Lipid Class Prediction from MS1 Data using Gaussian Graphical Models

Liquid chromatography-mass spectrometry (LC-MS) untargeted analysis enables comprehensive lipid profiling of biological samples. However, system-level interpretation is often limited by the large number of unannotated features. Assigning features to lipid classes provides a higher-level, yet informative, overview that complements detailed structural analysis and supports biological interpretation at the class level. Recent advances in the systematic prediction of chemical class using tandem mass spectrometry (MS2) help address this; however, a substantial proportion of features in untargeted LC-MS data sets are typically characterized only at the MS1 level. Here, we present a workflow to systematically predict the lipid class from MS1-only data in untargeted LC-MS, without requiring prior annotations or MS2. Motivated by previous research showing that Gaussian graphical models (GGMs) estimated from feature intensities can encode the lipid class structure, our method, GgmLipidClassifier (GLC), combines conventional accurate-mass database searching with a GGM-derived network structure in a unified scoring framework to predict lipid class according to the LIPID MAPS Structure Database (LMSD) ontology. Across three human serum and plasma data sets, GLC achieved overall accuracies of 82-90% at the LMSD main class-level and 72-86% at the lipid subclass level, with improved accuracy and reduced uncertainty compared to closest-m/z matching. GLC provides class predictions for most detected features and also generates prediction quality scores to support downstream interpretation. Applied to serum samples from an Alzheimer's disease study, lipid class enrichment based on GLC predictions was highly consistent with class enrichment derived from ground-truth lipid annotations. Importantly, GLC extended coverage to classes missing from the annotation set, revealing biologically plausible associations with Alzheimer's disease, including cholesterol and derivatives, vitamin D₃ and derivatives, and plasmalogen glycerophosphoethanolamines. Overall, GLC provides robust lipid class predictions from MS1-only data, generating lipid class assignments for most detected features and complementing conventional analysis to support broader system-level interpretation.