Establishing high-accuracy biomarkers for colorectal cancer by comparing fecal microbiomes in patients with healthy families

ABSTRACT

Colorectal cancer (CRC) causes high morbidity and mortality worldwide, and noninvasive gut microbiome (GM) biomarkers are promising for early CRC diagnosis. However, the GM varies significantly based on ethnicity, diet and living environment, suggesting varied GM biomarker performance in different regions. We performed a metagenomic association analysis on stools from 52 patients and 55 corresponding healthy family members who lived together to identify GM biomarkers for CRC in Chongqing, China. The GM of patients differed significantly from that of healthy controls. A total of 22 microbial genes were included as screening biomarkers with high accuracy in additional 46 cases and 40 randomly selected healthy adults in Chongqing (area under the receive-operation curve (AUC) = 0.905, 95% CI 0.832–0.977). The classifier based on the identified 22 biomarkers also performed well in the cohort from Hong Kong (AUC = 0.811, 95% CI 0.715–0.907) and French (AUC = 0.859, 95% CI 0.773–0.944) populations. Quantitative PCR was applied for measuring three selected biomarkers in the classification of CRC patients in independent Chongqing population containing 30 cases and 30 controls and the best biomarker from Coprobacillus performed well with high AUC (0.930, 95% CI 0.904–0.955). This study revealed increased sensitivity and applicability of our GM biomarkers compared with previous biomarkers significantly promoting the early diagnosis of CRC.

Introduction

Colorectal cancer (CRC) is the third most frequently diagnosed cancer in men and the second in women.^1,2 In China, the number of new CRC patients was estimated to be 376,300, and the CRC death toll was 191,000 in 2015.³ The American Cancer Society recommends a colonoscopy every 10 years after age 45 for the early diagnosis of colorectal diseases, especially CRC.⁴ However, colonoscopy is not widely acceptable in common health investigations given its invasive nature. Noninvasive tests, such as liquid biopsy and fecal immunochemical test, are promising for CRC screening but have limited detection precision and prediction accuracy.⁵

Emerging reports have demonstrated dysbiosis of gut microbiome (GM) in CRC patients and the potential of GM biomarkers in CRC screening.^6,7 The imbalanced GM includedFusobacterium nucleatum, Peptostreptococcus stomatis, Parvimonas micra, and Solobacterium moorei, which have been validated as biomarkers for CRC screening in several studies.^8–11 Microbial genes from the GM have also been increasingly accepted as early diagnostic CRC biomarkers.^5,12 However, dietary habits and exterior environments can impose long-term effects on GM configuration.¹³ Thus, we hypothesized that GM biomarkers in CRC screening differ based on region given the decreasing accuracy observed when biomarkers from Hong Kong (HK) population were applied to French and Austrian populations.⁸ Additionally, we hypothesized that high-accuracy CRC screening could be achieved using GM biomarkers that were identified by comparing the fecal microbiome of CRC patients with that of healthy family members.

To identify and investigate GM biomarkers for CRC screening in Chongqing (CQ), China, and test our hypothesis, we collected 107 fecal samples from 52 CRC patients and 55 corresponding healthy adult family members who lived together for at least 1 year. Metagenomic analysis was conducted to explore GM disequilibrium in CRC patients and associated biomarkers. We also assessed the performance of our GM biomarkers in additional 86 fecal samples in CQ and 269 microbial samples in different populations. Additionally, three biomarkers were validated with quantitative PCR (qPCR) in an independent Chongqing population with randomly selected 30 patients and 30 healthy controls.

Results

Sample information

Fecal samples from 193 Chinese subjects, including 98 CRC patients (68 males and 30 females) and 95 healthy controls (49 males and 46 females), were collected at the Southwest Hospital in CQ (Table 1). Of the recruited subjects, 52 cases and their 55 healthy family members were included to identify GM biomarkers (collectively regarded as Family cohort, Table S1). The remaining 86 fecal samples were selected randomly as the validation cohort (Table S1). Additionally, independent 30 CRC patients and 30 healthy controls were recruited from the same hospital for qPCR assessment (Table S2). We also collected published metagenomics data from 75 CRC patients and 53 healthy controls in HK (PRJEB10878) and 53 CRC patients and 88 healthy individuals in France (ERP005534).^8,14

Table 1.

Sample information.

	Family cohort			Validation cohort
	CRC (n = 52)	Controls (n = 55)	Adjusted P	CRC (n = 46)	Controls (n = 40)	Adjusted P	Overall (n = 193)
Gender (male:female)	35:17	26:29	0.08	33:13	23:17	0.23	117:76
Ages (years)	53 (33, 74)	42 (20, 72)	< 0.001	59 (28, 76)	42.5 (22, 65)	< 0.001	48 (20, 76)
BMI (kg/m2)	23.49 (17.51, 28.73)	23.21 (19.87, 27.6)	0.66	22.5 (17.92, 30.43)	22.82 (19, 28.27)	0.53	23.06 (17.51, 30.43)
GGT (IU/L)	21 (8, 71)	20 (11, 38)	0.55	25 (11, 166)	29.5 (11, 55)	0.56	22 (8, 166)
AST (IU/L)	20.6 (7.9, 47.9)	15.8 (9.2, 22.4)	< 0.001	20.8 (10, 105.8)	17.4 (9.3, 23)	0.03	17.95 (7.9, 105.8)
ALT (IU/L)	17.1 (3.7, 69.4)	15.5 (9, 22.9)	0.66	15.2 (1.1, 98.5)	15.5 (9, 23)	0.51	15.6 (1.1, 98.5)
ALB (g/L)	40.4 (27.8, 52.4)	41.9 (37.1, 49.9)	0.16	41.85 (31.1, 51.5)	42.4 (37, 49.8)	0.23	41.85 (27.8, 52.4)
Fasting glucose (mmol/L)	5.49 (4.16, 8.12)	4.65 (3.93, 6.08)	< 0.001	5.51 (3.9, 13.32)	5.31 (4.02, 6.06)	0.06	5.29 (3.9, 13.32)
HGB (g/L)	127 (65, 177)	135 (108, 160)	0.09	132 (75, 159)	142 (108, 160)	0.03	133 (65, 177)
Cr (mmol/L)	71 (40.5, 115)	79.7 (59.1, 102.5)	0.04	71.2 (33, 102.1)	80.9 (59.1, 102.8)	0.005	74.65 (33, 115)
CEA (ng/mL)	3.28 (0.32, 178.36)	-	-	2.84 (0.68, 397.6)	-	-	2.96 (0.32, 397.6)
AJCC stage (I:II:III:IV)	13:18:16:5	-	-	10:18:15:3	-	-	23:36:31:8
FOBT (positive:negative)	46:6	2:53	< 0.001	31:15	02:38	< 0.001	81:112
Localization (rectum:colon)	29:23	-	-	29:17	-	-	58:40

“-” represents no detection result; BMI, body mass index; GGT, γ-glutamyl transpeptidase; AST, aspartate aminotransferase; ALT, serum alanine aminotransferase; ALB, albumin; HGB, hemoglobin; Cr, creatinine; CEA, carcinoembryonic antigen; AJCC, American Joint Committee on Cancer staging system; FOBT, fecal occult blood test.

GM alterations in CRC patients compared with healthy family members

Metagenomic sequencing was performed for 52 CRC patients and 55 corresponding healthy family members who were recruited based on various inclusion criteria (Table 1, Table S1). We produced a total of 7.25 billion high-quality sequencing reads (37.56 million reads per individual on average, Table S3) using the Illumina HiSeq platform (Illumina, San Diego, California, USA). The rarefaction curve exhibited a plateau in all samples, suggesting a sufficient sequencing depth for the following analysis (Figure S1).

Age, aspartate aminotransferase, creatinine, fasting glucose and fecal occult blood test had significant differences between CRC patients and controls (Table 1). Permutational multivariate analysis of variance (PERMANOVA) indicated that diagnosis status (P = .015, Table S4) and age (P = .04, Table S4) contributed significantly to the differences of GM between patients and controls. Gender took a slight effect on the discrepancy of GM (P = .072, Table S4), while other clinical indicators, including body mass index, aspartate aminotransferase, fasting glucose, creatinine, etc, associated insignificantly with GM (P > .05, Table S4). We adjusted the confounding effect from age and gender with a generalized additive model. Microbial diversity slightly increased in the GM of CRC patients compared with that of healthy individuals (P > .05, Figure 1a). The phylum Bacteroides and Firmicutes dominated the GM of both healthy and diseased individuals, but the difference of them was not observed between healthy and patients (adjusted P > .05, Figure 1b). In addition, Proteobacteria was slightly enriched without significant difference in the GM of CRC patients (adjusted P > .05, Figure 1b). In addition, the difference in gene number was not obvious between two groups (P > .05, Figure 1c). Compared with that of healthy family members, the genera in CRC patients distinctly enriched by Coprobacillus, Burkholderia, Porphyromonas, Paracoccus, Peptoniphilus, Synechococcus and Cyanothece (adjusted P < .05, Figure 1d, Table S5). We also identified several accumulated microbial species in the GM of patients including Roseburia inulinivorans, Clostridium ramosum, Porphyromonas gingivalis, F. nucleatum, Gemella morbillorum, etc. (adjusted P < .05, Figure 1d, Table S5). We also observed that the correlations of the microbial species were obviously different in CRC patients and healthy people (Figure S2).

Figure 1.

Dysbiosis of GM in CRC cases compared with healthy family members. (a) α-Diversity of GM in genus and species between CRC patients (blue) and controls (pink). (b) Comparison of phylum in microbiomes of CRC (blue) cases and controls (pink). (c) Comparison of gene numbers in CRC cases (blue) and controls (pink). (d) Relative abundance of genera and species in the GM of CRC cases (blue) and healthy controls (pink). Only the microbial genus and species with more than 0.01% relative abundance are shown in the x-axis. The red name genus indicated significantly different between the case and control (adjusted P < .05). (e) Relative abundance of annotated genes with KEGG level 2 in CRC cases (blue) and healthy controls (pink). Only the functional categories with ≥0.1% relative abundance are shown in the x-axis. The red name function indicated significantly different between the case and control (adjusted P < .05).

On average, 624,404 microbial genes were observed in the GM of healthy controls, whereas only 585,092 genes were found in CRC cases (Figure 1c). In addition, a decreased proportion of functional category named amino acid metabolism was observed in the GM of inpatients (Figure 1e).

Discovery of CRC biomarkers in CQ family cohort

By comparing the GM of CRC patients with that of healthy individuals, we identified 22 microbial genes that were strongly associated with CRC (Table S6). Of the 22 genes, 20 enriched in the GM of CRC cases and 2 enriched in the GM of healthy family members (Figure 2a). Almost half of the biomarker genes (N = 8) were assigned as from phylum Firmicutes. Four biomarker genes were classified into Fusobacteria and 10 biomarker genes remained unknown. In eight genes from Firmicutes, five genes were classified as belonging to two microbial species, including Clostridium symbiosum (N = 4) and S. moorei (N= 1). Of the remaining three genes, one was assigned to the genus Coprobacillus and the other two were unknown. In four genes from phylum Fusobacteria, one gene was classified into F. nucleatum and the rest three genes were identified as genus Fusobacterium (Table S6). Of the 12 candidate genes with known functional annotation, four CRC-associated genes were annotated to quorum sensing (N= 2), butanoate metabolism (N= 1) and microbial metabolism in diverse environments (N = 1). The other eight candidate genes were annotated to acting on a sulfur group of donors, ribosome, genetic information processing, type II secretion system, porphyrin and chlorophyll metabolism, replication and repair, acting on peptide bonds (peptidases) and arginine biosynthesis (Table S6).

Figure 2.

Enrichment and performance of 22 biomarker genes among three cohorts. (a) Candidate genes are shown in the left vertical column, and corresponding phylogenetic and functional assignments are shown in the right column. The relative abundance is represented by different colors (log₁₀ (relative abundance); white, not detected; red, the most abundant).The red and green gene names indicate CRC enriched and health enriched genes, respectively. (b) The red curve indicates ROC in the validation cohort; the green and aquamarine curves indicate ROC in the HK and French cohorts, respectively.

The receiver operating characteristic (ROC) curve indicated that the 22 candidate genes significantly differentiate CRC cases and healthy families (area under the ROC (AUC) = 0.998, 95% CI 0.993–1.000). When applied to 46 additional CRC cases and 40 randomly selected healthy controls (validation cohort, Table S1), the 22 candidate genes also exhibited high sensitivity for CRC screening (Figure 2b, AUC = 0.905, 95% CI 0.832–0.977). In parallel, these 22 biomarker genes were similarly distributed in additional 86 CQ individuals as in the family cohort (Figure 2a).

Sensitivity of GM biomarkers for screening CRC revealed significant regional tendency

To test the performance of 22 biomarker genes in screening CRC in other regions, we collected published metagenomics data from HK (74 CRC cases and 54 healthy adults) and French (53 cases and 88 healthy adults) populations to serve as an external validation cohort (Table S7). Similar to CQ populations, 20 of 22 biomarker genes tended to accumulate in CRC patients in the HK and French cohorts while the remaining two genes enriched in healthy controls (Figure 2a). The ROC curve showed a high sensitivity of biomarker genes for classifying CRC patients in HK (AUC = 0.811, 95% CI 0.715–0.907) and French (AUC = 0.859, 95% CI 0.773–0.944) populations (Figure 2b), while these biomarker genes from CQ cohort seemed to exhibit reduced accuracy when applied to CRC screening in population with different living regions and diet habits (Figure 2b).

Living region was more significant for differentiating microbial samples compared to CRC risk

To understand the reduced accuracy of CQ GM biomarkers to detect CRC in other region populations, we analyzed the distribution of all microbial samples in three cohorts. Non-metric multidimensional scaling analysis indicated that microbial samples in the same region were clustered together (Figure 3a). In addition, microbial samples in CQ were more similar to those in HK than those in French populations (Figure 3a).

Figure 3.

Cluster of microbial samples from differed populations and distribution of biomarker bacteria in three cohorts. (a) Dot and triangle indicate microbiomes from CRC cases and healthy controls, respectively. Blue, green and red colors represent microbial samples in CQ, HK and French cohorts, respectively. (b) Relative abundance of the four biomarker microbes in three cohorts. The significant difference of abundance within the individual cohort is labeled with * (*, P < .05; **, P < .01; ***, P < .001).

Further analyses indicated the lowest GM diversity in the CQ cohort and the highest in the French cohort (Figure S3a, adjusted P < .05). Bacteroides dominated GM in three cohorts, whereas Prevotella represented the second highest abundance in the GM of the CQ cohort (Figure S4, Table S8). In addition, compared with the other two cohorts, these two dominant genera enriched higher in CQ cohort. In contrast, Escherichia and Ruminococcus represented the second most abundant microbial genera in the GM of the HK and French cohorts, respectively (Figure S4, Table S8). However, Escherichia had the lowest relative abundance in CQ cohort, compared with the other cohorts. The relative abundance of Ruminococcus was highest in French cohort and was different among the three cohorts. At the species level, Bacteroides dorei.vulgatus, Escherichia coli and Bacteroides uniformis were dominant in the GM of the CQ, HK and French populations, respectively (Figure S4, Table S8). Prevotella copri differed among three cohorts. Compared to the discrepancy in GM among different regions (Table S8), structural variations in GM between diseased and healthy individuals of the same region reduced slightly (Figure S4). For instance, the top five microbial genera in the GM of both inpatients and healthy controls in CQ included Bacteroides, Prevotella, Faecalibacterium, Eubacterium and Escherichia (Figure S4a, Table S8). Regarding the HK and French cohorts, the top 10 GM genera between cases and healthy individuals were similar (Figure S4b, c, Table S8).

Additionally, abundantce of biomarker gene-related microbes were different in three cohorts. Relative abundance of Coprobacillus differed between diseased and healthy people only in CQ cohort, and that of F. nucleatum differed between CRC patients and healthy controls in CQ, HK and French cohorts. The distribution of S. moorei showed a discrepancy between diseased and healthy subjects in French cohort. The relative abundance of C. symbiosum revealed a distinct difference between patients and healthy people both in HK and in French cohorts (Figure 3).

Three biomarker genes were verified by qPCR in independent chongqing population

According to the degree of gene’s contribution to the accuracy of classification CRC patients, we selected three genes from 22 biomarkers for further verification (gene 8122329, unknown function from Coprobacillus; gene 3742340, nitrilase from C. symbiosum; gene 5053929, peptide methionine sulfoxide reductase msrA/msrB from Fusobacterium). Their relative abundances were measured by qPCR in an independent Chongqing population (30 cases and 30 controls, age and gender were not significant different between patients and controls groups, Table S2). The results showed that the relative abundance of two biomarker genes (from Coprobacillus and C. symbiosum) differed significantly between CRC patients and healthy controls (adjusted P < .05, Figure 4a). But the biomarker gene from Fusobacterium did not show a significant difference (adjusted P > .05, Figure 4a). Density curve illustrated discrepant distribution of two biomarker genes’ (from Coprobacillus and C. symbiosum) relative abundance between CRC cases and healthy controls (Figure 4b). However, the biomarker gene from Fusobacterium displayed a similar distribution in the same population (Figure 4b). Therefore, the biomarker gene from Fusobacterium was not regarded as an appropriate independent CRC indicator and was excluded in the following analyses. Based on random forest classification algorithm, ROC curve indicated that the selected biomarker gene from Coprobacillus significantly classified CRC patients from healthy controls in 60 independent Chongqing population (Figure 4c, AUC = 0.930, 95% CI 0.904–0.955). In addition, the AUC value increased slightly after adding genes from C. symbiosum into classification model (Figure 4d, AUC = 0.935, 95% CI 0.883–0.987).

Figure 4.

Difference in three biomarker genes and performance of two selected biomarker genes in independent Chongqing population. (a) Relative abundance of three biomarker genes (8122329 indicates gene from Coprobacillus; 3742340 indicates gene from C.symbiosum and 5053929 indicates gene from Fusobacterium) in independence Chongqing cohort (30 CRC patients and 30 healthy controls) with qPCR. (b) Distribution of density curve of three biomarker genes based on qPCR △Ct value (gene from Coprobacillus, C. symbiosum and Fusobacterium). (c) ROC curve of the gene from Coprobacillus. (d) ROC curve of the gene from Coprobacillus and C.symbiosum.

Discussion

CRC is one of the most malignant tumors worldwide and early screening can help reduce associated mortality.^2,15 Currently available methods for early CRC diagnosis include fecal occult blood test, colonoscopy examination, fecal immunochemical test and carcino-embryonic antigen test. Such tests have many limitations, including low accuracy and invasive techniques.^16,17 Emerging studies have demonstrated imbalanced GM in CRC patients and the promise of GM biomarkers in CRC screening.^6,18,19

Prior studies documented that the low-abundance microbes in GM of CRC patients contributed to CRC development, such as F. nucleatum, P. stomatis, P. micra and S. moorei, most of which were recognized as potential GM biomarkers.^20,21 Previous reports suggested the potential role of our identified low-abundance microbial species and genera, including F. nucleatum, P. gingivalis, G.morbillorum, P. nigrescens, Porphyromonas and Coprobacillus, in CRC development.^21–23 Recent study found that F. nucleatum was able to promote intestinal tumorigenesis via adhering to cancer cells, modulating immune cells and modifying the tumor microenvironment.^24–26 In addition, F. nucleatum also stimulated Toll-like receptor 4 to modulate CRC patient’s chemotherapeutic response, which possibly influenced the treatment outcome.²⁷ The other low-abundant microbes, such as P. gingivalis, P. nigrescens, G. morbillorum, and porphyromonas, were also considered to be strongly associated with CRC development.^22,23,28 Additionally, further study should be conducted to analyze Coprobacillus, being rarely reported but identified as high-accuracy biomarker in CQ cohort.

Emerging studies indicated that GM-related metabolism, such as synthesis, deaminization and decarboxylation, affected cancerous conditions in the gut.^29,30 Previous researches identified increased abundance of microbial genes related to amino acid metabolism in GM of CRC patients,³¹ which was consistent with our results. The association between microbial metabolism of amino acids and CRC development may be established by producing toxic metabolites. For example, GM was able to degrade phenylalanine, being an aromatic amino acid, into CRC-associated toxic phenylacetate through catabolic pathways.^32,33 Additionally, GM could metabolize sulfur-containing amino acid, including cysteine and methionine, via sulfide-producing pathway and generate toxic H₂S which contributed to CRC incidence.³⁴

Given that GM could be shaped by various external factors, including living environments, diet, lifestylelife style and antibiotics exposure,^35–37 which should be considered when applying GM biomarkers in screening CRC. Our study selected CRC patients and their healthy families to decrease the confounding effects of living environment and lifestyle on GM differences. After adjusting the confounding effect of the age in the generalized additive model, identified GM biomarkers showed higher accuracy in screening CRC compared to published biomarkers.^8,14 Two biomarker genes from C. symbiosum and Coprobacillu were verified with high accuracy in Chongqing population via qPCR. C. symbiosum was confirmed to be related to CRC in previous studies and reported as a GM biomarker for detecting early and advanced CRC patients in Shanghai population in China.^14,38,39 In addition, another biomarker gene from genus Coprobacillus was enriched in GM of ethanol-related CRC patients,⁴⁰ which coincided with the fact that alcoholic beverages were widely accepted in Chongqing.⁴¹ This may partly explain that Coprobacillus performed well when screening CRC in Chongqing population. Recent studies also showed the improved accuracy of GM biomarkers when diminishing the confounding effect of environmental factors from various population cohorts.^22,23 However, these cross-cohort studies might neglect population-specific microbes or DNA fragments, such as Coprobacillus which was identified in our study and validated by qPCR in an independent cohort with high accuracy (AUC = 0.930).

Several limitations should be noted in our study. The family paired sampling design in our study is unable to totally eliminate effects from external factors, attributing to the difference of dietary, activity area and mental state in each family member. Therefore, the external factor contributing to confounding effect is supposed to be considered and adjusted in the following study, such as the age in our study. In addition, the accuracy of biomarkers identified from family-paired cohort is also limited by external factors in different populations. The recent cross-cohort studies give us a new angle to face this limitation: cross-cohort and family paired sampling method could be integrated for identifying higher accuracy and wilder applicable range biomarkers under appropriate inclusion criteria, sufficient clinical information and minimum sample size. Moreover, updated integrated gene catalog (IGC) database was employed for taxonomic annotation, which is the largest free accessible non-redundant gene catalog with 11,446,577 genes of human GM.⁴² But the limitations of this database are lacked verification from cultivation experiment and are not timely updated taxonomic and functional annotation.⁴³ In 2019, the culturable genome reference (CGR) was established by 1,520 cultivated and assembled bacteria genomes.⁴⁴ Although some strictly anaerobic bacteria might be inevitably omitted through cultivating, CGR database fills the limitation of IGC through cultivation experiments. We hope to revise our taxonomic annotation when CGR is available. CGR and ICG databases have a slight different scope of application: the CGR database might be fit to the study focused on specific species in GM, and the IGC database is more appropriate to the study of whole GM, such as identifying biomarkers from a dataset.

In general, GM from different population displayed respective characteristic because of a greatly confounding effect from regional environmental factors. However, regional specific biomarkers could be identified based on the GM and displayed high accuracy in a particular population. Therefore, regional specific GM biomarkers would be an efficient clinical tool for CRC diagnosis. Further analysis is required to promote the clinical application of GM biomarkers, including validating our findings in a large cohort study and exploring the diagnosing index integrating GM biomarkers and clinical indices. It is foreseen that, with the GM research developments, the regional specific GM biomarker will promisingly become conventional diagnosis method for CRC screening.

Materials and methods

Subject inclusion and sample collection

CRC patients were diagnosed histopathologically by colonoscopy at the hospital.⁴⁵ The exclusion criteria for patients were as follows: ≤ 18 years old or ≥ 76 years old, colorectal benign lesion, antibiotics exposure, radiotherapy therapy or continuous treatment by systemic corticosteroids 1 month prior to sampling and serious mental disorder. Healthy individuals were selected based on the following exclusion criteria: dysentery, chronic enteritis, inflammatory bowel disease, irritable bowel syndrome, Crohn’s disease, metabolic diseases (BMI ≥ 32, diabetes or malnutrition), long-term probiotics uptake, continuous treatment by systemic corticosteroids 1 month before sampling or serious mental disease. Selected cases’ healthy family members lived together with the patients for at least 1 year. The ethics committee of The First Hospital Affiliated to Army Medical University approved this study (Registration Number: KY201737). Informed consent were obtained from recruited patients and healthy individuals.

Fresh stool, blood and urine of each patient were collected in the early morning after admission to the hospital. All collected samples were frozen on dry ice within 30 min and then stored at −80°C until further analysis. Fresh stools from healthy subjects were collected during physical examination in the hospital.

Laboratory assessment

A blood auto-analyzer (Beckman Coulter AU5800, Brea, CA, USA) was used to analyze serum creatinine, fasting plasma glucose, serum alanine aminotransferase, aspartate aminotransferase, γ-glutamyl transpeptidase and albumin. Tumor markers, such as carcino-embryonic antigen, were tested in a detection system (Luminex200 xMAP, Austin, TX, USA). Fresh stools were processed on an automatic processing platform for standard stool examination, including fecal occult blood test, stool color and shape (JinHua JHSPSY-I, Nanchang, JiangXi, China).

DNA extraction, library construction and sequencing

DNA extraction was performed using the QIAamp DNA Stool Mini Kit (QIAGEN, 19593).⁴⁶ DNA concentration was measured using a NanoDrop spectrophotometer (Thermo Scientific, Scoresby, Victoria, Australia) and Qubit Fluorometer (Life Technologies, Grand Island, NY, USA). Molecular size was evaluated by agarose gel electrophoresis. All DNA samples were stored at −20°C before further process. DNA libraries were constructed following the manufacturer’s instruction (Illumina, San Diego, CA, USA) with an insert size of 350 bp. Paired-end sequencing was performed on the Illumina HiSeq platform (San Diego, CA, USA). Sequencing data were accessible in the GenBank database under accession number: SRP128485.

Phylogenetic and functional profiling

Paired-end reads were filtered using the following criteria: one of the paired reads contained 10% ambiguous Ns or contained 50% bases with a quality score <20. The filtered paired-end reads were mapped to the human genome (hg19) to eliminate genomic contamination with SOAPaligner 2.21 (-m 244 -x 455 -v 5 -r 1 -l 35 -M 4).⁴⁷ The remaining high-quality reads were aligned with an updated IGC⁴² using bowtie2 (v 2.3.0) (–very-sensitive-local – score-min L,0,1.6 -k 10 -p 8 -I 200 -X 500 – reorder – omit-sec-seq -N 1 – ignore-quals).⁴⁸ Three types of alignment results could be observed when paired-end reads aligned to genes in the database: (1) paired-end reads aligned to one gene, which was counted as two reads mapping; (2) paired-end reads aligned to more than one gene, the primary aligned gene was counted as two-read mapping, and the remaining genes were counted zero; or (3) one of paired-end reads aligned to one gene, which was counted as one-read mapping.

Then, the number of mapped paired-end reads to each gene in IGC was counted. For a given gene g, its relative abundance is Ab(g)relative, which was calculated using the following formula:

$$\mathrm{A}\mathrm{b}(\mathrm{g})\mathrm{r}\mathrm{e}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}=\mathrm{A}\mathrm{b}(\mathrm{g})\times 100/\mathrm{A}\mathrm{b}(\mathrm{G})$$

$$\mathrm{A}\mathrm{b}\left(\mathrm{g}\right)=\mathrm{U}/\mathrm{L}$$

Ab(g) is the abundance of gene g, Ab(G) is the sum of all genes abundance, U is the number of mapped reads in gene g and L is the length of gene g.⁴⁹

The relative abundance of taxonomy (Phylum, genus and species) and KEGG orthology was calculated by summing Ab(g) of corresponding genes based on the annotated results in the IGC database.

Statistical analysis

Based on taxonomy profiling, alpha diversity was assessed by Shannon index, and the rarefaction curve was generated with vegan package (version: 2.4–4; R: 3.3.1).⁵⁰ The confounding effect was tested by PERMANOVA. Generalized additive model was applied for removing the confounding effect from age and gender.⁵¹ ROC curves were generated with pROC in R (pROC: 1.10.0, R: 3.3.1). Density curve was drawn by R software (version: 3.3.1). Wilcoxon rank-sum test was applied to explore inter-group differences and the significance (P-value) was adjusted using the Benjamini and Hochberg method.⁵²

Biomarkers identification

Genes absent in more than 80% of the samples were removed. Generalized additive mode was employed for identifying different microbial genes between diseased and healthy subjects (P < .01). A two-step schema was then applied for biomarker identification. Firstly, we performed feature selection on these retained genes which were filtered using the mRMR algorithm⁵³ and the top 100 ones were enrolled as candidates for further analysis. In the second stage, random forest model was used to build the classifier based on the top k-th genes,⁴⁹ by varying k in the set 1 to 100. The biomarkers were identified based on the performance of the classifiers which was assessed by 5-fold cross-validation and ROC.¹⁴

Quantitative PCR

Primers for three microbial biomarker genes and bacterial universal 16S rDNA gene were synthesized and purified by Sangon Biotech (Shanghai, China). The primer sequences were presented as follows: 8122329 gene For-CTCTGTCAAGAGGAAAGTTCAATTCTTG; 8122329 gene Rev-CTCTCTTTCTTTGGATTCTGCAAGTG; 3742340 gene For-TTATGTGGGGACGGATAATGCG; 3742340 gene Rev-GGTATTTTATGGTATTTGGCCGCC; 5053929 gene For-TGGAAAAGATGGAAAACCAACTTATGTT; 5053929 gene Rev- CACGAACATTTAATATCTTTGATAATTCACCTTT. Internal reference primer for total bacterial DNA was determined by 16s rRNA using the following primers: 16s rRNA For-GTTGTCGTCAGCTCGTGTCG. 16s rRNA Rev-GCAGTCTCGCTAGAGTGCCC. qPCR amplifications were performed in a 20-μl reaction system containing 10 μl 2 × SYBRTM Premix Dimer Eraser (Takara Bio, RR091A), 2 μl of extracted fecal DNA and 1.2 μl of primers. Amplification and detection of DNA were performed with the C1000 Thermal cycler (Bio-Rad, Hercules, CA, USA) via the following reaction conditions: 95℃ for 3 min, followed by 40 cycles of 95°C for 5 s and 60°C for 30 s. Each sample was assayed in triplicate and the mean of the three-cycle threshold (Ct) values was used for subsequent analysis. The Ct value is defined by the number of cycles for the fluorescent signal reaching the threshold in qPCR. The abundance of the biomarkers in each sample was calculated by 2^−ΔΔCt methods (ΔCt = Ct_target – Ct_control, ΔΔCt = ΔCt – ΔCt_maximum).

Disclosure of potential conflict of interests

The authors declare that they have no competing interests.

Supplementary Material

Supplemental data for this article can be accessed in the publisher’s website.