doi: 10.1016/j.isci.2022.105868.
eCollection 2023 Jan 20.
Quantitative analysis of disease-related metabolic dysregulation of human microbiota
Affiliations
- PMID: 36624837
- PMCID: PMC9823209
- DOI: 10.1016/j.isci.2022.105868
Item in Clipboard
Quantitative analysis of disease-related metabolic dysregulation of human microbiota
iScience.
.
Abstract
The metabolic activity of all the micro-organism composing the human microbiome interacts with the host metabolism contributing to human health and disease in a way that is not fully understood. Here, we introduce STELLA, a computational method to derive the spectrum of metabolites associated with the microbiome of an individual. STELLA integrates known information on metabolic pathways associated with each bacterial species and extracts from these the list of metabolic products of each singular reaction by means of automatic text analysis. By comparing the result obtained on a single subject with the metabolic profile data of a control set of healthy subjects, we are able to identify individual metabolic alterations. To illustrate the method, we present applications to autism spectrum disorder and multiple sclerosis.
Keywords: Computational bioinformatics; Microbiome.
© 2022 The Author(s).
Conflict of interest statement
The authors declare no competing interests.
Figures
Schematic of the algorithm The OTU abundances matrix obtained from microbiota sequencing experiments are combined with information on the metabolic pathways present in each OTU, retrieved from the Macadam database, and with the stoichiometry of metabolites produced and consumed in each pathway, obtained from the Metacyc database. All the information is combined to obtain a metabolite-patient matrix for each experimental dataset.
Algorithm validation (A) Heatmap of the -scores of the metabolites recorded in and (B) the corresponding predictions made by the STELLA algorithm. (C) Confusion matrix for the algorithm predictions.
Batch effect removal (A and B) Correlation matrix of OTUs abundances obtained merging the datasets in the study by Kang et al. and Kong et al. before (A) and after (B) batch effect removal. Hierarchical clustering on correlation matrix shows that differences between the two datasets (gray, red) is predominant in the merged dataset, and the two major blocks disappear after batch removal. Hierachical clustering does not allow us to distinguish between patients and healthy controls (yellow, blue) suggesting a more elaborated procedure is needed. (C and D) Panels show the data from the intersection of data from the study by Kang et al. and Kong et al. projected onto the first two principal components and divided into healthy controls (red) and autistic patients (blue) before (C) and after (D) one-step batch removal. In panel (C), the variance between the data from different datasets is higher than the distance between the autistic patients and the healthy controls. Different symbols represent the two studies (circles, triangles as in legend).
LDA algorithm (A) Schematic of the LDA algorithm. After removal of highly correlated OTUs and metabolites, LDA feature selection is applied in order to obtain an optimal set of OTUs and metabolites to discriminate between patients (yellow circles) and healthy controls (blue circles). (B) Heatmaps show the ensemble of selected OTUs and metabolites for each replica of LDA feature selection for the combined datasets in the study by Kang et al. and Kong et al. after batch removal. The presence of a specific OTU or metabolite in a given replica is represented in blue and its absence in white and ordered using hierarchical clustering. (C and D) Performance of the models obtained from LDA feature selection. Correctness rate is evaluated after the inclusion of each new (C) OTU or (D) metabolite. The plots show, for each replica and in chronological order of selection, the increasing correctness rate of the model after the addition of every taxonomy.
LDA feature selection results for ASD Panel (A) show heatmaps representing the selected taxonomies for each replica of LDA feature selection reordered using hierarchical clustering and co-occurrence matrix representing the number of replica two compounds are selected as significant by LDA. The presence of a specific taxonomy in a given iteration is represented in blue and its absence in white. Panels (B and C) Illustrative reconstruction of reverse metabolic network. Plot shows the five most frequently selected metabolites during LDA analysis of metabolite-host matrix and the pathways in which they are mapped (B). For a specific metabolite (2-methyladenine), we report the obtained pathway-compound network (C). The network includes 2-methyladenine and its correlated metabolites (r1-r47) and the related pathways (p1-p35). The complete list of labels is reported Data S1. Lines represent the links between the metabolites (green, light green) and the pathways (gray). For all the panels, data are reported for the combined datasets from the study by Kang et al. and Kong et al. after batch removal.
LDA feature selection result for MS (A) Heatmaps representing the selected taxonomies for each replica of LDA feature selection reordered using hierarchical clustering and co-occurrence matrix representing the number of replica two compounds are selected as significant by LDA. The presence of a specific taxonomy in a given iteration is represented in blue and its absence in white. (B) Illustrative reconstruction of reverse metabolic network. Plot shows the five most frequently selected metabolites during LDA analysis of metabolite-host matrix and the pathways in which they are mapped. Lines represent the links between the metabolites (green, light green) and the pathways (gray) For all the panels, data are reported for the dataset in the study by Cekanaviciute et al. Metabolites names are reported in the figure, while full pathways names are only reported in Data S1.
