A systematic review of microbiome-derived biomarkers for early colorectal cancer detection

Abstract

Increasing evidence suggests a role of the gut microbiome in the development of colorectal cancer (CRC) and that it can serve as a biomarker for early diagnosis. This review aims to give an overview of the current status of published studies regarding the microbiome as a screening tool for early CRC detection. A literature search was conducted using PubMed and EMBASE in August 2022. Studies assessing the efficacy of microbiome-derived biomarkers based on noninvasive derived samples were included. Not relevant studies or studies not specifying the stage of CRC or grouping them together in the analysis were excluded. The risk of bias for screening tools was performed using the QUADAS-2 checklist. A total of 28 studies were included, ranging from 2 to 462 for CRC and 18 to 665 advanced adenoma patient inclusions, of which only two investigated the co-metabolome as biomarker. The diagnostic performance of faecal bacteria-derived biomarkers had an AUC ranging from 0.28-0.98 for precursor lesions such as advanced adenomas and 0.54-0.89 for early CRC. Diagnostic performance based on the co-metabolome showed an AUC ranging from 0.69 – 0.84 for precursor lesions and 0.65 – 0.93 for early CRC. All models improved when combined with established clinical early detection markers such as gFOBT. A high level of heterogeneity was seen in the number of inclusions and methodology used in the studies. The faecal and oral gut microbiome has the potential to complement existing CRC screening tools, however current evidence suggests that this is not yet ready for routine clinical use.

Introduction

Colorectal cancer (CRC) is the third commonest cancer worldwide [1] and a leading cause of cancer-associated deaths. Despite this, CRC is often treatable with a high overall survival when detected and treated in its early stages [2]. This has led to the formation of global screening programs for CRC [3], which typically employ faecal biomarkers like the guaiac faecal occult blood test (gFOBT), followed by endoscopy [4]. The gFOBT, is a fast, noninvasive and inexpensive test, however, its diagnostic sensitivity ranges from 7 to 21% for the detection of early cancer (Stage 0 and I) and is unsuitable for the detection of complex adenomas [5,6]. Faecal immunochemical test (FIT) offers a slight improvement in its diagnostic yield for early cancer (maximum sensitivity of 25%), however, with higher associated costs [7]. The FIT-DNA test (or stool DNA test), a test that combines the FIT with DNA markers associated with CRC, shows a further increasement in sensitivity, however at the expense of a decrease in specificity, resulting in more false positive results [8].

Blood-based biomarkers such as Septin 9 and protein-based markers (e.g. CEA, TIMP-1 etc.) have also failed to provide diagnostic utility for early detection [9]. Novel data from recent trials suggest that circulating DNA (cDNA) biomarkers for colorectal cancer have diagnostic and prognostic utility, particularly for the stratification of chemotherapeutic strategies [10] and the prediction of cancer recurrence. However, despite the promise of these approaches for early detection, very limited data exist for the identification of high-risk adenomas.

Increasing evidence points to the gut microbiome as a critical mediator of adenoma formation and CRC risk through parallel processes of co-metabolic dysfunction [11], inflammation, epigenetic programming and DNA damage [12]. Multiple studies have identified the presence of intra-tumoural bacteria in precursors lesions of cancer [13], [14], [15], [16] such as Fusobacterium nucleatum, Escherichia coli, and Bacteroides fragilis [17]. A loss of microbial diversity and increases in the abundance of both mucosal and faecal pathobionts can distinguish healthy persons from patients with CRC [18], [19], [20]. Importantly, microbiome changes can already be observed in very early stages of CRC [21]. It has therefore been hypothesised that the microbiome has value as a biomarker for early cancer detection or to identify those individuals at risk of this disease [22]. This systematic review aims to provide a comprehensive overview of the current state of microbiome-derived biomarkers, based on their genes and their products, for early CRC detection.

Methods

Search strategy

A literature search was carried out from inception to the 3^rd of August 2022, using the databases PubMed and EMBASE (via OVID). A combination of the following terms was used: colorectal cancer, screening/early detection, and microbiome. No filters or restrictions were applied. Literature was reviewed following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [23]. The complete search strategy can be found in Supplementary Table S1.

Study selection

Results were screened independently by two authors (F.Z. and H.S.) based on title and abstract for relevance and assessment for inclusion based on full text. Articles were evaluated via Rayyan, developed by Qatar Computing Research Institute, a platform that allows blinded assessment of articles based on title, abstract, and key-words by multiple reviewers [24]. Disagreement of articles screened for possible eligibility based on title and abstract was resolved by discussion. Conflicts over the inclusion of articles for assessment following full-text screening were resolved by consulting a third reviewer (J.K.). If studies used the same datasets, the study most appropriate to our study was selected.

Eligibility criteria

Articles were included for assessment if they assessed the efficacy and/or use of microbiome-derived biomarkers for early CRC detection, including precancerous lesions: complex adenoma > 1 cm, high-grade dysplasia and early cancer: carcinoma in situ (CIS), malignant polyps and stage 0, I or II with N0, M0, according to the American Joint Committee on Cancer (AJCC) TNM Classification of Malignant Tumors [25]. Articles were excluded based on: (I) viral, fungal or solely human-derived biomarkers; (II) investigating prognostic biomarkers; (III) disease other than CRC; (IV) advanced stage CRC; (V) inflammatory bowel disease (IBD)-related CRC or CRC as part of a syndrome; (VI) in vitro or animal studies; (VII) language other than English; (VIII) small sample size (<10); (IX) no full text available; (X) case report or case series; (XI) review; (XII) no clear distinction of the CRC stages assessed; (XIII) all CRC stages grouped in the analysis.

Since this study investigates the use of microbial-derived biomarkers as a screening tool for the early detection of CRC, all sorts of invasive screening studies, for example, studies assessing microbiome composition of the mucosal tissue based on biopsies, were excluded. Studies assessing saliva, urine and blood were regarded as minimally invasive procedures and were included.

Data extraction and assessment

Extraction of data from the eligible articles was performed independently by the two authors (F.Z. and H.S.). Confirmation of the extracted data was performed by a third author (K.Z.). Data extraction included primarily: year of publication, country, number of participants, age, percentage of male, stage of disease, sample origin, biomarker found, technique used for discovery, diagnostic performance of studied biomarker (based on sensitivity, specificity or area under the curve (AUC)) and, if described, comparison with currently used screening tools (e.g. gFOBT, FIT). If any of the data was not described directly, additional information was used to calculate the necessary values if possible. Data was rounded off at either two decimals (AUC) or no decimals (sensitivity, specificity). If stated, specific bacteria or their molecular products and functions per cohort and methodology were reported.

Comparison with healthy controls

Comparison with healthy controls was based, if applicable, via colonoscopy and defined as subjects without prior gastro-intestinal disease(s) or colonoscopic findings with the exception of a few polyps (<5mm) present.

Assessment

Assessment of the strength of overall data regarding individual biomarkers based on their risk of bias and applicability was undertaken using the QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies checklist) [26].

Included studies were independently evaluated for the risk of bias and applicability by two authors (F.Z. and H.S.). The risk of bias was assessed across four domains: patient selection, index test, reference standard, flow and timing (i.e. an appropriate interval between index test and reference as well as the inclusion of all patients in the analysis).

Concerns regarding applicability were assessed on three domains: patient selection, index test and reference standard. Signalling questions of all domains were tailored for this review. Based on these questions, the risk of bias and applicability was classified as low, high or unclear for each domain.

Results

Study selection

A literature search was performed using the databases PubMed and EMBASE (via OVID). A total of 3859 records were retrieved, 1427 from PubMed and 2432 from EMBASE. After removing duplicates, 3006 records were screened based on title and abstract, resulting in 513 records being evaluated for eligibility based on full-text screening. A total of 69 studies evaluating microbiome-derived indicators for early CRC diagnosis were again carefully screened, resulting in another 41 records to be excluded based on: a lack of CRC stage distinction (n = 8), all stages grouped together in the analysis (n = 11), screening based on tumour tissue (n = 2), meta-analysis or usage of the same dataset in another article already included (n = 19) and insufficient data reported (n = 1). A total of 28 studies remained to be included in this systematic review. Study selection and exclusion were based on the PRISMA guidelines [23]. See Fig. 1 for a detailed overview of the selection process.

Fig. 1.

PRISMA flow diagram of exclusion and included of studies.

Study characteristics

For microbiome-derived biomarkers based on bacteria for the detection of early CRC, 26 articles were found and two articles were found investigating the co-metabolome solely as a biomarker. Seventeen of the studies were conducted in East Asia [19,[27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37], [38], [39], [40], [41], [42]], one in South Asia [43], eight in Europe [44], [45], [46], [47], [48], [49], [50], [51] and two in North America [52,53]. The number of inclusions of these studies ranged between 2 and 462 for CRC, 18 and 665 for adenoma and/or polyp and 24 and 788 for healthy controls. Regarding the stage used in the articles for early CRC detection, five studies were based on stage I or II [29,30,37,40,50] and two used stage 0 [19,38]. The other studies investigated a precursor lesion of CRC, either advanced adenomas or polyps (AP) [28,[31], [32], [33], [34], [35], 39,41,[44], [45], [46], [47], [48], [49], [51], [52], [53]]. Only five articles evaluated both precursor lesions of CRC and stages I and II [27,36,38,43,54]. All articles used faecal samples, except for four: two studies used the microbiota collected from the oral mucosa instead of or next to faecal samples [31,46] and two used serum-based samples next to faecal samples [36,54]. Whenever possible, the mean age and percentage of men per group were extracted. The reported mean age ranged from 59 to 73 in the CRC group, from 48 to 68 in the AP group and from 49 to 67 in the healthy controls. The percentages of men in each group ranged from 32 to 67, 19 to 78 and 17 to 56 per cent for CRC, AP and healthy controls, respectively. See Table 1 for the study characteristics as well.

Table 1.

Study characteristics.

Author, year	Country	Sample	Examined samples (CRC/AP/HC)1			CRC stage according to TNM ((0)3/I/II/III/IV)
(n)	Age (mean (±SD))2	Gender, m (%)
Bacterial taxa
Zhang Y, 2021 [32]	China	Faeces	N/A/268/788	N/A/61.7 (6.2)/59.9 (6.2)	N/A/74/48	N/A
Zhang S, 2020 [31]	China	Oral mucosa	161/34/58	59.2 (10.7)/51.8 (7.7)/50.7 (11.3)	67/59/53	24/66/60/11
Yachida, 2019 [19]	Japan	Faeces	258/67/251+40 with history of colorectal surgery	0 64.8 (7.5), I/II: 63.8 (9.3), III/IV: 59.1 (11) /63.2 (9.1)/ 64.7 (10.6) + 59.2 (12.3)	0 62, I-II 66, III-IV 58/72/28 + 33	73/75/36/52/22
Yao, 2021 [30]	China	Faeces	206/ N/A /112	62.2 (8.6)/ N/A /56.2 (12.8)	61/ N/A /56	98/47/49/12
Xie, 2017 [29]	China	Faeces	445/288/304	I-IV 63.5 (10.2), I-II 64 (8.9)/61.9 (8.7)/ 60.1 (8.4)	I-IV 61, I-II 65/63/ 55	I/II: 142; III/IV: 303
Wei, 2022 [28]	Taiwan	Faeces	20/67/60 + 40 COB4	64 (43–88)/48 (39–60)-/ 61 (31–72) + 52 (35–63)	35/42/41 + 45	Not distinguished
Mo, 2021 [27]	China	Faeces	108/18/36	58 (26-86) of n = all	58 of n = all	20/39/37/12
Liu, 2021 [41]	China	Faeces	60/37/42	64.1 (11.1)/66.5 (3.5)/61.0 (7.9)	32/19/31	Not distinguished
Liang, 2017 [40]	China	Faeces	170-2005/97/33-365	67.2 (11.6)- 63.4 (9.6)/60.5 (4.7)/59.3 (5.8)-53.2 (12.2)	59-52/52/39-28	Stage subset (n=111) 17/42/43/9/111
Liang, 2021 [39]	China	Faeces	210/115+866/265	67.0 (11.3)/61.1 (6.8) + 60.2 (5.0)/58.1 (7.7)	59/57 + 57/42	31/68/73/28
Konishi, 2022 [38]	Japan	Faeces	462/240/317	56/ - /66	52 / - /53	68/107/77/130/44
Guo, 2018 [37]	China	Faeces	215/ N/A /156	61.2 (12.3)/ N/A / 48.6 (10.3)	55/ N/A / 45	38/59/75/43
Gao, 2022 [36]	China	Faeces	35/31/34	66.1 (10.8)/64.4 (8.4)/57.8 (5.7)	57/58/17	3/21/9/2
Goedert, 2015 [47]	China	Faeces	2/20/24	65 (61–69) of n = all	N/A/60/29	1/0/1/0
Gao, 2020 [35]	China	Faeces	55/85/110	N/A /63.1 (12.8)/65 (10.4)	N/A /63/29	N/A
Coker, 2022 [34]	China	Faeces	118/140/128	73.2 (10.4)/65.8 (5.5)/64 (6.8)	54/59/46	N/A
Ai, 2017 [33]	China/France	Faeces	42/47/52	62.9 (1.5)/58.9 (1.5)/52.3 (1.5)	43/51/40	I/II 12 III/IV 30
Rezasoltani, 2018 [43]	Iran	Faeces	20/42/31	60.9 (13.5)/58.6 (13.5)/59.8(17)	56/55/52	20/0/0/0
Zeller, 2014 [50]	France	Faeces	53/42/61	70.5-65.0*/62.0-68.0⁎⁎/63.0	55/71/46	15/7/10/21
Young, 2021 [49]	UK	Faeces	430/665/6667	68.1 (5.0)/66.3 (4.7)/66.6 (4.3)	67/65/52	Not distinguished
Tarallo, 2019 [48]	Italy	Faeces	29/27/24	Not disclosed	Not disclosed	Not distinguished
Flemer, 2017 [46]	Ireland	Faeces	69/23/62	65.3 (10.8)/60.4 (13.4)/63.9 (11.1)	67/78/51	Not distinguished
		Oral mucosa	45/21/25	65.7 (10.9)/59.2 (15.1)/51.5 (12.4)	56/71/38
Eklöf, 2017 [45]	Sweden	Faeces	39/134/65	N/A	51/60/54	2/21/8/7
Clos-Garcia, 2020 [44]	Spain	Faeces	All: 99/69/77 Analysis: 83/62/74	70.2/68/65 (all)	62/60/45 (all)	3/22/22/30/6 (analysis)
Zackular, 2014 [53]	North America	Faeces	30/30/30	59.4 (11.0)/ 61.3(11.1)55.3 (9.2)	70/60/37	Not distinguished
Baxter, 2016 [52]	USA/Canada	Faeces	120/198/172	60 (median, range 29-89)	Not disclosed	Not distinguished
Co-metabolome (different cohort)
Yachida 2019 [19]	Japan	Faeces	178/45/149+34 with history of colorectal surgery	0 63.7 (8.5), I/II 63.6 (8.8), III/IV59.5 (11.5)/64.4 (8.5)/64.1 (10.9) + 60.6 (11.9)	0 50, I-II 65, III-IV 53/ 78/58 + 47	30/51/29/44/24/34
Gao, 2022 [36]	China	Serum	35/31/34	66.1 (10.8)/64.4 (8.4)/57.8 (5.7)	57/58/17	3/21/9/2
Chen, 2022 [54]	China	Serum Faeces	84/19/53	58 (7.2)3/53.8 (7.3)	723/38	8/9/18/20/19 (29 no record)
Bosch, 2022 [51]	the Netherlands	Faeces	N/A /19/19	N/A /73 (6.1)/68 (10.4)	N/A /90/68	N/A

¹CRC = colorectal cancer, AP = adenoma and/or polyp, HC = healthy controls

²Unless differently defined

³Stage 0 = Intramucosal carcinoma/polypoid adenomas with high-grade dysplasia, pTis/Stage 0 CRC (as defined by Yachida et al. [19])

⁴Colonic occult blood loss

⁵HC and CRC based on two separate cohorts

⁶Asymptomatic and symptomatic patients with AP. Analyses used in this review only on asymptomatic patients.

⁷Normal and non-neoplastic colonoscopy grouped together

^⁎Divided in stage I/II and III/IV,

^⁎⁎Divided in small and large adenoma

Diagnostic performance of bacteria-derived biomarkers

The overall performance of bacteria-derived biomarkers for the detection of precursor lesions showed an AUC ranging from 0.28-0.98, a sensitivity ranging from 18-100 per cent and specificity ranging from 39-97 per cent. Notably, the high AUC of the overall range for precursor lesions was determined by two studies that used oral mucosa instead of or in conjunction with faecal samples. These studies by Flemer et al. and Zhang S. et al. [31,46], reported the highest AUC found of 0.98 (95% CI 0.95-0.98), with a sensitivity of 88 per cent and AUC of 0.95 (95% CI 0.91-0.99), respectively, suggesting a high accuracy for the use of microbiome-derived biomarkers based on the oral mucosa for precursor lesions. The same study by Flemer et al. evaluated the performance based solely on bacteria from faecal samples with a reported AUC of 0.87 (no 95% CI available) [46].

For early stages of CRC, the overall performance of solely bacteria-derived biomarkers, based on AUC as well, showed a range from 0.54 to 0.89 and sensitivity ranging from 60 to 100 per cent. The studies of Yachida et al. and Konishi et al. were the only ones that included stage 0 in their analyses and reported an AUC of 0.73 and 0.54, respectively [19,38]. For stage I, an AUC range between 0.74 and 0.97 was seen, with a sensitivity ranging from 90 to 100 per cent, and stage I/II or only stage II showed a range between 0.82-0.89, with a sensitivity ranging from 60 to 100 per cent. Although some articles demonstrated very high performance in patients with stage I, performance generally increased with increasing stage. The study of Rezasoltani et al. reported the highest performance, with an AUC of 0.97 (no 95% CI available) and a sensitivity and specificity of 100 and 77, respectively [43]. This was based, however, on a very small sample size of only 20 patients with stage I CRC combined with 43 patients with adenomas.

Most panels were based on multiple bacteria ranging from single up to 50; however, no better performance was seen for the detection of precursor lesions or CRC when more bacteria in the panels were applied. However, the performance of bacterial panels was generally enhanced when combined with other characteristics such as body mass index (BMI), age, sex and ethnicity. Likewise, the addition of other faecal markers like FIT or gFOBT strengthened the accuracy of the models overall. The use of only two bacteria (Fusobacterium nucleatum and Parvimonas micra) combined with FIT, faecal DNA methylation (i.e. Septin9, NDRG4, BMP3) or mutation markers (i.e. KRAS, BRAF, PI3KCA) resulted in a reported increase in AUC from 0.57 to 0.73 for the detection of precursor lesions when compared to the bacterial model alone [27]. However, sensitivity and specificity of respectively 39 and 83 for only the bacteria model changed in respectively 28 and 94 when combined. Liu et al. added Fusobacterium nucleatum and pks+ Escherichia Coli to the gFOBT and two blood-based markers (carcinoembryonic antigen (CEA) and carbohydrate antigen (CA) 19-9) with a reported AUC of 0.85 (95% CI 0.57-1.0) [41]. Again, this was accompanied by a drop in sensitivity from 91 per cent based on a measurement of bacteria alone to 67 per cent when combined. However, there was an increase in specificity from 75 to 90 per cent for the combined model. CEA as an additional marker was also used by Xie et al. with FIT and two bacteria (Clostridium symbiosum and Fusobacterium nucleatum) with a reported AUC of 0.82 (95% CI 0.77–0.87) [29]. A summary of the performance capacity of each study is outlined in Table 2.

Table 2.

Included studies assessing microbiome-derived biomarkers for early CRC detection

Reference	Comparison of interest1	Biomarker	Performance for early CRC detection		Analytical method2
AUC (95% CI)	Sens/spec
Zhang Y, 2021 [32]	AP vs HC	13 OTU3	0.61 (0.55–0.66)	34 / 80	16S
	AP vs HC	13 OTU + FIT	0.64 (0.58–0.69)	41 / 80
	AP vs HC	13 OTU + FIT + APCS4	0.71 (0.65-0.75)	46 / 80
Zhang S, 2020 [31]	AP vs HC	5 OTU	0.95 (0.91-0.99)		16S
Yachida, 2019 [19]	Stage 0 vs HC	29 species	0.73		Metagenomics
		16 KO genes	0.75
		29 species + 24 metabolites + 16 KO genes	0.78
Yao, 2021 [30]	Stage I/II vs HC	5 species	N/A	I 62 / -, II 72 / -	qPCR
	Stage I/II vs HC	5 species + FIT	N/A	I 70 / -, II 89 / -
Xie, 2017 [29]	Stage I/II vs HC	Fusobacteria nucleatum	0.59 (0.54-0.64)		qPCR
		Clostridium symbiosum	0.73 (0.68-0.77)
		Clostridium symbiosum + FIT	0.81 (0.75-0.86)
		Clostridium symbiosum and Fusobacteria nucleatum + FIT + CEA	0.82 (0.77-0.87)
Wei, 2022 [28]	AP vs HC	5 OTUs	0.71 (0.63-0.80)		16S
Mo, 2021 [27]	AP vs HC	Fusobacterium nucleatum and Parvimonas micra	0.57	39 / 83	qPCR
		Fusobacterium nucleatum and Parvimonas micra + DNA methylation/mutation5 + FIT	0.73	28 / 94
	Stage I/II	Fusobacterium nucleatum and Parvimonas micra + DNA methylation/mutation + FIT	N/A	I 60 / -, II 85 / -
Liu, 2021 [41]	AP vs HC	Fusobacterium nucleatum	0.74 (0.56-0.91)	71 / 65	qPCR
		pks+ Escherichia coli	0.82 (0.64-0.98)	91 / 75
		Fusobacterium nucleatum, pks+ Escherichia coli + CEA + CA19-9 + FOBT	0.85 (0.57-1.0)	67 / 90
		4 species + FIT	N/A	49 / 81
Liang, 2017 [40]	Stage I/II vs HC	4 species + FIT	N/A	I 77 / -, II 100 / -	qPCR
Liang, 2021 [39]	AP vs HC	4 species	0.67 (0.62-0.72)	39 / 83	qPCR
		4 species + FIT	N/A	49 / 81
Konishi, 2022 [38]	AP vs HC	50 OTUs	0.52 (0.45 – 0.60)		16S
	Stage 0 vs HC		0.54 (0.45 – 0.62)
	Stage I vs HC		0.74 (0.68 – 0.80)
	Stage II vs HC		0.86 (0.81 – 0.90)
Guo, 2018 [37]	Stage I vs HC	3 taxa	0.80	90 / 60	qPCR
	Stage I/II vs HC		0.89	90 / 76
Goedert, 2015 [47]	AP vs HC	5 taxa + 7 genera	0.77		16S
Gao, 2022 [36]	AP vs HC	12 species	0.66 (0.57-0.75)		Metagenomics
	Stage I vs HC		0.88 (0.72-1.00)
	Stage II vs HC		0.96 (0.91-1.00)
Gao, 2020 [35]	AP vs HC	18 genera	0.62 (0.52-0.71)	84 / 39	Metagenomics
		18 genera + FIT	0.72 (0.63-0.81)
Coker, 2022 [34]	AP vs HC	14 species	0.84 (0.80-0.89)		Metagenomics
		14 species + 2 metabolites	0.88 (0.84-0.92)
Ai, 2017 [33]	AP vs HC	6 species	0.87		16S
Rezasoltani, 2018 [43]	AP + Stage I vs HC	5 taxa	0.97	100 / 77	qPCR
Zeller, 2014 [50]	CRC vs HC^	22 species	0.84		Metagenomics
	CRC vs HC^	22 species + FOBT	0.87
Young, 2021 [49]	AP vs HC*	15 taxa	0.72 (0.68-0.75)		16S
	AP vs HC⁎⁎	15 taxa	0.82 (0.79-0.84)
	AP vs HC*	15 taxa + age + sex	0.71 (0.67-0.74)
	AP vs HC⁎⁎	15 taxa + age + sex	0.84 (0.82-0.86)
Tarallo, 2019 [48]	AP vs HC	hsa-miRNA + bsRNA + mbDNA6	0.47		sRNA7/Metagenomics
Flemer, 2017 [46]	AP vs HC	16 OTUs (faecal)	0.87		16S
		12 OTUs (oral)	0.89 (0.80-0.89)	67 / -
		16 OTUs (faecal) + 12 OTUs (oral)	0.98 (0.95-0.98)	88 / 94
Eklöf, 2017 [45]	AP vs HC	clbA+ bacteria, Fusobacterium nucleatum	N/A	47 / 63	qPCR
		clbA+ bacteria, Fusobacterium nucleatum + FIT	N/A	61 / 63
Clos-Garcia, 2020 [44]	AP vs HC	16 genera	0.28		16S
Zackular, 2014 [53]	AP vs HC	5 OTUs	0.84 (0.74–0.94)		16S
		5 OTUs + age + race + BMI	0.90 (0.82–0.98)	90 / 80
Baxter, 2016 [52]	AP vs HC	22 OTUs	0.67	18 / 97	16S
		23 OTUs + FIT	0.76

¹CRC = colorectal cancer, AP = adenoma and/or polyp, HC = healthy controls

²Analytical method: 16S = 16S rRNA gene sequence, qPCR = real-time quantitative PCR, metagenomics

³OTU = operational taxonomic unit

⁴APCS = Asia-Pacific Colorectal Screening, a validated risk-stratification tool

⁵Methylated genes: Septin9, NDRG4, BMP3. Mutated genes: KRAS, BRAF, PI3KCA

⁶hsa = Homo sapiens microRNAs, bsRNA = bacterial small RNAs, mbDNA = microbial DNA

⁷sRNA = small RNA sequencing

^{^}Based on CRC vs HC, but similar changes in microbial abundance observed in early-stages of CRC

^⁎Healthy controls with normal colonoscopy

^⁎⁎Healthy controls, blood negative in stool

Analytical methodology of bacteria-derived biomarkers

Biomarkers in faecal samples for early CRC detection were typically based on the relative abundance of bacteria. Relevant bacteria were either preselected from relevant literature or associations were discovered with selected bacteria between the CRC, precursor lesions and healthy controls. (See Supplementary file S4 for the bacteria and the analytical methods used for the detection of early CRC). For the detection of these bacteria, different techniques were used. 11 articles used 16S rRNA gene sequencing [28,[31], [32], [33], 38,44,46,47,49,52,53], whereas nine used qPCR [27,29,30,37,[39], [40], [41], 43,45] and six used metagenomic shotgun sequencing [19,[34], [35], [36], 48,50]. Universal primers were used in most articles using 16S rRNA gene sequence analysis. However, a large range was seen in the variable region used for the analysis per study, ranging from V1 to V4. For qPCR, different primers were sometimes used (See Supplementary file S4). It is important to note that studies using 16S were only capable of giving bacterial taxa, whereas studies employing qPCR or metagenomics were able to identify or use bacteria on their species level for the detection of early CRC. Furthermore, between the included articles, a high heterogeneity was seen between sample collection methods as well as processing kits for DNA extraction which (see Supplementary Table S2 for a summarised description of sample collection and DNA processing kits per article).

Statistical approaches for the classification models for the prediction of early CRC were based on different models. Besides simple logistic regression models (LRM) [27,29,30,35,[37], [38], [39], [40], [41], 43,45], a random forest model (RFM) was often used [28,31,34,36,[47], [48], [49], 52], followed by LASSO logistic regression models [32,44,50] and Bayesian methods [33,53]. Two studies even used two classification models [19,46]. Only the study by Ai et al. investigated on forehand the best classification model, comparing Bayesian methods, RFM, LRM, F measure and Matthews correlation coefficient, finding the best model based on Bayesian methods [33].

Microbiome-derived biomarkers based on co-metabolome

The metabolome may provide an alternative source of novel biomarkers. The metabolome, defined as the collection of all small low-molecular-weight compounds (<1kDa) in a biological sample, is a time-dependent, multiparametric analysis of the functional output of a biological system and its exposome [55]. The metabolome expression will therefore vary with subtle or early changes in a given disease state [56,57]. The co-metabolome, defined as the set of compounds which are the product of more than one genome interacting in a symbiotic system, is of particular pertinence in the diagnosis of colorectal cancer [55]. Five articles investigated the diagnostic performance of the faecal co-metabolome next to, or in conjunction with, bacteria-derived biomarkers [19,28,34,36,44]. Only two articles were found that solely investigated the diagnostic performance of the co-metabolome for the detection of early CRC [51,54]. The overall performance for precursor lesions based on the co-metabolome reported a range in AUC from 0.69 to 0.84, whereas stage I and II of CRC ranged between 0.65 to 0.93. Chen et al. demonstrated the best performance for precursor lesions and stage I and II of CRC of 0.84 (sensitivity 63 per cent, specificity 85 per cent) and 0.93 (sensitivity 88 per cent, specificity 85 per cent), respectively, based on a model of eight gut-microbiome associated serum metabolites (GMSM-panel) in a matched faeces and serum cohort [54]. Their panel, based on bacterial metabolism, emphasised on secondary bile acids and described an upregulation in serum concentrations of unconjugated primary bile acids (CA) and deoxycholic acid (DCA) in CRC patients (see Supplementary file S4 for all specific metabolites used). Interestingly, the study of Yachida et al. [19] reported an alteration in bile acid metabolism as well, with a significant increase in DCA in patients with stage 0 CRC. These findings suggest a role for bacterial metabolism, especially bile acid metabolism, in early CRC. Four articles described the performance of metabolites compared with bacterial species. Yachida et al. [19] reported an increase in performance, based on AUC, from 0.73 based on 29 bacterial species and 0.65 based on metabolites to 0.78 when combined. Gao et al. [36], interestingly the only study who looked solely at the serum metabolomic profile instead of the faecal metabolomic profile, reported an AUC of 0.91 when combined with bacterial species for the detection of precursor lesions. Coker et al. [34] reported an AUC of 0.69 based on metabolites and an AUC of 0.84 based on 14 bacterial species. They reported an AUC of 0.88 when these 14 species were combined with the two most distinctive metabolites (L-Asparagine and Phenyllactic acid). Finally, the study of Clos-Garcia et al. [44] reported a low AUC of 0.30 when metabolites and bacteria were combined. This study showed the lowest AUC as well based only on bacteria of 0.28. A summary of the results is presented in Table 3.

Table 3.

Co-metabolome as screening tool for early CRC detection.

Reference	Comparison of interest	Biomarker	Performance for early CRC detection		Analytical method1
wAUC (95% CI)	Sens/Spec
Yachida, 2019 [19]	Stage 0 vs HC	24 metabolites	0.65		CE-TOFMS
Wei, 2022 [28]	AP vs HC	4 metabolites	0.73 (0.63-0.82)		UPLC-MS/MS
		4 OTUs + 5 metabolites	0.90 (0.85-0.96)
Gao, 2022 [36]	Stage I vs HC	3 metabolites	0.82 (0.68-0.97)		UPLC-MS/MS
	Stage II vs HC		0.87 (0.79-0.95)
	Stage I/II vs HC		0.85 (0.76-0.93)
	AP vs HC	12 OTUs + 3 metabolites	0.91 (0.83-0.99)
Coker, 2022 [34]	AP vs HC	11 metabolites	0.69 (0.62-0.81)		GC-TOFMS
		11 metabolites + age + gender + obesity	0.75 (0.69-0.81)
		4 metabolites + 6 species	0.75 (0.69-0.81)
Clos-Garcia, 2020 [44]	AP vs HC	16 genera + 6 metabolites	0.30		UHPLC-MS
Chen, 2022 [54]	AP vs HC	8 metabolites	0.84	63 / 85	UPLC-MS
	Stage I/II vs HC		0.93	88 / 85
Bosch, 2022 [51]	AP vs HC	3 metabolites	0.79 (0.64–0.94)	79 / 74	UPLC-MS

¹CE-TOFMS = capillary electrophoresis time-of-flight mass spectrometry, U(H)PLC-MS/MS = Ultra-(High)-Performance Liquid Chromatography Tandem Mass Spectrometry

Analytical methodology of co-metabolomic markers

The methods used to detect metabolites were all based on mass spectrometry (MS), with variations in chromatography techniques. Most used liquid chromatography MS [28,36,44,51,54], and only one used gas-chromatography GS-MS [34], and one applied capillary electrophoresis MS [19]. Most studies investigated the faecal co-metabolome [19,28,34,44,51], with only two studies investigating the serum co-metabolome either solely or in conjunction with the faecal metabolome [36,54].

It has been reported that volatile components of the co-metabolome are more prone to variations in storage conditions and sampling methodologies than the microbiome [58,59]. The sampling methodologies were extracted per study, although not all studies disclosed details of sample collection, showing a large heterogeneity between trial designs (see Supplementary Table S2 for a detailed description). The study of Yachida et al. was the only study to collect faeces on site and immediately stored for analysis under appropriate conditions (dry ice and -80°C for long-term storage) [19]. Two studies let the patient collect the faeces at home and stored in their own freezer, leaving time until long-term storage unclear [44,51].

Risk of bias and applicability

The QUADAS-2 tool was used to assess the risk of bias and applicability. The results are summarised in Supplementary Table S3 as well as presented in Supplementary Fig. S1.

Overall the risk of bias for the first domain, 'patient selection,' was low in 17 studies, high in 10 and unclear in one since most of the studies were cohort studies. For the second domain ‘index test', the risk of bias was very high, with 23 studies high risk, zero low and five unclear. The third domain 'reference test' showed a low risk of bias with zero high, one unclear and 27 low risk. For the last domain 'flow and timing', low risk was seen in 19 studies, high risk in two and unclear in seven studies. Since we preselected articles only applicable to the detection or description of early CRC, all studies showed a high degree of applicability.

Discussion

A total of 28 studies assessed microbiome-derived biomarkers for the detection of early CRC. Despite this, a large degree of heterogeneity was observable in cohort sizes, demographic properties, microbiome sequencing analyses and statistical models used, making a robust interpretation of these data challenging and preventing meta-analysis. Therefore, results were only reported in a descriptive manner, summarising the diagnostic performance of microbiome-derived biomarkers for early CRC detection. The biomarkers identified in these analyses are also yet to be externally validated. Despite this, it appears that the faecal and oral microbiome may have some diagnostic utility for the early detection of CRC.

Diagnostic performance of microbiome-derived biomarkers

The diagnostic performance demonstrated a wide confidence interval across the reported studies for the measurement of the relative abundance of faecal bacteria (0.28 – 0.98 for precursor lesions and a range of 0.54-0.89 for early CRC, based on AUC). Diagnostic performance increased when the bacterial panel was added to existing screening tests, like the gFOBT or FIT, or added to DNA-makers for methylated (Septin9, NDRG4, BMP3) or mutated genes (KRAS, BRAF, PI3KCA), suggesting that bacteria-derived biomarkers could be used to augment current screening programs. Diagnostic performance did not significantly vary in those studies that used small numbers of bacteria as compared to studies using a multitude of bacteria in their panels. This is in contrast to recent literature investigating the microbiome as a biomarker in (all stages of) CRC [60,61]. Many of these studies were not prospective or powered to a primary diagnostic endpoint, or performed in clinically representative populations (e.g. within population demographics typically called for screening).. Moreover, many did not make a distinction between stage or anatomical location (e.g. right or left colon) of CRC in their analysis [62], [63], [64], [65]. Similarly, some studies simply group cancer stages together to produce summary diagnostic statistics [66], [67], [68], [69], [70], [71]. As expected, studies that did compare the performance of the gut microbiome across AJCC different stages often report an increase in performance in stage III and IV cancers [36], [37], [38].

The optimal diagnostic performance across all studies showed an AUC of 0.98, with a sensitivity and specificity of 88 and 94 per cent, and employed an oral mucosa-based detection of bacterial strains [46]. Bacterial taxa statistically associated with CRC, such as Fusobacterium, Peptostreptococcus, Porphyromonas, and Parvimonas, were consistently detected in the oral cavity of those patients with CRC [46]. However, this observation is not CRC-specific, as these bacteria have also been observed to be enriched in other gastro-intestinal (GI) tract cancers, such as oesophageal [42] and pancreatic cancer [72]. Interestingly, these bacteria were however abundant in precursor lesions for bowel cancer. This is important as colonic adenoma detection rates range from 26 to 47% in the screening populations [73] and prevalence rates of cancer within these polyps have been reported in 0.9–2.8 % of polyps ≤5 mm and 5.3–15.5 % of polyps between 6 and 9 mm [74]. The risk of adenoma progression to cancer is dependent on their histological subtype and anatomical location (right vs left) but also on environmental factors (e.g. diet or obesity) that are co-regulated by the gut microbiome [75]. A recent prospective analysis of 231 patients undergoing screening colonoscopy analysed faecal swabs using 16S RNA and demonstrated a classification accuracy for adenomas >75% for Naïve Bayes and Neural Network models using informative OTUs [76]. Previous analysis of faecal samples by 16 sRNA of patients with conventional adenomas is depleted in a network of Clostridia operational taxonomic units from families Ruminococcaceae, Clostridiaceae, and Lachnospiraceae and enriched in the classes Bacilli and Gammaproteobacteria, order Enterobacteriales, and genera Actinomyces and Streptococcus [15].

Of all the screened articles, only seven investigated the potential of the co-metabolome as a diagnostic tool for early CRC detection. A similar effect was reported with an AUC ranging from 0.69 – 0.84 for precursor lesions and 0.65 – 0.93 for stages I and II of CRC. However, there were insufficient articles on the efficacy of the co-metabolome for early CRC detection to form any proper conclusions. Oncogenic co-metabolites produced by luminal and mucosal gut microbiota (e.g. bile acids) are over-abundant in those populations at high risk of colonic cancer, while beneficial microbial co-metabolites, like short-chain fatty acids (SCFAs), are often decreased [76,77]. Changes in metabolic profiles may therefore reflect an early pro-oncogenic environment making them valuable in screening programs, however this hypothesis is yet to be tested at scale.

Currently, most research regarding the investigation of a microbiome-derived biomarker for (early) CRC is based on differences in taxonomic variation. However, some studies suggest a higher sensitivity for the detection of CRC based on bacterial genes [78]. A direct comparison of bacterial taxa vs bacterial genes for the detection of CRC has been recently investigated by Norouzi-Beirami et al. [79]. This study noticed a small increase in sensitivity and specificity for the detection of CRC based on genes in comparison to bacterial taxa. However, the increase was almost gone when tested in an independent cohort. Superior accuracy was however detected based on functional features (AUC 0.71) compared to taxa or gene level (with a corresponding AUC of both of 0.59). Most studies have been referring to bacterial taxonomy because this is the most standardised technique in this field, mainly based on only one bacterial gene, the 16S rRNA gene. To use other bacterial genes as biomarkers requires standardised procedures and bioinformatic pipelines as well as the establishment of universally usable primers for these specific genes, which are also very challenging at this specific moment considering the advances in whole genome shotgun sequencing. Nevertheless, one major limitation of taxonomic identification as well as gene level, especially in regard to early CRC, is that it does not give any insight into, or is necessarily associated with, functional variation of the genetic content. When investigating a more accurate and sensitive biomarker for the prediction of early CRC, it is the functional level rather than bacterial taxa or genes that could provide insight into microbiome changes associated with the development of CRC and therefore be used as a possible biomarker for the detection of early CRC.

Limitations of microbiome-derived biomarkers

The gut microbiome demonstrates significant inter-individual variation and macro and microscopic anatomical heterogeneity, governed by factors such as age, gender, BMI, diet and antibiotic use. When comparing healthy controls with colorectal patients, these factors should ideally be incorporated in the analyses via matching of the groups of interest. Unfortunately, most of the included articles only reported possible influencing factors on the microbiome composition without any matching between the two groups. Pilot studies in CRC have also identified significant variation between cancers [80], in part because the gut microbiome is dynamic and evolves with the pathology, in part because it is subject to confounding environmental factors, such as diet, medication, smoking and other lifestyle factors [81]. These were rarely accounted for in the analyses, and most were often not reported. The heterogeneity in the diversity of the faecal microbiome exists at both a phyla and strain level between geographically discrete populations and across countries [82,83]. The majority of included articles were performed in Asia, potentially limiting the translation to other regions, and it is unclear therefore how many of these studies are translatable.

One other major limitation in these studies has been the lack of standardisation in both sample collection and processing. Sample collection showed a wide heterogeneity between studies (e.g. before or after bowel-cleansing agents required for the colonoscopy) as well as temperature and time until long-term storage at -80°C. These factors are known to influence microbiome composition [84,85]. Furthermore, DNA extraction was carried out using a range of DNA extraction kits, adding an additional variable influencing the outcome of microbiome composition (see Supplementary Table S2 for a summary of sample collection and DNA processing of the included studies). A lack of standardisation was also seen in the analytical approaches employed by the included studies. All studies used different analytical methods to identify bacteria, either 16S rRNA, qPCR or metagenomics. 16S rRNA analysis is largely restricted to taxonomic analysis with an inferred functional interpretation. Metagenomics permits species-level identification and deep functional insights, bypassing some of these challenges [86], although it is not yet affordable for population-level analyses and its interpretation requires significant computing power. Within these methods, there was again a large heterogeneity in the choice of primers used, leading to potential biases and making comparisons between them almost impossible [87].

Research priorities and the ideal microbiome biomarker study

If the microbiome is to be realistically mined for its biomarker potential in CRC, there is an urgent need to move away from small-scale pilot studies into prospectively, adequately powered, multicentred trials that are quality assured from a sampling and analytical methodology. Currently, there is a lack of consensus on the most robust method of either qualitatively or quantitively analysing the gut microbiome for early cancer detection in terms of its function, absolute or relative abundance, its diversity or species richness. Given the significant inter-individual variation in the global microbiome and its evolutionary instability within and on CRC, it is not yet clear if the microbiome is a viable biomarker for early cancer. Therefore future studies must seek to test biomarkers that are mechanistically linked to established molecular pathways and which are established to exist across vulnerable populations. However, this is potentially where the microbiome has its greatest potential, as it can provide future clinicians with noninvasive, actional information on the anatomical location or progression of the disease.

Conclusion

Gut microbial-derived biomarkers could be leveraged to enhance current screening programs for CRC. However, significant barriers must be overcome before this can be achieved. Current analyses cannot be meta-analysed due to large observed variations in study design and analytical precision. Future CRC microbiome studies therefore require precision and must account for environmental and sampling confounding and bias, and provide methodological quality assurance and establish cost-effectiveness.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Supplementary materials

Supplementary material of this systematic review can be found in the online version. The following supplementary material of this systematic review can be found in the online version. Fig. S1. QUADAS-2 overview, Table S1. Search strategy, Table S2. Sample collection and DNA processing, Table S3. QUADAS-2 results, Excel file S4. Additional characteristics of analytical method and biomarker used.

CRediT authorship contribution statement

Florine H. Zwezerijnen-Jiwa: Conceptualization, Writing – original draft. Hugo Sivov: Data curation. Petra Paizs: Writing – review & editing. Konstantina Zafeiropoulou: Visualization, Writing – review & editing. James Kinross: Supervision, Writing – review & editing.

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.