High-throughput non-targeted PFAS detection at scale: Algorithm optimization, benchmarking, and automated workflows for environmental monitoring

Per- and polyfluoroalkyl substances (PFAS) represent a structurally diverse class of synthetic chemicals that resist environmental degradation and accumulate in water, soil, and biological tissues, posing well-documented risks to human health and ecosystems. Targeted analytical methods, which screen only for a predefined list of regulated compounds, systematically miss the broader PFAS chemical space and fail to detect emerging, uncharacterized contaminants. This paper presents the design, implementation, and evaluation of a high-throughput non-targeted PFAS detection framework that integrates high-resolution mass spectrometry (HRMS) with an automated, algorithm-optimized data processing pipeline. The framework was applied to a dataset of approximately 1 TB of LC-HRMS data drawn from water, soil, and biota matrices, processing over 2,400 environmental samples. Three core technical contributions are reported. First, a stochastic parameter optimization protocol for feature detection, modeled on the approach of Sadia et al. (2024), reduced false-positive rates by 38% relative to default XCMS settings while recovering 94% of true PFAS signals in spiked validation samples. Second, parallel cloud-based processing reduced per-sample computational time from approximately 47 minutes to 6.2 minutes, enabling near-real-time throughput across the full dataset. Third, automated feature prioritization using the PFΔScreen scoring scheme (Zweigle et al., 2024) combined with a trained Random Forest classifier achieved a positive predictive value of 91% on a held-out test set of 380 samples. Benchmarking against parallel targeted analysis demonstrated a 2.3-fold increase in the number of PFAS features detected per sample, including 47 tentatively identified compounds absent from current EPA regulatory lists. Intra-laboratory repeatability, assessed across triplicate injections, yielded a median coefficient of variation of 4.7% for peak area, confirming analytical consistency. These results establish the framework as a scalable, reliable tool for large-scale environmental surveillance, regulatory decision-support, and emerging-contaminant discovery.