previously undetected PFAS in environmental matrices.
• This method provides a new avenue for high-throughput pollutant annotation, advancing environmental monitoring and risk assessment.