Evaluation of data analysis pipelines for ambient ionization mass spectrometry: Bridging analytical chemistry and...

Ambient Ionization Mass Spectrometry (AIMS) enables rapid, high-throughput chemical analysis with minimal sample preparation, but produces complex, information-rich spectra that require robust preprocessing and multivariate analysis. Numerous normalization, scaling, and transformation strategies are routinely applied, yet their selection is often empirical and poorly justified. Using edible oil authentication and type-determination as a model system, this work systematically evaluates ten preprocessing strategies combined with multiple dimensionality reduction and classification methods. The problem addressed is the lack of a quantitative, evidence-based framework for selecting optimal preprocessing and analysis workflows for AIMS data.

Ten preprocessing approaches were quantitatively assessed in combination with Principal Component Analysis (PCA), sparse PCA (sPCA), and Partial Least Squares Discriminant Analysis (PLS-DA), followed by Naïve Bayes, Support Vector Machines (SVM), and Linear Discriminant Analysis (LDA) classifiers. Performance was evaluated using explained variance, clustering index, projected chromatographic resolution (Rs), and classification accuracy. No universally optimal workflow was identified; instead, performance depended strongly on data characteristics. Log₂ transformation consistently outperformed alternative preprocessing methods, yielding superior cluster separation and classification accuracy for this dataset. PCA and PLS-DA performed comparably and outperformed sPCA, while LDA achieved the highest classification accuracy. In contrast, commonly used Total Ion Count (TIC) normalization degraded performance and, in some cases, produced misleading classifications. Two visualization tools are introduced: an importance spectrum highlighting discriminative m/z features, and a synthetic chromatographic projection enabling intuitive interpretation of multivariate separability. Although demonstrated using LMJ-SSP AIMS data, the framework is broadly applicable to conventional MS workflows.

This study provides the first comprehensive, quantitative framework for optimizing preprocessing, dimensionality reduction, and classification for AIMS data analysis. By demonstrating that commonly used methods can underperform or mislead, it shifts workflow selection from convention to evidence. The proposed metrics and visualization tools improve interpretability, rigor, and reproducibility, enabling more reliable classification and feature discovery in mass spectrometry-based research.