Coupling Mixed Mode Chromatography/ESI Negative MS Detection with Message-Passing Neural Network Modeling for Enhanced Metabolome Coverage and Structural Identification

Abstract

A key unmet need in metabolomics continues to be the specific, selective, accurate detection of traditionally difficult to retain molecules including simple sugars, sugar phosphates, carboxylic acids, and related amino acids. Designed to retain the metabolites of central carbon metabolism, this Mixed Mode (MM) chromatography applies varied pH, salt concentration and organic content to a positively charged quaternary amine polyvinyl alcohol stationary phase. This MM method is capable of separating glucose from fructose, and four hexose monophosphates a single chromatographic run. Coupled to a QExactive Orbitrap Mass Spectrometer with negative ESI, linearity, LLOD, %CV, and mass accuracy were assessed using 33 metabolite standards. The standards were linear on average >3 orders of magnitude (R2 > 0.98 for 30/33) with LLOD < 1 pmole (26/33), median CV of 12% over two weeks, and median mass accuracy of 0.49 ppm. To assess the breadth of metabolome coverage and better define the structural elements dictating elution, we injected 607 unique metabolites and determined that 398 are well retained. We then split the dataset of 398 documented RTs into training and test sets and trained a message-passing neural network (MPNN) to predict RT from a featurized heavy atom connectivity graph. Unlike traditional QSAR methods that utilize hand-crafted descriptors or pre-defined structural keys, the MPNN aggregates atomic features across the molecular graph and learns to identify molecular subgraphs that are correlated with variations in RTs. For sugars, sugar phosphates, carboxylic acids, and isomers, the model achieves a predictive RT error of <2 min on 91%, 50%, 77%, and 72% of held-out compounds from these subsets, with overall root mean square errors of 0.11, 0.34, 0.18, and 0.53 min, respectively. The model was then applied to rank order metabolite IDs for molecular features altered by GLS2 knockout in mouse primary hepatocytes.

Retention Time Prediction through Learning from a Small Training Data Set with a Pretrained Graph Neural Network

“…A GNN can directly process the graph representation. , Previous studies have attempted to build GNNs as RT prediction models to achieve an improved predictive performance. Various GNN architectures have been evaluated for their suitability for RT prediction, including graph convolutional network (GCN) and its variants, , message passing neural network (MPNN), , and many others …”

“…25,26 Previous studies have attempted to build GNNs as RT prediction models to achieve an improved predictive performance. Various GNN architectures have been evaluated for their suitability for RT prediction, including graph convolutional network (GCN) 16 and its variants, 17,18 message passing neural network (MPNN), 19,20 and many others. 21 In practice, the predictive performance of a prediction model primarily depends on various factors, including the quantity and diversity of the training data set and the level of difficulty of the target prediction task.…”

“…Machine learning approaches have been adopted to accurately predict the RT of a molecule in a data-driven manner. ,,− These approaches build a prediction model that uses the representation of a molecule as the input to predict the RT. The model learns from a training data set containing a number of molecules and their RT measurements according to a target chromatographic system.…”