Included studies and their characteristics

Figure 1 shows the study selection process. A total of 5168 potentially relevant records were identified from PubMed, ScienceDirect and Web of Science. After removal of duplicate publications and non-research articles such as reviews, 1218 records remained, of which 284 were identified for full-text assessment. During the full-text assessment, 128 studies were excluded, for example for not reporting cut-off criteria, sample sizes or direction of dysregulation. As a result, 156 studies met the eligibility criteria for meta-analysis, with a combined sample size of >15,000. Of the 156 eligible studies (listed in electronic supplementary material [ESM] Appendix), most reported only type 2 diabetes microRNA expression profiles in humans; approximately 30 studies were based on both individuals with type 2 diabetes and animals models of diabetes. For the present meta-analysis, only human data were used. Details of the study characteristics are shown in ESM Table 1.

Fig. 1

Flow diagram of study selection. T2D, type 2 diabetes

Quality assessment

The MIAME guideline 2.0 [19], MINSEQE guideline and MIQE guideline [20] were used to assess study quality. Figure 2 shows the results of the quality assessment process, mainly according to the domains of the MIAME guideline. The detailed quality assessment of the individual studies is shown in ESM Table 2. The overall assessment found that 85% of the included studies did not report raw data on hybridisation, which was rated as high risk, and only 17% and 23% of studies provided sufficient information (for replicability and reproducibility) about annotation of array design and experiment design, respectively (Fig. 2).

Fig. 2

Quality assessment according to the MIAME guideline

Meta-analysis of differentially expressed microRNAs

The outcomes from REML estimation of DEMs are presented in the main text and ESM Tables 3–21, while the outcomes from EB estimation are presented in ESM Tables 22–35. Of the 404 DEMs reported in the 156 studies that compared type 2 diabetic samples with non-diabetic control samples, 205 (51%) were reported in at least two substudies. Among the 205 DEMs, meta-analysis identified 60 statistically significant dysregulated DEMs (i.e. 31 upregulated and 29 downregulated), as shown in ESM Table 3; the remaining 145 DEMs were of no statistical significance (adjusted p>0.05). The most frequently reported upregulated microRNA was miR-320a, which was reported in 14 substudies (logOR 5.2885, 95% CI 2.2857, 8.2914; adjusted p=4.96×10–2). The most frequently reported downregulated microRNA was MiR-30c-5p (logOR 7.4205, 95% CI 5.7924, 9.0485; adjusted p=3.68×10–17), which was reported in six substudies.

Subgroup analysis

A total of 118 of 156 studies investigated circulating microRNAs in plasma, serum, peripheral blood mononuclear cells (PBMCs) or whole blood. Seven studies investigated skeletal muscle tissue, five pancreatic tissue, seven adipose tissue, five blood vessels, five heart, three foot skin, two vitreous of the eye, two kidneys, one epithelial breast cancer tissue, one bone cells, one liver and two gingival crevicular fluid. Among the five pancreatic profiling studies, four studies used pancreatic islets and one used pancreatic tissue. Details are provided in ESM Table 1. Statistically significant dysregulation of microRNAs in different tissue types is shown in ESM Tables 4–9. In total, one statistically significant microRNA was found in the pancreas, one in the kidney, two in the heart, five in skeletal muscle, six in adipose tissue and 50 in blood. No statistically significantly dysregulated microRNAs were identified in foot skin, vitreous and gingival crevicular fluid. In addition, no statistically significantly dysregulated microRNAs were identified in multiple tissues.

In subgroup analysis of blood fractions, 15 studies extracted RNA from PBMCs, 39 studies used serum as the RNA source, 43 studies focused on plasma RNA and 26 studies used whole blood as the RNA source. Subgroup analysis identified 12, 38, 25 and 25 statistically significantly dysregulated microRNAs in PBMCs, serum, plasma and whole blood, respectively (ESM Tables 10–13). ESM Table 14 shows that 87 statistically significant microRNAs were identified from the four RNA sources. Of these, 76 microRNAs were identified in only one of the four RNA sources; 10 microRNAs (e.g. miR-125b-5p and miR-130b-3p) were upregulated in one source but downregulated in another; and one microRNA (miR-150-5p) was upregulated in both whole blood and serum.

In total, 150 studies detected microRNAs using PCR-based methods, three studies used sequencing technologies, two studies screened for microRNAs using sequencing technologies and validated the results using PCR-based methods, and one study used the NanoString assay. Subgroup analyses of microRNAs detected using PCR-based methods and sequencing technologies identified 61 and 11 statistically significantly dysregulated microRNAs, respectively, which are shown in ESM Tables 15 and 16, respectively. Two microRNAs (miR-144-3p and miR-30b-5p) were upregulated in PCR-based studies but downregulated in sequencing-based studies.

Sensitivity analysis

Sensitivity analysis was conducted to examine the robustness of the findings. We first excluded those studies with sample sizes <25, and then further excluded studies whose sample sizes were <50, after which 114 and 78 studies remained, respectively. Analysis of the 114 and 78 studies identified 53 and 37 microRNAs, respectively, that were significantly differentially expressed (ESM Table 17). Some microRNAs were statistically significantly dysregulated both in sensitivity analysis and in the overall analysis, whereas others were not. For example, miR-93-5p was statistically significant in sensitivity analysis but not in the overall analysis, as several studies with small sample sizes (<25) but large effect sizes were excluded in sensitivity analysis. ESM Table 17 shows that the number of significant microRNAs decreased when the sample size increased. These data indicate that the small sample sizes used in microRNA profiling studies may explain some of the differences seen in the results.

Publication bias

Funnel plots, Begg’s tests and Egger’s tests were performed to evaluate publication bias. The results of the analysis of publication bias for the top three most reported microRNAs according to the number of studies (miR-126-3p, miR-15a-5p, miR-155-5p) and the top three most reported upregulated microRNAs and the top three most reported downregulated microRNAs (according to both the number of studies and the statistical significance) are presented in the main text. The four substudies of the most reported downregulated microRNA, miR-593, were part of the same study, which does not fit the models in Egger’s and Begg’s tests; therefore, only two of the top three most reported downregulated microRNAs were tested and a total of eight microRNAs are presented in Table 1. Typical funnel plots are presented in ESM Fig. 1, showing various levels of asymmetry across the studies and indicating some publication bias in the case of miR-126-3p, miR-320a, miR-29a-3p, miR-29c-3p and miR-30c-5p. Begg’s tests and Egger’s tests confirmed the statistical significance of the publication bias in miR-126-3p, miR-320a, miR-29a-3p, miR-29c-3p and miR-30c-5p (Table 1).

Table 1 Results of Begg’s and Egger’s tests MicroRNA pathway modelling and biomarker selection

A total of 384,579 microRNA–target interaction pairs were identified from the miRTarBase, miRecords and TransmiR databases (ESM Table 18). After alignment of microRNA names to avoid duplication, 382,633 microRNA–target interaction pairs remained. After filtering by the number of supporting articles, the top 1.4% (108,812/7,765,056) of prediction pairs from TargetScanHuman and the 1203 articles on microRNAs from Retraction Watch, 3290 robust interaction pairs were identified from 1966 articles (ESM Table 36). The 3290 interactions were further combined with KEGG pathways, producing 225 microRNA-regulated pathways (see https://osf.io/e9v7f). Extensive meta-analyses identified 138 statistically significantly dysregulated microRNAs, of which 124 microRNAs were dysregulated in a consistent direction in various meta-analyses. Pathway analysis found that the 124 dysregulated microRNAs were statistically significantly enriched in type 2 diabetes-related pathways (ESM Table 19), such as diabetic cardiomyopathy, insulin resistance, advanced glycation end products (AGE)/receptor for advanced glycation end products (RAGE) signalling-mediated diabetic complications and the type 2 diabetes pathway. The priority verification order (according to the following order of importance: [1] detectable in blood or blood fractions; [2] statistically significance in different analyses) for the 16 microRNAs enriched in the type 2 diabetes pathway and meeting the criteria for biomarker selection (i.e. statistically significant and biologically relevant) is as follows: miR-29a-3p, miR-221-3p, miR-126-3p miR-26a-5p, miR-503-5p, miR-100-5p, miR-101-3p, mIR-103a-3p, miR-122-5p, miR-199a-3p, miR-30b-5p, miR-130a-3p, miR-143-3p, miR-145-5p, miR-19a-3p and miR-311-3p (ESM Table 20).