Faecal microRNAs as a non-invasive tool in the diagnosis of colonic adenomas and colorectal cancer: A meta-analysis

MicroRNAs (miRNAs) are proposed as potential biomarkers for the diagnosis of numerous diseases. Here, we performed a meta-analysis to evaluate the utility of faecal miRNAs as a non-invasive tool in colorectal cancer (CRC) screening. A systematic literature search, according to predetermined criteria, in five databases identified 17 research articles including 6475, 783 and 5569 faecal-based miRNA tests in CRC, adenoma patients and healthy individuals, respectively. Sensitivity, specificity, positive/negative likelihood and diagnostic odds ratios, area under curve (AUC), summary receiver operator characteristic (sROC) curves, association of individual or combinations of miRNAs to cancer stage and location, subgroup, meta-regression and Deeks’ funnel plot asymmetry analyses were employed. Pooled miRNAs for CRC had an AUC of 0.811, with a sensitivity of 58.8% (95% confidence interval [CI]: 51.7–65.5%) and specificity of 84.8% (95% CI: 81.1–87.8%), whilst for colonic adenoma, it was 0.747, 57.3% (95% CI: 40.8–72.3%) and 76.1% (95% CI: 66.1–89.4%), respectively. The most reliable individual miRNA was miR-21, with an AUC of 0.843, sensitivity of 59.3% (95% CI: 26.3–85.6%) and specificity of 85.6% (95% CI: 72.2–93.2%). Paired stage analysis showed a better diagnostic accuracy in late stage CRC and sensitivity higher in distal than proximal CRC. In conclusion, faecal miR-21, miR-92a and their combination are promising non-invasive biomarkers for faecal-based CRC screening.

Subgroup meta-regression analysis in colonic adenoma was also performed, looking at differences in sample size, pooled individual/combination miRNAs and quantitation method ( Table 4). The meta-regression indicated that a small sample size significantly affected both the specificity (Z-value: −2.011, P = 0.044) and sensitivity (Z-value: −1.309, P < 0.001). With respect to the pooled individual/combination miRNAs, meta-regression analysis did not show significant effects in both the pooled sensitivity (Z-value: −10.85, P = 0.853) and specificity (Z-value: −0.194, P = 0.846). For the qPCR relative quantitation method, a significant effect was observed in the sensitivity (Z-value = −3.356, P < 0.001) of pooled miRNAs. Meta (Table 6 and Suppl. Fig. 2). the detection accuracy of individual microRNAs. Each individual miRNA reported by more than one research group was pooled for an accuracy estimation (Table 6). miR-21 was reported by five CRC and three colonic adenoma studies 11,16,18,20,22 (Table 6 and Suppl. Fig. 2). publication bias evaluation. The Deeks' funnel plot asymmetry test was utilised to evaluate the potential publication bias from each included faecal-based miRNA study. The slope coefficient was associated with P = 0.03 in the pooled miRNAs in CRC, and P = 0.61 in pooled miRNAs in colonic adenoma studies, indicating that significant asymmetry was found in the CRC dataset (Fig. 6A) but not in the colonic adenoma dataset (Fig. 6B). The combination of CRC and colonic adenoma resulted in a P = 0.11 (Suppl. Fig. 3).

Discussion
To evaluate the diagnostic value of faecal-based miRNAs, data from 17 eligible publications, including 46 studies on miRNAs in CRC and 10 studies on colonic adenoma, corresponding to 6475, 783 and 5569 faecal-based miRNA tests in CRC patients, adenoma and healthy control volunteers, respectively, were subjected to meta-analysis. Our study reveals that pooled faecal-based miRNAs have a relatively high detection accuracy for  Table 1. The circular regions (95% confidence contour) contain likely combinations of the mean value of sensitivity and specificity. Sen = sensitivity; Spe = specificity SOP, summary operating point. # Early stage CRC includes TMN stages 0 + I + II or Dukes' stage A + B; late stage CRC includes CRC stages III + IV or Dukes' stage C + D. Proximal CRC is defined as from cecum to transverse colon, and distal CRC is defined as from the splenic flexure to the rectum.
www.nature.com/scientificreports www.nature.com/scientificreports/ CRC. However, the lack of consensus regarding the optimal quantitation method, data normalisation, and selection of control subjects, may present obstacles to clinical application.
qPCR-based studies were the subject of our meta-analysis analysis. This approach for miRNA level quantification is advantageous compared to others in that is fast and easily adoptable in a clinical setting. However, it comes with limitations that relate to the method for miRNA isolation and the selection of the appropriate reference/normalisation control. Although reference quantitation method demonstrated a better diagnostic accuracy compared to absolute quantitation, it is important to acknowledge that a variety of internal controls were used as references, including RUN6B(U6) 11,[15][16][17][18]24 , miR-24 23 , miR-200b-3p 19 , miR-378 12 , miR-1202 27 and miR-4257 27 . Increasing evidence suggests that RUN6B may not be a suitable endogenous control for miRNAs 31,32 due to its rapid degradation in faeces 20 . miRNAs used as internal controls also have functions in the host cell, and their deregulation could interfere with the detection accuracy. For example, miR-24, a proposed tumour suppressor miRNA in CRC, controls cellular proliferation independently of p53 by targeting the 3'UTR of dihydrofolate reductase (DHFR) mRNA 33,34 . Deregulation of plasma miR-378 was also found in CRC patients 35 . miR-4257 has been reported to be down-regulated in bladder cancer cell lines and up-regulated in the plasma of patients with recurrence of non-small cell lung cancer (NSCLC) 36,37 . Peripheral levels of miR-1202 predicts and mediates the response to anti-depressants, specifically regulating the expression of metabotropic glutamate receptor-4 (GRM4) with levels correlating to changes in brain activity [38][39][40] . miR-1202 is deregulated in different types of cancers, such as breast cancer 41 , gastric cancer 42 and clear cell papillary renal cell carcinoma 43 . Absolute quantitation was employed in several studies for faecal-based miRNA screening 13,14,20,26 (Table 1), however, this necessitates a standard curve which depends on the quantification detection method and does not eliminate potential contamination by gut bacteria DNA/RNA 14 .
The combination of FIT and stool-based miRNA markers may increase detection accuracy to overcome this problem. A previous study indicated that the combination of miR-21 and miR-92a with FIT had a specificity of 96.8% and sensitivity of 78.4% while FIT alone only had a specificity of 98.4% and sensitivity of 66.7% 22 . A parameter that should be considered is the presence/absence of occult blood in samples as miRNAs expressed in blood cells may interfere with the assay, altering the levels of specific miRNAs. In an effort to assess the potential contribution of blood in faecal miRNA levels we have retrieved a list of circulating miRNAs 44 . Comparison showed that 8 miRNAs are detected in both blood and faecal specimens (Suppl. Fig. 4). This finding does not imply that blood cells are responsible for the alterations in the levels of these miRNAs, as their origin may as well be the tumour. Optimally, a controlled study including comparisons between samples positive and negative for FOBT/ FIT could address the relative contribution. A more inclusive approach employing miRNA analyses and comparisons between matched blood, faecal specimens and tumours or colonic tissues would be most informative about the source of changes in miRNA levels. Furthermore, other colonic pathologies like inflammatory bowel diseases are characterised by deregulation of miRNAs detectable in tissues and serum [45][46][47][48] and the presence of occult blood in faeces. A comprehensive analysis would include samples from different pathologies of the colon, assess and identify disease-specific miRNA signatures and their diagnostic/prognostic properties.
A clear conclusion on which quantitation method is more suitable cannot be drawn with the currently available data. The use of multiple internal controls or the geometric mean, using a multiplex screening method, such as microarrays or next generation sequencing, would provide the optimal means of normalisation. Alternatively, the NanoString nCounter technology enables profiling of around 800 molecular targets in one single reaction by utilising molecular "barcodes". This approach normalises the data by using multiple targets, and more importantly quantifies multiple miRNAs which can be used simultaneously as biomarkers to improve detection accuracy. In addition, this platform overcomes the need for data processing and bioinformatic analysis expertise, as in the case of microarrays or high-throughput sequencing, thus may be easily utilised in a clinical setting 46 .
To evaluate the potential detection efficiency for each individual miRNA, individual miRNAs reported in more than one study were grouped to evaluate its detection accuracy. In this meta-analysis, miR-21 and miR-92a were the most commonly reported faecal-based miRNAs ( Table 6). Numerous studies have characterised the functional roles of these two miRNAs in CRC pathogenesis and aggressiveness. Up-regulation of miR-21 and miR-92a promotes CRC cell migration, invasion and proliferation 11,16,18,20,22 , and inhibition of apoptosis [49][50][51] . Several significant targets of miR-21 are associated with CRC malignancy, such as phosphatase and tensin homolog (PTEN) 49 www.nature.com/scientificreports www.nature.com/scientificreports/ programmed cell death protein 4 (PDCD4) 53,54 , and ras homolog gene family member B (RhoB) 55 . Among these, PTEN was reported frequently silenced in CRC by miR-21, resulting in PI3K/AKT pathway activation and induction of tumour formation 49,52 . Recently, a long non-coding RNA (LINC00312) suppressed in CRC was shown to regulate miR-21 levels through its function as a miRNA sponge, thereby regulating PTEN expression 56 . miR-92a has been shown to disrupt the expression of several tumour suppressors such as PTEN 57,58 , Dickkopf WNT Signalling Pathway Inhibitor 3 (DKK3) 57 , Kruppel-like factor 4 (KLF4) 59 and mothers against decapentaplegic homolog 7 (SMAD7) 60 . Hence, miR-92a activates the PI3K/AKT, WNT/β-catenin and BMP/Smad pathways and enhances  Table 6. Diagnostic accuracy of individual microRNAs and microRNA combinations. DOR, Diagnostic odds ratio; +LR, Positive likelihood ratio; −LR, Negative likelihood ratio. and colonic adenoma (n = 2). The number next to the dot/triangle corresponding to the study ID in Table 1 (Blue dots: CRC) or www.nature.com/scientificreports www.nature.com/scientificreports/ tumorigenesis. Subject to this analysis five studies reported the use of miR-21 in the identification of CRC, and three studies reported its use in identification of adenomas 11,16,18,20,22 . Four studies reported the utility of miR-92a in the identification of CRC, and two studies in identification of adenomas 11,17,20,22 . miR-21 had a better detection accuracy range compared with miR-92a, with a DOR of 9.28 (95% CI: 2.97-28.97) and summary AUC of 0.843. Panels including a combination of either miR-21 or miR-92a, as well as panels including both miR-21 and miR-92a demonstrated a small improvement in detection ( Fig. 5 and Table 6). However, due to the small number of published studies, with each having wide confidence intervals, a direct comparison between two faecal-based miRNAs may not be accurate. Additional data are needed to limit potential errors.
The FOBT or FIT, have limited sensitivity for detecting proximal compared with distal CRC 61,62 . This is due to the degradation of haemoglobin. Hence, tumour location analysis for faecal-based miRNA detection was also considered and reported by several studies -with none of them reporting a statistical difference. In this study, the results between pooled miRNAs for proximal and distal CRC reveal differences associated with tumour location, with an AUC of 0.719 versus 0.818, and DOR of 3.44 (95% CI: 2.53-4.66) versus 8.51 (95% CI: 4.97-14.57) (Fig. 4B and Table 5).
Our study is characterised by many strengths but should be interpreted in the context of specific shortcomings. Firstly, subgroup analysis suggested that the combination of faecal miRNAs exhibited a good accuracy for CRC and colonic adenoma patients screening (Tables 3, 4 and Fig. 3). However, certain combinations of miRNAs may not significantly improve the detection accuracy. For example, the panel containing miR-223, miR-92a, miR-16 and miR-106b had a sensitivity of 73.9%, specificity of 82.2% and AUC of 0.84 13 , whereas the combination of miR-18a and miR-135b only had a sensitivity of 66%, specificity of 72% and AUC of 0.75 26 . Therefore, an optimal miRNA combination panel should be prioritised. Secondly, the majority of studies were performed in East Asia (Hong Kong, Taiwan, China, Japan and Singapore) ( Table 1) with only one study in the USA, Europe and the Middle East, making it unclear whether the ethnic background of participants has an influence on the expression of miRNAs in CRC. Thirdly, due to the high cost of colonoscopy, the majority of test subjects were recruited from the corresponding clinics. This may result in a degree of bias, since the subjects are not representative of the general population. Last but not least, the publication bias analysis revealed that pooled miRNAs in CRC have a significant asymmetry (P = 0.03). This may be due to file-drawer effects, bias from the studies with small same sizes, lack of clarity in reporting the results for some publications, or the level of detail provided being lower than the one required for our analysis. Consequently, some studies were excluded, resulting in a possible bias in our meta-analysis (Fig. 6A).
In conclusion, faecal-based miRNAs show a relatively high accuracy for the non-invasive detection of colonic adenomas and CRC in the studied population. The use of a panel of miRNAs as biomarkers may result in a higher CRC detection rate, while the combination of miRNA biomarkers with FOBT or FIT may increase the detection accuracy. Large, ideally multi-centre, double-blinded randomised controlled trials are needed to establish the value of miRNAs as biomarkers in CRC screening within the general population.

Methods
overview. The study protocol followed the Cochrane Handbook for Diagnostic Test Accuracy Review 63 and the Preferred Reporting Items in Systematic Reviews and the Meta-Analysis statement (PRISMA) 64 . Investigators of each of the original studies obtained approval from their local ethics committee and had written, informed patient consent.
Literature search strategy. The search strategy was designed to identify any studies describing the diagnostic value of faecal-based miRNA for CRC and colonic adenoma patients. After an initial search for articles in PubMed, assessments of key terms within the title and abstract were conducted. A full systematic search using the established key terms was adopted for the following databases: PubMed, Ovid Embase, The Cochrane Library, Scopus and Web of Science. The search terms used were "miRNA OR microRNA OR miR" AND "colorectal www.nature.com/scientificreports www.nature.com/scientificreports/ cancer OR colorectal tumor OR colorectal adenocarcinoma OR colorectal carcinoma OR colorectal neoplasm OR colon cancer OR colonic adenoma OR colonic adenocarcinoma OR stomach cancer OR rectal cancer OR CRC" AND "stool OR feces". The search formulas are available as supplementary data (Supplementary data 1). Manual searching of related citations and reference lists was undertaken. Book chapters, letters to editors, commentaries, editorials, patents, and non-peer reviewed articles were excluded. Two investigators independently screened the search results, initially through articles' title and abstract. The filtered candidate articles were then scrutinised independently through full-text reading. Discrepancies were resolved through discussion between the two investigators. study selection criteria. All research articles in any language published up to November 17, 2017 were eligible for inclusion. An electronic data extraction form was developed, and pre-tested, with data extracted by two researchers. Eligible studies included in this meta-analysis adhered to the following criteria: (1) studies evaluated the diagnostic value of miRNAs for detecting human CRC or colonic adenomas; (2) all CRC patients involved in the study had been confirmed by histology; (3) studies contained data on miRNAs' sensitivity, specificity, and sample size to enable reconstruction of the diagnostic 2 × 2 contingency table. Exclusion criteria were set as follows: (1) duplicated studies, the later ones were excluded; (2) publications that were unrelated to the diagnostic value of miRNAs for CRC; (3) incomplete data reporting. The detection accuracy of miRNAs between proximal (from cecum to transverse colon) and distal (from splenic flexure to rectum) CRC, as well as between early (CRC stages 0 + I + II or Dukes' stages A + B) and late (CRC stages III + IV or Dukes' stages C + D) stage CRC were evaluated separately if investigators reported the location and stage. Each individual miRNA and miRNA combinations were grouped together if found in more than two studies.

Risk of bias.
The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) was utilised to assess the quality of included publications, evaluating four key domains ("Patient Selection", "Index Test", "Reference Standard", and "Flow and Timing") in two categories (risk of bias and applicability of diagnostic accuracy studies) 30 . Each category in all publications was judged as low, high or unclear based on the assessment criteria provided. Assessment of each included study was performed by two investigators, with disagreements resolved by consensus after discussion.

Data synthesis.
A meta-analysis of diagnostic test accuracy was conducted on faecal-based non-invasive miRNA tests through a bivariate random effects modelling approach. The bivariate model accounts for the correlation between the studies' sensitivity and specificity in two different levels. The first level represents a variability between sensitivity and specificity within one study; the second level represents the heterogeneity in diagnostic performance of the index test across the testing studies. Random effects meta-analysis methods were applied in our study as heterogeneity is presumed to exist.
Statistical analyses in this study were performed using the statistical package mada version 0.4.8 in R (version 3.4.3) to implement the bivariate normal approach of Reitsma et al. 65 . Sensitivity, specificity, diagnostic odds ratio (DOR), log DOR, positive likelihood ratio (+LR) and negative likelihood ratio (-LR) were calculated along with their 95% confidence interval (95% CI) based on the random effects model (DerSimonian and Laird method) with continuity correction 66,67 . Summary receiver operating characteristic (sROC) curves, area under the curve (AUC) and partial AUC were also utilised to examine the pooled faecal-based miRNAs in CRC, adenoma and the subgroups. Potential sources of heterogeneity were investigated using subgroup and bivariate meta-regression (restricted maximum likelihood (REML) estimators) analysis. The Deeks' funnel plot asymmetry test was examined using the midas package in Stata (version 12).