Correlation Between Fecal Microbiota and Corticosteroid Responsiveness in Primary Immune Thrombocytopenia: an Exploratory Study

Abstract

Corticosteroids (CSs) are the initial therapy for immune thrombocytopenia (ITP); however, their efficacy is not adequately predicted. As a novel biomarker, the composition of the gut microbiota is non-invasively tested and altered in patients with ITP. This study aims to develop a predictive model that leverages gut microbiome data to predict the CS response in patients with ITP within the initial four weeks of treatment. Metagenomic sequencing is performed on fecal samples from 212 patients with ITP, 152 of whom underwent CS treatment and follow-up. Predictive models are trained using six machine-learning algorithms, integrating clinical indices and gut microbiome data. The support vector machine (SVM) algorithm-based model has the highest accuracy (AUC = 0.80). This model utilized a comprehensive feature set that combined clinical data (including sex, age, duration, platelet count, and bleeding scales) with selected microbial species (including Bacteroides ovatus, Bacteroides xylanisolvens, and Parabacteroides gordonii), alpha diversities, KEGG pathways, and microbial modules. This study will provide new ideas for the prediction of clinical CS efficacy, enabling informed decision-making regarding the initiation of CS or personalized treatment in patients with ITP.

Keywords: biomarker; gut microbiome; gut microbiota; immune thrombocytopenia; machine learning.

Figure 1

Flow diagram of study enrolment, sample collection, follow‐up, and data analysis. To evaluate the association between the gut microbiota and clinical indices and corticosteroid response, we profiled stool metagenomes of patients with primary ITP and healthy volunteers. Individuals were excluded if they were treated with any medication for thrombocytopenia in the previous 6 months; or exposed to antibiotics, prebiotics, or probiotics within 4 weeks before fecal sampling. In 2018 and 2021, individuals including patients with ITP (N = 142) and healthy volunteers (N = 62) were enrolled in a single center (cohort 1 and cohort 2) in Beijing, China. During 2021–2022, 70 participants with ITP were enrolled in the multicenter cohort (cohort 3). Patients with ITP subsequently received personalized treatment under clinical supervision, among which a total of 152 participants proceeded with standard corticosteroid therapy, completing an efficacy evaluation within four weeks post‐treatment initiation. Upon enrollment, all participants submitted baseline stool samples (ITP, N = 212; HC, N = 62), and 46 participants with ITP provided stool samples after the fourth week following the start of corticosteroid therapy. This study used all baseline sample data to conduct cross‐sectional descriptive studies and correlation analysis with population characteristics and clinical metrics. Key clinical indicators included the duration from initial ITP diagnosis at enrollment, WHO bleeding score, complete blood count parameters, and medication history over the previous six months. For the 46 paired samples before and after corticosteroid treatment, the interaction between corticosteroids and the gut microbiota was explored through microbial abundance and network analyses. Machine learning techniques were utilized for the 152 patients assessable for corticosteroid efficacy, merging baseline fecal microbiome data with clinical indicators to predict treatment outcomes. The training set included 100 samples from two cohorts in Beijing, while the testing set comprised 52 samples from the cohort across 20 centers. The predictive model, which integrates microbial taxonomy, functional components, and clinical indicators, offers valuable assistance in clinical treatment decisions for ITP. Abbreviations: ITP, immune thrombocytopenia; HC, healthy controls; CBC, complete blood count; WBC, white blood cells; WHO, World Health Organization; STAMP, statistical analysis of metagenomics profile.

Figure 2

Gut microbiota profiles associated with the clinical indicators of ITP. a) The PCoA of β‐diversity based on order distribution by Bray‒Curtis dissimilarity in the HC group (n = 62), as well as the ND/persistent (n = 108) and chronic (n = 104) ITP groups. b) Comparison of the α‐diversity measured in the Chao1, Dominance, and Shannon index at the order level among the HC group (n = 62), as well as the ND/persistent (n = 108) and chronic (n = 104) ITP groups. Each box represents the IQR with the midpoint of the data. Whiskers indicate the upper and lower values within 1.5 times the IQR. c) The PCoA of β‐diversity based on order distribution by Bray‒Curtis dissimilarity in HC (n = 62), N‐CS (n = 97), and Y‐CS (n = 115) groups. d) Comparison of the α‐diversity measured with the Chao1, Dominance, and Shannon index at the order level among the HC (n = 62), N‐CS (n = 97), and Y‐CS (n = 115) ITP groups. e) Significant associations between gut microbial taxa and ITP grouping at the species level analyzed by MaAsLin2. P‐values are calculated using Kruskal–Wallis test, followed by the Dunn post‐hoc test across groups, ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001. Data from the figures was presented in Tables S5−S9 (Supporting Information). Abbreviations: CS, corticosteroid; ITP, immune thrombocytopenia; HC, healthy controls; PCoA, principal co‐ordinates analysis; LDA, Linear discriminant analysis; LEfSe, LDA Effect Size; IQR, interquartile range.

Figure 3

Gut microbiota variations throughout CS therapy. a) The relative abundance of predominant genera in the stool samples from patients with ITP (n = 46), collected both before (group PRE) and after (group POST) CS treatment. b) Gut microbial community networks constructed from genera‐level analysis of stool samples, from PRE and POST ITP groups that received CS treatment. A correlation coefficient >0.7 with statistically significant (p < 0.01) connections is shown (positive correlation, red edge; negative correlation, blue edge). Each node represents a genus, node size represents the relative abundance of the genus, and color represents the affiliated phylum. c) Relative abundance of the 20 species most associated with CS therapy between PRE and POST groups in STAMP analysis. P‐values are calculated using paired Wilcoxon test, ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001. d) The differential gut microbial species in the pre‐ and post‐CS treatment samples of ITP. PRE, n = 46; POST, n = 46. The four listed species were validated using both paired Wilcoxon and MaAsLin analysis. Each box represents the interquartile range (IQR, the range between the 25th and 75th percentiles) of the relative abundance with the mid‐point of the data. Dotted lines connect the points corresponding to samples collected before and after treatment from the same patient. (data was shown in Table S11, Supporting Information) Abbreviations: CS, corticosteroid; ITP, immune thrombocytopenia; STAMP, Statistical Analysis of Metagenomics Profile; MaAsLin, Multivariable Association with Linear Models.

Figure 4

The baseline gut microbiota composition of patients with ITP was correlated with CS response. a) PCoA of β‐diversity based on order distribution by Bray‒Curtis distance between responders (group R, n = 71) and non‐responders (group NR, n = 29) to CS therapy. Adonis analysis was used to test the statistical significance of β dissimilarities between groups. b) Comparison of the α‐diversity measured with the Chao1, dominance, and Shannon indices at the order level between the R (n = 71) and NR (n = 29) groups. Each box represents the interquartile range (IQR, the range between the 25th and 75th percentiles) of the relative abundance with the mid‐point of the data. P‐values are calculated using Mann–Whitney U test, ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001. c) The overall composition and relative abundance of the bacterial community at the phylum level between the R (n = 71) and NR (n = 29) groups. d) Circos plot of the distribution of abundant genera in the R (n = 71) and NR (n = 29) groups. Groups are represented above, genera are represented below, chords connect different genera and samples, and the outer circle indicates the relative abundance of the genus. e) Gut microbial community networks based on genera in the R (n = 71) and NR (n = 29) groups, constructed by using Spearman correlations. A correlation coefficient >0.7 with statistically significant (p < 0.01) connections is shown (positive correlation, red edge; negative correlation, blue edge). Each node represents a genus, node size represents the relative abundance of the genus, and color represents the affiliated phylum. f) Volcano plot of the distribution of all differentially enriched species between the R (n = 71) and NR (n = 29) groups. g) Box plots of the top 10 species differentially enriched between the R (n = 71) and NR (n = 29) groups at baseline. Each box represents the IQR with the midpoint of the data. Whiskers indicate the upper and lower values within 1.5 times the IQR. P‐values are calculated using Mann–Whitney U test, ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001. Data from the figures was presented in Tables S13−S15 (Supporting Information). Abbreviations: ITP, immune thrombocytopenia; CS, corticosteroids; PCoA, principal coordinates analysis; R, responders; NR, non‐responders; IQR, interquartile range.

Figure 5

Functional alterations in the fecal metagenome of ITP were related to the CS response. a) Principal component analysis (PCA) of sample distribution based on KEGG gene annotation between responders (group R, n = 71) and non‐responders (group NR, n = 29) to CS therapy. Adonis analysis was used to test the statistical significance of β dissimilarities between groups. b) Dot plot of differentially annotated KEGG pathways. The dot size represents the log₂FoldChange in the R group, and the color scheme represents p values. c) Box plots of the top 10 modules differentially enriched between the R (n = 71) and NR (n = 29) groups at baseline. Each box represents the IQR with the midpoint of the data. Whiskers indicate the upper and lower values within 1.5 times the IQR. P‐values are calculated using Mann–Whitney U test, ^* p < 0.05, ^** p < 0.01, ^*** p < 0.001. Data from the figures was presented in Tables S16−S19 (Supporting Information). Abbreviations: ITP, immune thrombocytopenia; CS, corticosteroids; R, responders; NR, non‐responders; PCA, principal component analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; FC, fold change; IQR, interquartile range.

Figure 6

Predictive models for CS response in patients with ITP based on taxonomic and functional components. a) Machine learning framework for predicting CS response: training set, n = 100; testing set, n = 52. b−c) AUROC, PR AUC, and MCC values for CS response based on a combination of clinical indices, alpha diversities, taxonomic components, and functional components in the training b) and testing c) sets using six models. Abbreviations: ITP, immune thrombocytopenia; CS, corticosteroids; AUC, the area under the curve; SVM, support vector machines; AUROC, receiver operating characteristic curve; PR AUC, the precision‐recall area under the curve.