Machine Learning–Enhanced LC–HRMS Workflows for Suspect and Nontargeted Screening of Pesticide Residues in Fruits and Vegetables

Abstract

Analytical methodologies are essential for detecting and quantifying contaminants. While methods for known compounds are well established, identifying unexpected or unknown compounds remains challenging due to the lack of reference data. Several strategies have been proposed to integrate analytical information into data analysis workflows, but their implementation often requires programming skills and results are rarely presented in a format familiar to analytical chemists, such as the uncertainty budgets used in quantitative analysis. We developed an integrated workflow for suspect screening and nontargeted identification of pesticides in fruits and vegetables using QuEChERS extraction and HPLC-ESI-HRMS (Orbitrap) analysis. The workflow was developed using variable data-independent acquisition (vDIA) data and evaluates protonated and deprotonated species for compound identification. The workflow estimates missing identification properties using machine learning and combines their contributions into a single confidence value. Validation with fortified matrix extracts at different mass concentrations and with real samples containing pesticide residues showed performance comparable to existing software, particularly improving nontargeted identification at high concentrations (>6% increase in identified compounds). This approach reduces the input required for suspect screening and estimates properties for candidate compounds in nontargeted analysis. Results are reported with explicit consideration of the influence of identification properties, analogous to uncertainty budgets in chemical metrology. This workflow improves the interpretability and reliability of chemical identification and supports data-driven decision-making in routine analysis.