The impact of gut microbiome enterotypes on ulcerative colitis: identifying key bacterial species and revealing species co-occurrence networks using machine learning

Figures & data

Table 1. Information on the studies included in the meta-analysis.

Study	Accession number	Design	Sample grouping	Age	Gender	Country	Sampling	DNA extraction	16S region	Seq Tech
Alam et al.	PRJEB33711	Case–control	UC: 11 Health: 10	NA	NA	UK	Collect feces in sterile tubes and immediately freeze at −80°C	FastDNA Spin Kit for Soil	V1–3	454 Sequencing
NIH HMP	PRJNA50637	Cross-sectional	UC: 601	NA	NA	USA	NA	NA	V1–3	454 Sequencing
NA	PRJDB6133	Case–control	UC:30 Health:23	NA	NA	JAPAN	NA	MiSeq 600 cycle v3 kit	V3–4	MiSeq
NA	PRJNA368966	Case–control	UC:26 Health:13	NA	NA	CHINA	NA	NA	V5-V6	MiSeq
NA	PRJNA296920	Case–control	Health: 362	NA	NA	USA	NA	NA	V5-V6	MiSeq
Weng et al.	PRJNA431126	Case–control	UC:107 Health:42	43.4±0.94	Male:97 Feaml:52	CHINA	Collect fecal samples with the help of a stool sampling kit and stored at home at 4 to 8°C for up to 24 hours	DNA Extraction Kits	V4	MiSeq
Maldonado‐Arriaga et al.	PRJNA596546	Case–control	UC:18 Health:15	37.1±1.36	Male:19 Female:14	Mexico	Collect the first-morning stool in a sterile stool collection tube and store the collected sample at −80°C	AllPrep PowerFecal DNA	V3-V6	MiSeq
Dai et al.	PRJNA681685	Case–control	UC:10 Health:14	30±0.88	Male:12 Femal:12	CHINA	Stool samples were collected in the Gastroenterology Department of Xijing Hospital and stored at −80°C	E.Z.N.A. stool DNA Kit	V3-V4	MiSeq
Wan et al.	PRJNA753210	Case–control	UC:29 Health:30	39 people>age 45	Male:39 Female:20	CHINA	Stool samples were collected in the Gastroenterology Department of Xijing Hospital and stored at −80°C	E.Z.N.A. stool DNA Kit	V3-V4	MiSeq
Cao et al.	PRJNA541040	Cross-sectional	UC:11	15–65	Male:7 Female:4	CHINA	NA	NA	NA	HiSeq 2500
Lloyd-Price J et al.	PRJNA398089	Case–control	UC:30 Health:26	28.8±4.04	Male:10 Female:45	USA	NA	NA	NA	HiSeq 2000
NA	PRJNA386260	Case–control	Health:211	55.3±1.39	Male:73 Female:138	USA	NA	NA	V4	Illumina

The table describes the design format in each study, the nationality of the study subjects, the amplified segment of 16s rRNA, and the system used for sequencing. Parts not provided are denoted by NA. Ulcerative colitis, UC; healthy control, HC.

Figure 1. Characterization and distribution of three enterotypes in UC and HC subjects based on the gut microbiome. Ulcerative colitis, UC; healthy control, HC; ET-B, enterotype Bacteridaceae; enterotype ET-L, Lachnospiraceae; ET-C, enterotype Clostridiaceae. (a) The principal component analysis (PCA) diagram of the three enterotypes was drawn based on the fecal microbes at the family level. (b) The number of UC and HC subjects in each enterotype. The Chi-square test was used to count the significant differences in the number of people among the various enterotypes. (c) The relative abundance of the top 6 family taxon in each enterotype, we employed Tukey’s post-hoc test to identify significant differences among the enterotypes. Taxa that exhibited significant differences were annotated with English letters for clarity. (d) The relative abundance of the top 6 genus taxon in each enterotype, we employed Tukey’s post-hoc test to identify significant differences among the enterotypes. Taxa that exhibited significant differences were annotated with English letters for clarity.

Figure 2. α-diversity by Shannon index and β-diversity by Bray–Curtis method. (a) α-diversity Shannon index in total participants. (b) α-diversity Shannon index in ET-B. (c) α-diversity Shannon index in ET-L. (d) β-diversity in total participants. (e) β-diversity in ET-B. (f) β-diversity in ET-L. Shannon index used the Wilcoxon test for significance analysis, and β-diversity used permutation multivariate analysis of variance (PERMANOVA) for significance. The numbers presented a significance level. Ulcerative colitis, UC; healthy control, HC; ET-B: Enterotype Bacteroidaceae cohort, ET-L: Enterotype Lachnospiraceae cohort; PCoA: Principal coordinate analysis.

Figure 3. The SHapley additive exPlanations (SHAP) interpreter and species co-occurrence network (SCN) of the enterotype Lachnospiraceae (ET-L) queue deep neural network (DNN). (a) The SHAP interpreter was used to conduct the microbial-specific importance analysis in the DNN classifier. The middle line is biased to the left for the healthy controls (HC) classification and vice versa for the ulcerative colitis (UC) classification. The color of the scatter points represents the influence of the relative abundance of the feature on the classification. The variables used to train the DNN network were species of microorganisms significantly different in the UC and HC groups. (b) Box plot of the 20 taxons included in the SHAP beeswarm plot and the numbers indicated the significance level in the corresponding Wilcoxon rank sum test. (c) The SCN was constructed using important gut microbes in the DNN classifier. The SHAP importance was used to determine the top 20 bacteria, and the network diagram was drawn. It was determined whether it belonged to UC or HC according to the SHAP bee colony diagram and the mean value of UC and HC. All connections with a sparse correlation for compositional data (SparCC) correlation coefficient less than 0.1 were removed in the SCN. The red edges represent positive correlation, the blue represents negative correlation, the thickness represents the size of the absolute correlation coefficient, the yellow node represents the HC group, the purple represents the UC group, and the node size represents the relative abundance.

Figure 4. The SHapley additive exPlanations (SHAP) interpreter and species co-occurrence network (SCN) of the enterotype Bacteroides (ET-B) queue deep neural network (DNN). (a) The SHAP interpreter was used to conduct the microbial-specific importance analysis in the DNN classifier. The middle line is biased to the left for the healthy controls classification (HC) and vice versa for the ulcerative colitis (UC) classification. The color of the scatter points represents the influence of the relative abundance of the feature on the classification. The variables used to train the DNN network were species of microorganisms significantly different in the UC and HC groups. (b) Box plot of the 20 taxons included in the SHAP swarm plot and the numbers indicated the significance level in the corresponding Wilcoxon rank sum test. (c) The SCN was constructed using important gut microbes in the DNN classifier. The SHAP importance was used to determine the top 20 bacteria, and the network diagram was drawn. It was determined whether it belonged to UC or HC according to the SHAP bee colony diagram and the mean value of UC and HC. All connections with a sparse correlation for compositional data (SparCC) correlation coefficient less than 0.1 were removed in the SCN. The red edges represent positive correlation, the blue represents negative correlation, the thickness represents the size of the absolute correlation coefficient, the yellow node represents the HC group, the purple represents the UC group, and the node size represents the relative abundance. (d) Linear discriminant analysis (LDA) scores of the HC and UC groups in LDA effect size (LEfSe). (e) Differential analysis of PICRUSt2-predicted gut microbiome function between UC and HC in the ET-B cohort was performed using LEfSe. The correlation analysis between the functions that LEfSe showed significant differences and the common pro-inflammatory functions and the important bacteria of HC and UC in the SHAP analysis was drawn into a heatmap.

Figure 4. (Continued).

Table 2. The relative abundance of ET-B group UC bacteria and HC bacteria in each enterotype.

	Gut microbes	ET-L(n=742)	ET-B(n=424)	ET-C(n=162)
UC	Bacteroides fragilis	0.49±0.23^b	5.12±0.51^a	0.55±0.26^b
	Desulfovibrio simplex	0.16±0.03^a	0.19±0.02^a	0.01±0^b
	Enterocloster clostridioformis	0.08±0.03^b	2.23±0.25^a	0.19±0.06^b
	Lachnoclostridium pacaense	0.55±0.08^a	0.33±0.05^b	0.12±0.06^c
	Phocaeicola dorei	0.53±0.14^b	5.25±0.55^a	0.13±0.05^b
	Ruminococcus gnavus	0.57±0.13^b	2.95±0.31^a	3.47±0.44^a
HC	Anaerostipes hadrus	2.77±0.33^a	0.51±0.06^b	0.08±0.04^b
	Bifidobacterium longum	1.01±0.28^a	0.07±0.02^b	0.79±0.26^a
	Blautia luti	1.07±0.19^a	0.19±0.03^b	0±0b
	Fusicatenibacter saccharivorans	2.26±0.28^a	0.23±0.04^b	0.13±0.07^b
	Odoribacter splanchnicus	0.17±0.04^a	0.07±0.01^b	0±0c
	Bacteroides uniformis	2.4±0.9^a	6.17±0.87^a	0.007±0.03^b
Clostridium	Clostridioides difficile	0±0	0.03±0.0^Citation3	0.07±0.03
	Clostridium butyricum	0.03±0.02^b	0.78±0.25^b	8.52±1.35^a
	Clostridium paraputrificum	0.04±0.04^b	0.4±0.07^b	1.65±0.3^a
	Clostridium perfringens	0.29±0.12^b	0.29±0.06^b	6.48±0.89^a
	Clostridium saudiense	0.03±0.02^b	0.24±0.08^b	3.4±0.69^a
	Clostridium tertium	0.12±0.08^b	0.07±0.04^b	0.7±0.16^a

^{a, b, c}Different superscript letters in each gut microbe indicated significant difference among the groups by Tukey post hoc test at P < 0.05.

The relative abundance of ulcerative colitis (UC) microbes and healthy control (HC) microbes obtained in the ET-B cohort in each enterotype disease group was compared.

Figure 5. Gut microbiota composition of volunteer subjects and results of in vitro experiments. (a) The relative abundance of the top 5 intestinal flora at the family level of 5 subjects. Subjects with a higher abundance of Bacteroidaceae and lower numbers of Lachnospiraceae and ruminococcaceae were classified as enterotype bacteridaceae (ET-B). Subjects with a higher abundance of Lachnospiraceae and ruminococcaceae were classified as enterotype Lachnospiraceae (ET-L). (b) Enumeration of C. difficile in ET-B and ET-L feces cultured in vitro. ETL-CD: ET-L enterotype feces co-cultured with C. difficile, ETL-Con: ET-L enterotype feces co-cultured with C. difficile, ETB-CD: ET-B enterotype feces co-cultured with C. difficile, ETB-Con: ET-B enterotype feces cultured alone. Statistical differences in Tukey’s test are expressed in letters. Different letters indicate significant differences between groups, and the same letters indicate no significant differences.

Figure 6. Workflow. Enterotype classification adopted by Arumugam et al.⁹ was used. ET-B: enterotype bacteridaceae, ET-L: enterotype Lachnospiraceae, ET-C: enterotype Clostridiaceae. Deep neural network (DNN). SHapley additive exPlanations (shap).

Choi CH, Moon W, Kim YS, Kim ES, Lee B-I, Jung Y, Yoon YS, Lee H, Park DI, Han DS, et al. Second Korean guidelines for the management of ulcerative colitis. Intest Res. 2017;15(1):7. doi:10.5217/ir.2017.15.1.7.

 PubMed Web of Science ®Google Scholar

Arumugam M, Raes J, Pelletier E, Le Paslier D, Yamada T, Mende DR, Fernandes GR, Tap J, Bruls T, Batto J-M, et al. Enterotypes of the human gut microbiome. Nature. 2011;473(7346):174–180. doi:10.1038/nature09944.

 PubMed Web of Science ®Google Scholar