Exploratory analysis of metabolic changes using mass spectrometry data and graph embeddings

Abstract

Mass spectrometry (MS)-based metabolomics analysis is a powerful tool, but it comes with its own set of challenges. The MS workflow involves multiple steps before its interpretation in what is denominate data mining. Data mining consists of a two-step process. First, the MS data is ordered, arranged, and presented for filtering before being analyzed. Second, the filtered and reduced data are analyzed using statistics to remove further variability. This holds true particularly for MS-based untargeted metabolomics studies, which focused on understanding fold changes in metabolic networks. Since the task of filtering and identifying changes from a large dataset is challenging, automated techniques for mining untargeted MS-based metabolomic data are needed. The traditional statistics-based approach tends to overfilter raw data, which may result in the removal of relevant data and lead to the identification of fewer metabolomic changes. This limitation of the traditional approach underscores the need for a new method. In this work, we present a novel deep learning approach using node embeddings (powered by GNNs), edge embeddings, and anomaly detection algorithm to analyze the data generated by mass spectrometry (MS)-based metabolomics called GEMNA (Graph Embedding-based Metabolomics Network Analysis), for example for an untargeted volatile study on Mentos candy, the data clusters produced by GEMNA were better than the ones used traditional tools, i.e., GEMNA has [Formula: see text], vs. the traditional approach has [Formula: see text].

Keywords: Graph embeddings; Graph neural networks; Mass spectrometry; Metabolomic networks.

Fig. 1

Performance of biological repetitions (Br) and analytical repetitions (Ar). The red dotted line represents the number of common edges found by a greedy algorithm, which does not eliminate edges considered as noise (outlier). The solid grey line represents the number of common edges based on the inverse of the covariance matrix (Precision matrix).

Fig. 2

Number of common edges in filtering networks.

Fig. 3

MAD values on Mentos dataset.

Fig. 4

Runtimes to generate embeddings on Synthetic networks.

Fig. 5

Similarity analysis on Mutant network.

Fig. 6

Heatmap on Leaf filter data.

Fig. 7

PCA on Mentos filter data workflow.

Fig. 8

Mass spectrometry data details. (a) is Alignment ID, is Retention time, is Average Mz, is Metabolite name, , , ..., are Biological repetitions. In addition, each biological repetition has two analytical repetitions at least.

Fig. 9

Pipeline 1: Network generation. Note: none means, network without variation.

Fig. 10

Pipeline 2: Network filtering. Note: The edges in red color have positive correlation and the edges in blue color have negative correlation.

Fig. 11

Pipeline 3: Similarity analysis. Note: the “?” mark means that there is no correlation.

Fig. 12

Our method with a toy example. Note: , , ..., are Biological repetitions.