Gaining Insights Into Metabolic Networks Using Chemometrics and Bioinformatics: Chronic Kidney Disease as a Clinical Model

Abstract

Because of its ability to generate biological hypotheses, metabolomics offers an innovative and promising approach in many fields, including clinical research. However, collecting specimens in this setting can be difficult to standardize, especially when groups of patients with different degrees of disease severity are considered. In addition, despite major technological advances, it remains challenging to measure all the compounds defining the metabolic network of a biological system. In this context, the characterization of samples based on several analytical setups is now recognized as an efficient strategy to improve the coverage of metabolic complexity. For this purpose, chemometrics proposes efficient methods to reduce the dimensionality of these complex datasets spread over several matrices, allowing the integration of different sources or structures of metabolic information. Bioinformatics databases and query tools designed to describe and explore metabolic network models offer extremely useful solutions for the contextualization of potential biomarker subsets, enabling mechanistic hypotheses to be considered rather than simple associations. In this study, network principal component analysis was used to investigate samples collected from three cohorts of patients including multiple stages of chronic kidney disease. Metabolic profiles were measured using a combination of four analytical setups involving different separation modes in liquid chromatography coupled to high resolution mass spectrometry. Based on the chemometric model, specific patterns of metabolites, such as N-acetyl amino acids, could be associated with the different subgroups of patients. Further investigation of the metabolic signatures carried out using genome-scale network modeling confirmed both tryptophan metabolism and nucleotide interconversion as relevant pathways potentially associated with disease severity. Metabolic modules composed of chemically adjacent or close compounds of biological relevance were further investigated using carbon transfer reaction paths. Overall, the proposed integrative data analysis strategy allowed deeper insights into the metabolic routes associated with different groups of patients to be gained. Because of their complementary role in the knowledge discovery process, the association of chemometrics and bioinformatics in a common workflow is therefore shown as an efficient methodology to gain meaningful insights in a clinical context.

Keywords: bioinformatics; chemometrics; chronic kidney disease; integrative data analysis; metabolic networks; metabolomics.

FIGURE 1

Schematic diagram of the dataset structure composed of 40 data matrices involving the five different groups of individuals, the four blocks of variables, and repeated measurements. CTRL, control group; ICKD, intermediate chronic kidney disease; HD, hemodialysis; KG, kidney graft; DV, living kidney donors. Sharps indicate the numbering of the tables.

FIGURE 2

Circular layout graphs describing network connections between (A) groups of samples and (B) blocks of variables. The numbering of the data matrices corresponds to that given in Figure 1.

FIGURE 3

NetPCA score plot of the two first global components. CTRL, dark green crosses; ICKD, filled orange triangles; preHD, filled red squares; postHD, empty red squares; preKG, filled violet diamonds; postKG1, light violet diamonds; postKG2, empty violet diamonds; preDV, filled brown circles; postDV1, light brown circles; postDV2, empty brown circles.

FIGURE 4

Over-representation analysis (–log10 of the right-tailed Fisher test p-value corrected using the Benjamini-Hochberg False Discovery Rate procedure). The –log10 value of the 5% significance threshold is 1.301.

FIGURE 5

Pink pathway: Tryptophan metabolism. Blue pathway: Nucleotide interconversion (purine and pyrimidine metabolism). Metabolites related to the first NetPCA component and the corresponding subnetwork are shown in bold.

FIGURE 6

HCA dendrogram and heatmap of the metabolic fingerprint distance matrix, with the five clusters highlighted as relevant biological modules. The darker, the shorter reaction path.

FIGURE 7

Box-plots showing nine metabolites with representative alteration patterns from altered pathways.