Robust Feature Prioritization in LC-HRMS Non-Target Screening: A Multi-Method Workflow Applied to Municipal Wastewater

Abstract Feature prioritization is a critical step in high-dimensional liquid chromatography−high-resolution mass spectrometry (LC-HRMS)-based non-target screening (NTS), yet alternative statistical prioritization strategies have not been systematically compared in this context. We implemented a multi-method workflow combining robustness-aware prioritization, operationalized through within-method feature-selection stability assessment, with explicit evaluation of cross-method support, and applied it to 24 municipal wastewater influent samples collected from four wastewater treatment plants (WWTPs) across six sampling periods. After pre-processing, 3,636 ESI(+) and 2,053 ESI(−) features were evaluated. Principal component analysis (PCA) and self-organizing maps (SOM) revealed a pronounced exploratory temporal gradient with weaker site-related structure. Features were prioritized using partial least-squares-discriminant analysis/variable importance in projection (PLS-DA/VIP) with model population analysis (MPA), support vector machine-recursive feature elimination (SVM-RFE) with repeated stratified cross-validation, and false discovery rate (FDR)-adjusted univariate volcano filtering. The resulting prioritized sets showed limited and contrast-dependent overlap, demonstrating substantial method dependence. Cross-method support was summarized using consensus levels (CL1−CL3), representing selection by one, two, or three approaches, respectively. Temporal comparisons contained no CL3 features, whereas spatial contrasts showed greater convergence, reaching 50.0% CL3 at one WWTP in ESI(+). Design-preserving sample-composition and parameter-sensitivity analyses showed that the main agreement patterns were generally retained, although individual feature support varied. Association fingerprints with physicochemical parameters and quantified biomarkers also differed across consensus levels. Overall, combining feature-selection stability with cross-method comparison provides a transparent basis for prioritizing LC-HRMS NTS candidates for downstream identification and wastewater-based epidemiology interpretation.