Exploring non-target screening variability in unsupervised multivariate time trend analysis of LC-HRMS data.

Area of Science:

• Environmental Chemistry
• Analytical Chemistry
• Mass Spectrometry

Background:

• Non-target screening (NTS) with liquid chromatography-high-resolution mass spectrometry is crucial for identifying unknown contaminants in complex samples like industrial wastewater.
• Accurate detection and interpretation of temporal trends, particularly spill events, are key objectives in NTS applications.
• The performance of multivariate models in NTS is significantly influenced by the quality of feature lists generated by various software tools.

Purpose of the Study:

• To evaluate five peak picking tools (MarkerView, MZmine3, XCMS, OpenMS, SIRIUS) for unsupervised time trend exploration in industrial wastewater NTS data.
• To assess the impact of different software tools on feature list quality and subsequent multivariate analysis using sparse principal component analysis (SPCA).
• To enhance the reliability of time trend detection by combining SPCA with stratified bootstrapping (SBS-SPCA).

Main Methods:

• Utilized sparse principal component analysis (SPCA) for unsupervised time trend exploration, focusing on its ability to select informative features and improve model interpretability.
• Employed two datasets: a controlled validation set with spiked compounds and a real-world dataset of 52 daily industrial wastewater samples.
• Implemented stratified bootstrapping with SPCA (SBS-SPCA) to evaluate the robustness of detected temporal trends.

Main Results:

• SPCA effectively distinguished spiking patterns in the validation set, highlighting tool-specific differences in feature/artifact prioritization.
• XCMS, MZmine3, and OpenMS demonstrated higher consistency and were selected for further analysis.
• Under optimized conditions, five out of nine persistent markers were robustly detected across the selected tools using SBS-SPCA (selection frequency > 70%).

Conclusions:

• Interpretable, sparse models enhance marker detection in unsupervised NTS, particularly for time-series data.
• The choice of peak picking software significantly impacts the structure of feature lists and the outcomes of multivariate analyses.
• These findings are vital for advancing high-throughput NTS applications in temporally dynamic environmental exposure scenarios.