Ambient and direct mass spectrometry for untargeted metabolomics: experimental design, preprocessing and data modelling

Ambient and direct mass spectrometry (MS) have become increasingly attractive for untargeted metabolomics due to their short analysis times, minimal sample preparation and capacity to generate rich spectral fingerprints. These advantages, however, introduce distinct analytical and computational challenges arising from matrix effects, high dimensionality, sparsity and instrumental drift. As a result, robust outcomes depend as much on experimental design and data handling as on advances in ionisation hardware. This review summarises practical considerations for planning ambient and direct MS metabolomics studies, including appropriate sample numbers, construction and placement of pooled quality-assurance samples, and strategies for monitoring and correcting analytical drift. We outline common approaches for feature alignment, sparsity reduction, filtering and preprocessing tailored to centroided high-resolution MS data. Linear chemometric models (Principal Components Analysis (PCA), Partial Least Squares – Discriminant Analysis (PLS-DA), Orthogonal Partial Least Squares -Discriminant Analysis (OPLS-DA), Least Absolute Shrinkage and Selection Operator (LASSO), Ridge Regression (RR) and Elastic Nets (EN)), non-linear methods and deep learning algorithms are compared with respect to validation, risk of overfitting and interpretability. Particular emphasis is placed on cross-validation, permutation testing and model-agnostic measures of feature importance such as Shapley values. The aim is to provide practical guidance for generating, processing and modelling ambient and direct MS datasets so that rapid analysis can be paired with rigorous statistical and chemometric practice.