Baseline Assessment of Circulating MicroRNAs Near Diagnosis of Type 1 Diabetes Predicts Future Stimulated Insulin Secretion

Type 1 diabetes is a chronic autoimmune disease leading to progressive (yet heterogeneous) loss and dysfunction of pancreatic β-cells (1). The TrialNet organization has completed clinical trials in participants with newly diagnosed type 1 diabetes (<100 days) and stimulated peak C-peptide >200 pmol/L. A 2-year follow-up of 191 participants in placebo groups demonstrated greater C-peptide loss during the 1st year postdiagnosis (2), yet 88% and 66% had stimulated peak C-peptide ≥200 pmol/L 1 year and 2 years after diagnosis, respectively. Several agents preserve insulin secretion in individuals with newly diagnosed type 1 diabetes, in some cases for up to 2–7 years after treatment (3). Novel biomarkers to stratify trial participants at baseline and predict decline of their insulin secretion would afford gains in trial design and efficiency, facilitating the identification of effective therapies.

MicroRNAs (miRNAs) are small, noncoding RNAs that regulate gene expression (4) and are emerging as disease biomarkers. Circulating miRNAs are stable and measurable in serum and plasma with similar results (5). Twenty-nine circulating miRNAs were associated with type 1 diabetes by two to eight studies (6–24) (Supplementary Table 1), suggesting reproducible associations despite heterogeneity in study design, cohorts, assays, and analysis methods. Cellular miRNAs were also linked to both human and experimental diabetes (23,25–30).

An outstanding question is whether circulating miRNAs predict C-peptide decline after diagnosis. Samandari et al. (12) assessed plasma levels using Exiqon RT-PCR assays for 179 miRNAs and found that 3-month visit plasma levels of several miRNAs (miR-24-3p, miR-146a-5p, miR-194-5p, miR-197-3p, miR-301a-3p, and miR-375) correlated with residual β-cell function at the 6- and 12-month visits; miR-197-3p levels at the 3-month visit predicted β-cell function 12 months after diagnosis. Garavelli et al. (24) reported that miR-23a-3p, miR-23b-3p, miR-24-3p, miR-27a-3p, and miR-27b-3p predicted fasting C-peptide loss of <10% or >90% 12 months after diagnosis. In other studies, Let-7g was associated with C-peptide levels during the 1st year postdiagnosis (19), levels of the β-cell–enriched miR-204 correlated with C-peptide area under the curve (AUC) at diagnosis (31), and levels of miR-142-5p, miR-29c-3p, and miR-320 differed in children with recent-onset type 1 diabetes according to residual fasting C-peptide (23).

We assessed levels of 2,083 miRNAs in baseline serum samples from 53 participants with newly diagnosed type 1 diabetes randomized to placebo groups in TrialNet clinical trials. We report miRNA associations with stimulated C-peptide, which are maintained at the 12-month visit and stratify participants for future severity of C-peptide loss.

We examined baseline serum samples from 53 individuals with newly diagnosed type 1 diabetes who were enrolled in TrialNet clinical trial placebo groups <100 days from diagnosis and with stimulated peak C-peptide >200 pmol/L. Table 1 shows their baseline characteristics. Subjects were included in this study on the basis of availability of a fasting serum sample that could be used for miRNA assessment obtained on the day of the baseline mixed meal tolerance test (MMTT). Supplementary Fig. 1 illustrates C-peptide AUC and peak levels during the 2-h MMTTs performed at the baseline, 6-month, and 12-month visits (2). The baseline MMTT was performed within an average ± SD of 1.9 ± 0.1 months from diagnosis, or 58 days.

The MMTT was described previously (32). Serum C-peptide levels were measured using a TOSOH AIA-900 Automated Immunoassay Analyzer. The trapezoidal rule was used to calculate the C-peptide AUC in nmol/L; peak levels are reported in pmol/L (2).

We tested 15-μL serum aliquots. Samples were collected and processed according to TrialNet protocols (Supplementary Material). miRNAs were assayed using the HTG EdgeSeq miRNA Whole Transcriptome Assay (HTG Molecular Diagnostics) (33), which combines a quantitative nuclease protection assay with next-generation sequencing. It does not require miRNA isolation, reverse transcription, adenylation, or ligation, which could introduce bias. The assay has a broad dynamic range with high reproducibility, sensitivity, and specificity (33). Testing of blind replicate samples from four individuals with type 1 diabetes provided by the JDRF Biomarker Working Group Core for Assay Validation confirmed excellent reproducibility (r = 0.94–0.96; coefficient of variation 1.06%). Undiluted RNA is bound to corresponding target-specific nuclease protection probes after treatment with lysis buffer. The probe set contains complementary sequences for 2,083 specific miRNAs or ∼78% of the published mature transcripts in miRBase V22 (34). Probes hybridized to cognate miRNAs are protected from S1 nuclease digestion, amplified with the addition of barcodes, and sequenced. After amplification, the library was quantified according to the HTG EdgeSeq KAPA library quantification protocol for Illumina sequencing.

Processing controls include four negative and one positive control and a human brain RNA standard. All samples were run as singletons, except the standard was run in triplicate. Samples were randomized before placement to reduce interplate and intraplate biases, which were assessed using both Pearson and Spearman correlation coefficients.

Each well in HTG EdgeSeq assays included four negative control probes and one positive control probe with unique sequences. All samples and controls were quantified in triplicate with the inclusion of no template control reactions during the quantitative PCR process. A PhiX control adaptor-ligated library was spiked into the pooled library to confirm labeling efficiency, and each well was spiked with four unique plant sequences that were digested by the S1 nuclease during the protection assay.

The output was a read count, as in small RNA sequencing (RNA-seq), but unlike small RNA-seq, the read count reflects the quantity of probes bound by miRNAs and protected from digestion. The HTG EdgeSeq Parser aligned the FASTQ files to the probe list to collate the data. Data tables included raw, quality control raw, log₂ counts per million (CPM), and median normalized.

CPM standardization was used for evaluation between samples, replicate comparisons, batch effects, and quality control metrics. The log₂ transformation was used for standardizing, or scaling, gene-level data within a sample. CPM standardization allows the evaluation of probe-level expression as a proportion of total counts on a sample level and between samples, as described in the limma package (35,36). Intersample normalization was achieved by scaling raw read counts in each lane by a single lane-specific factor reflecting its library size (37). Gene counts were divided by the median of mapped reads (or library size) associated with their lane and multiplied by the median total count across all samples. Normalization was performed using the DESeq2 package from Bioconductor and the R statistical program (https://www.r-project.org).

We investigated whether C-peptide levels (fasting, AUC, and peak) were associated with miRNA levels at the baseline MMTT. We next examined whether baseline associations were maintained at the 12-month MMTT.

We estimated associations between baseline MMTT outcomes (fasting C-peptide, AUC, and peak levels) and miRNA levels. For each miRNA, we fit a linear model to MMTT outcomes, including a nominal indicator of quartile of expression, with adjustments for age, BMI, and sex. Global F tests of significance determined a list of miRNAs where variability in baseline MMTT outcome was significantly explained by miRNA quartile. Bootstrapped resampling with 1,000 replications provided correction for multiple comparisons and estimated comparisons among quartiles for differences in MMTT outcomes. For those miRNAs having significant bootstrapped associations between expression quartiles and MMTT outcomes, we evaluated specific comparisons between quartiles for significant differences in MMTT outcomes.

Those miRNAs with baseline associations with MMTT outcomes were investigated to evaluate maintained associations with 12-month C-peptide AUC and peak. This was accomplished by fitting a linear model to each MMTT, considering quartile of miRNA expression with identified association with baseline MMTT, and adjusting for corresponding baseline C-peptide, age at draw, sex, and BMI. Results illustrating evidence of longitudinal associations with specific miRNAs were then considered in stepwise regression to identify those miRNAs that, in combination, had significant association and best predicted 12-month MMTT after adjustment for baseline C-peptide AUC, age, sex, and BMI.

Receiver operating characteristic (ROC) curves were constructed from predicted values from logistic regression models fit to binary outcome defined by percentage decline from baseline (<25% vs. >25%). These models were fit with/without miRNA information in addition to baseline C-peptide AUC and were adjusted for BMI and sex. ROC curves were compared to determine whether miRNAs improved prediction of C-peptide decline above the baseline C-peptide AUC using the DeLong test for correlated ROC curves (38).

We plotted glucose against C-peptide values (30–120 min) from baseline, 6-month, and 12-month MMTTs for the two groups of subjects defined by baseline miRNA levels. Changes/differences in the curves’ position (shifts to the left, lower C-peptide; upward, higher glucose), directionality, and shape (a narrower horizontal spread and upward straightening of the curve or a monotonic shape) illustrate progressive worsening over time and differences between miRNA-stratified groups. We used the t test to assess statistical significance in comparisons of baseline mean AUC C-peptide/AUC glucose ratios of the two miRNA expression groups (curves) with 6- and 12-month ratios and in comparisons of miRNA-stratified groups (curves) within each MMTT.

For statistical comparisons involving binary outcomes, groups were compared using the two-tailed Fisher exact test.

We interrogated the reference Kyoto Encyclopedia of Genes and Genomes database (https://www.genome.jp/kegg) and used miRWalk 2.0 (39) (http://zmf.umm.uni-heidelberg.de/apps/zmf/mirwalk2) to identify gene pathways predicted to be targeted by miRNAs of interest (accessed April–May 2020). miRWalk 2.0 hosts predicted and experimentally validated miRNA-target interaction pairs, documents miRNA-binding sites within the complete sequence of a gene, and combines this information with a comparison of binding sites resulting from use of miRanda-rel2010, TargetScan 6.2, miRWalk 2.0, and RNA22 v2. Statistical significance of these predictions is reported after the Benjamini-Hochberg correction for multiple comparisons.

The data sets generated and/or analyzed during the current study were deposited in the Gene Expression Omnibus repository (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE157177). No applicable resources were generated or analyzed during the current study.

The EdgeSeq assays collectively detected all 2,083 miRNAs. Detection rates were >95% for all but four miRNAs (miR-128-2-5p 2%, miR-1282 27%, miR-4525 6%, miR-6752-3p 38%). The average ± SD detection rate was 2,077 ± 12 miRNAs, which accounts for 99.7 ± 0.5% of the panel. The mean normalized log₂ CPM values were 7.92 ± 0.33 (range 0.5–22.4) for all miRNAs. We observed very robust raw counts (mean 3.2 ± 1.05, range 1.4–7.7 million reads) in 15 μL of serum.

We estimated associations of baseline MMTT outcomes (fasting C-peptide, AUC, and peak levels) with miRNA expression quartiles. Bootstrapped resampling with 1,000 replications provided adjusted P values for associated miRNAs to correct for multiple comparisons and estimated comparisons among quartiles for differences in MMTT outcomes. Among statistically significant associations of C-peptide with mRNAs after bootstrapping, lowest or highest quartiles versus the other three quartiles were generally best at discriminating C-peptide differences. Table 2 lists miRNAs and quartile comparisons that identified significant differences in baseline C-peptide AUC and/or peak. Differences in C-peptide AUC or peak were identified by 25 and 22 miRNAs, respectively, among which 16 miRNAs had associations with both outcomes, which are naturally correlated. There were nine and six miRNAs associated with either C-peptide AUC or peak levels. There were no significant miRNA associations with fasting C-peptide. miRNA expression quartiles identified baseline C-peptide AUC differences ranging from 25.92 to 42.3 nmol/L and peak C-peptide differences ranging from 276 to 442 pmol/L after bootstrapping. The miRNA that identified the largest C-peptide AUC and peak differences was miR-3187-3p (42.3 nmol/L and 442 pmol/L, respectively); the previously reported miR-197-3p (6) identified C-peptide AUC and peak differences of 33.58 nmol/L and 351 pmol/L, respectively. All values are reported in Table 2. Figure 1 shows C-peptide AUC and peak levels according to miRNA expression quartiles for six representative miRNAs. Supplementary Table 5 reports total raw counts, raw CPM, and normalized CPM (log₂) for the 31 miRNAs associated with C-peptide.

We investigated whether baseline miRNAs predicted C-peptide at the 12-month MMTT. Specifically, we examined whether any miRNA associated with baseline C-peptide AUC and/or peak (Table 2) remained associated and predicted C-peptide AUC and/or peak at the 12-month MMTT after adjusting for baseline C-peptide, age, sex, and BMI. Two miRNAs remained associated with both C-peptide AUC (miR-3187-3p P = 0.037, miR-4302 P = 0.047) and peak (miR-3187-3p P = 0.038, miR-4302 P = 0.039) at the 12-month MMTT (Table 3). Baseline expression quartiles of miR-3187-3p and miR-4302 defined groups of participants with a mean difference in the 12-month C-peptide AUC of 22.49 and 21.05 nmol/L, respectively; mean differences for 12-month peak C-peptide were 236 and 229 pmol/L, respectively. miR-1292-5p was associated with peak C-peptide only (mean difference 212 pmol/L). Figure 2 illustrates longitudinal associations and C-peptide decline for representative miRNAs miR-3187-3p and miR-4302 (which stratified subjects with significantly different 12-month C-peptide AUC and peak) and miR-103a-3p and miR-197-3p (which stratified subjects only at baseline).

miRNAs associated with baseline C-peptide AUC were evaluated in a stepwise regression to identify miRNA combinations with improved prediction of 12-month C-peptide AUC after adjustment for baseline C-peptide AUC, age, sex, and BMI. Twelve of 25 miRNAs with baseline associations were included in the model. The combination of miR-3187-3p and miR-103a-3p discriminated C-peptide AUC (Fig. 3A) and peak levels (Fig. 3B) at 12 months. Stratification according to baseline miRNA expression quartiles for this combination demonstrated differences in the baseline to 12-month AUC and peak C-peptide levels of 37.95 nmol/L (P = 0.001) and 39 pmol/L (P = 0.001) between groups (Supplementary Table 2). Eleven subjects with low expression (first quartile) of miR-3187-3p combined with high expression (second–fourth quartile) of miR-103a-3p had higher 12-month C-peptide AUC compared with the other 42 subjects. This combination was superior to miR-3187-3p alone because two more subjects were stratified to the lower C-peptide group. The miR-3187-3p/miR-4302 combination identified differences in AUC C-peptide decline of 44.86 nmol/L (P = 0.001) (Fig. 3C and Supplementary Table 2) and assigned two additional individuals to the lower C-peptide group than miR-3187-3p alone.

We also investigated whether any single miRNA or combination could stratify subjects at the 12-month visit by peak C-peptide levels above/below the clinical trial eligibility threshold. No individual miRNA from Table 2 was predictive. However, the combination of miR-3187-3p and miR-4723-5p predicted that 94% (17 of 18) of the participants with baseline expression levels in the lower quartile for both miRNAs would have a peak C-peptide level >200 pmol/L at the 12-month visit compared with 64% (22 of 34) of those with miRNA expression levels in the second to fourth quartiles (P = 0.021 by two-tailed Fisher exact test, relative risk 1.4 [95% CI 1.086–1.993], sensitivity 0.4359 [95% CI 0.2930–0.5902], specificity 0.9231 [95% CI 0.6669–0.9961]).

We examined whether miRNAs improved prediction of C-peptide decline compared with baseline C-peptide AUC. The ROC curve in Fig. 3D illustrates an improved ability to separate between groups with decline greater/lower than 25% when the miR-3187-3p/miR-103a-3p combination is considered in the model; ROC AUCs with/without miRNAs were 0.82 and 0.70, respectively (P = 0.04).

C-peptide/glucose response curves for baseline, 6-month, and 12-month MMTTs stratified subjects into two groups (curves) by their baseline levels of associated miRNAs (Fig. 4). Curves evolved at 6 months and 12 months, demonstrating progressive worsening of insulin secretion (shift to the left) and higher glucose levels (shift upward), with greater separation of the curves for the miRNA combinations. Overall, subject stratification by baseline miRNA expression quartiles demonstrated differences in disease severity at diagnosis that persisted during further progression and involved both C-peptide and glucose responses.

Longitudinally, we compared baseline mean AUC C-peptide/AUC glucose ratios of the two miRNA expression groups (curves) with the ratios of the corresponding curves at later time points; for example, the ratio from the first quartile curve at baseline was compared with the ratio of the first quartile curve at 6 months, and the same comparison was made between 6 and 12 months. These were all significantly different from one another, demonstrating worsening in both groups (range P < 0.0001–0.02 by two-tailed paired t test) (Supplementary Table 3). However, 6- and 12-month curves of miR-3187-3p/miR-4302 were statistically different from each other for subjects in combination group 0 but not for those in combination group 1, suggesting that the latter did not experience significant worsening in this time interval. Cross-sectionally, we compared AUC C-peptide/AUC glucose ratios of the two groups (curves) in each panel; these were significantly different at all time points for all miRNAs or combinations analyzed (range P < 0.0001–0.0397 by two-tailed unpaired t test) (Supplementary Table 4); the only exception was the miR-103a-3p 12-month curves (P = 0.05). The findings suggest significant differences in disease progression identified by stratification in groups defined by baseline miRNA levels.

We used miRWalk 2.0 (39) to examine whether any of the 31 miRNAs associated with C-peptide AUC and/or peak at baseline are predicted to modulate gene pathways relevant to type 1 diabetes. Results are reported in Table 4 where we list four major gene pathways relevant to type 1 diabetes and/or type 2 diabetes, specifically the insulin signaling, SNARE interactions in vesicular transport, type 2 diabetes, and the T-cell receptor (TCR) signaling pathways. Nineteen miRNAs were predicted to target either the insulin or the TCR signaling pathways, and remarkably, nine miRNAs were predicted to target both (miR-103a-3p, miR-193b-5p, miR-197-3p, miR-3187-3p, miR-4302, miR-622, miR-6748-3p, miR-1208, and miR-1292-5p). Among the five miRNAs that alone or in combination predicted 12-month C-peptide outcomes (miR-3187-3p, miR-4302, miR-1292-5p, miR-103a-3p, and miR-4723-5p), four targeted both insulin and TCR signaling pathways and potentially may modulate a large number of genes (71–104 of 139 genes and 65–80 of 110 genes, respectively). For miR-3187-3p, the TCR signaling pathway was predicted as the first of 16 pathways.

Despite the emergence of reproducible associations of circulating miRNAs with type 1 diabetes, there are limited data about miRNA prediction of C-peptide decline after diagnosis. Moreover, virtually all published studies of circulating miRNAs in islet autoimmunity and type 1 diabetes used RT-PCR assays investigating a fraction of the known miRNAs (Supplementary Table 1). Nielsen et al. (6) sequenced pooled samples to identify differentially expressed miRNAs between individuals with type 1 diabetes and control subjects and then assessed levels of 24 miRNAs by RT-PCR. With 2,656 transcripts in miRBase V22 (34), there is much potential for discovery. Thus, we profiled 2,083 miRNAs using RNA-seq technology. To date, our study has examined the largest number of miRNAs concerning residual C-peptide at diagnosis.

To investigate whether circulating miRNAs are associated with and predict loss of insulin secretion after diagnosis, we examined fasting serum samples obtained on the same day of the baseline MMTT from 53 individuals. Several miRNAs were associated with C-peptide AUC and/or peak at the baseline MMTT (Table 2 and Fig. 1). miRNA expression quartiles identified subjects with better or worse residual insulin secretion after adjustment for age, sex, and BMI. The observed differences were not explained by variation in time from diagnosis to baseline MMTT (data not shown). These associations survived correction for multiple comparisons by bootstrapping.

In longitudinal analyses, baseline levels of five of these miRNAs, alone or in combination, predicted MMTT C-peptide outcomes at the 12-month visit: miR-3187-3p and miR-4302 predicted C-peptide AUC, miR-3187-3p/miR-103a-3p predicted AUC and peak, and miR-1292-5p predicted peak C-peptide; miR-3187-3p/miR-4723-5p predicted subjects being above or below the peak C-peptide trial eligibility threshold at the 12-month visit. The miR-3187-3p/miR-103a-3p combination improved prediction of C-peptide decline compared with baseline C-peptide AUC alone (Fig. 3D). Baseline differences in C-peptide AUC or peak were maintained on follow-up after correction for baseline C-peptide, age, sex, and BMI, and decline occurred with similar slopes (Figs. 2 and 3). We cannot discern whether this reflects differences in physical/functional β-cell mass at baseline, in the severity of the autoimmune process, or both.

A prior study of relatives at risk for type 1 diabetes showed that plotting C-peptide AUC against glucose AUC values at the time points of the oral glucose tolerance test helps with assessing metabolic impairment during progression to clinical diagnosis (40). Changes in the curves’ position, shape, and direction demonstrated progressive worsening during the progression. For the first time, we applied this approach to visualize these relationships during the MMTT and analyze differences in metabolic responses at baseline and on follow-up in groups stratified by baseline miRNA levels (Fig. 4).

Subjects having higher baseline C-peptide AUC after miRNA stratification had less pathological curves: Their position on the grid indicated higher C-peptide and lower glucose levels, and their shape indicated more C-peptide secretion relative to glucose levels. On follow-up, their curves remained distinct from those of the other participants. The comparisons of the AUC C-peptide/AUC glucose ratios from baseline ratios quantified that the significant worsening and the differences among the groups stratified by the baseline miRNA levels persisted over time.

For most of the 31 miRNAs associated with C-peptide at baseline, there is no prior involvement in disease-relevant pathways. This is not surprising given that we identified many miRNAs never before examined in this setting. However, miRWalk 2.0 predicted that 19 of 31 miRNAs may target disease-relevant gene pathways; 7 of 19 had previous disease-relevant literature associations (4 with type 1 and type 2 diabetes, 2 with type 2 diabetes only, and 1 with β-cell differentiation). Remarkably, 19 miRNAs could target either the insulin or the TCR signaling pathway; 9 miRNAs may target both. These included four of five miRNAs associated with C-peptide at the 12-month MMTT alone or in combination (miR-3187-3p, miR-4302, miR-103a-3p, miR-1292-5p, and miR-4723-5p). There were no previous associations, except for miR-589-5p, for 13 miRNAs with no relevant predictions.

miR-3187-3p had the strongest association with C-peptide. The TCR signaling pathway ranked first of 16 predicted pathways. It could target genes involved in AKT (serine/threonine kinase)/phosphatidylinositol 3-kinase signaling, which is critical for the development, differentiation, and function of effector (41) and regulatory (42) T cells. Other predicted targets include the CD3 γ-chain, the MAPK13 and MAPK14 genes in the mitogen-activated protein kinase signaling pathway, LAT (linker for the activation of T cells), SOS1 (son of sevenless factor 1, an exchange factor recruited by LAT), NFAT (transcription factor nuclear factor of activated T cells), and PTPRC (the tyrosine phosphatase CD45) (43).

Plasma levels of miR-103a-3p were increased in individuals with type 1 diabetes (<5 years duration compared with healthy subjects) (18). We show that higher miR-103a-3p levels are associated with higher residual insulin secretion near diagnosis and that baseline miR-103a-3p levels can aid in predicting C-peptide AUC at 12 months; the combination miR-3187-3p/miR-103a-3p was the stronger predictor of C-peptide AUC. This miRNA was linked to type 2 diabetes, obesity, and HFN1A maturity-onset diabetes of the young (44–46). In the Coronary Diet Intervention with Olive Oil and Cardiovascular Prevention study, low circulating levels of miR-103a-3p were associated with increased likelihood of type 2 diabetes (47). miR-107 and miR-103a-3p are negative regulators of insulin sensitivity (48). Validated gene targets of miR-103a-3p include SFRP4 (the secreted frizzled-related protein 4), which suppresses insulin exocytosis (49), and CAV1 (caveolin 1), which inhibits insulin signaling by decreasing insulin receptors in caveolae-enriched plasma membrane domains (48). miR-103a-3p regulates the autophagy gene ATG5 (50), and autophagy regulates transport-competent secretory peptide precursors, including proinsulin (51). In our analysis, this miRNA may target both the TCR and the insulin signaling pathways.

Other miRNAs were associated with C-peptide at baseline but not on follow-up. These included miR-197-3p, which in a previous report predicted future C-peptide AUC (12); the different outcomes may reflect assay type, sample size (most likely), or sample type (serum vs. plasma). miR-197-3p is predicted to target several genes in both the insulin and the TCR signaling pathways. Its plasma levels were reduced in subjects with type 2 diabetes (52).

The gene coding for miR-342-3p on 14q32 contains a cluster of glucose-responsive miRNAs expressed in pancreatic islet cells (53); miR-342-3p also regulates the expression of the autoantigen IA-2β (54). We previously reported that miR-342-3p levels were associated with an increased risk of progression to type 1 diabetes among autoantibody-positive relatives, and levels correlated with oral glucose tolerance test outcomes (13); in other studies, miR-342-3p was differentially expressed in individuals with type 1 diabetes compared with healthy subjects and at-risk relatives (14); its levels were reduced in regulatory T cells in affected individuals compared with healthy subjects (25).

miR-127-3p is enriched in human β-cells (55) and involved in endocrine differentiation (56); miR-99a-5p exhibited increased levels during the first 12 months postdiagnosis in children with recent-onset type 1 diabetes (20) and targets the mammalian target of rapamycin pathway (57). miR-589-5p (58) and miR-193b-5p were associated with type 2 diabetes and prediabetes, respectively; miR-193b-5p was linked to islet autoimmunity because it was differentially expressed in autoantibody-positive versus -negative individuals (10).

In closing, trial subjects with higher or lower MMTT C-peptide AUC and peak were stratified by baseline miRNA levels. Selected miRNAs/miRNA combinations predicted future decline of C-peptide AUC and peak. Predicting future C-peptide at baseline is critical for subject stratification early after diagnosis when impactful decisions about trial participation or treatment need to be made. An miRNA combination predicted 12-month C-peptide peak above/below the clinical trial eligibility threshold, which is of particular importance given increased consideration for trial enrollment up to 2 years from diagnosis if meeting the peak C-peptide threshold. Many associated miRNAs were examined for the first time, but some were previously linked to type 1 diabetes; several are predicted to affect gene pathways relevant to β-cell function and T cells, both critical to disease pathogenesis. Future studies may explore possible links of miRNAs with disease endotypes. Limitations of this study are the limited sample size and the lack of a validation cohort, which require future investigations. These miRNAs are excellent candidates for validation studies and may become useful biomarkers for advancing therapeutic discoveries for type 1 diabetes.

Acknowledgments. We acknowledge Simi Ahmed (JDRF) for programmatic support.

Funding. The study was supported by JDRF (2-SRA-2015-122-Q-R) and the Diabetes Research Institute Foundation, Hollywood, Florida. We acknowledge the support of the Type 1 Diabetes TrialNet Study Group, which identified study subjects and provided samples and follow-up data for this study. The Type 1 Diabetes TrialNet Study Group is a clinical trials network funded by the National Institutes of Health through the National Institute of Diabetes and Digestive and Kidney Diseases, the National Institute of Allergy and Infectious Diseases, and the Eunice Kennedy Shriver National Institute of Child Health and Human Development through cooperative agreements U01-DK-061010, U01-DK-061034, U01-DK-061042, U01-DK-061058, U01-DK-085453, U01-DK-085461, U01-DK-085465, U01-DK-085466, U01-DK-085476, U01-DK-085499, U01-DK-085504, U01-DK-085509, U01-DK-103180, U01-DK-103153, U01-DK-103266, U01-DK-103282, U01-DK-106984, U01-DK-106994, U01-DK-107013, U01-DK-107014, UC4-DK-106993, UC4-DK-11700901, and U01-DK-106693-02 and through JDRF.

The contents of this article are solely the responsibility of the authors and do not necessarily represent the official views of the National Institutes of Health or JDRF.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. I.S. generated the data. I.S., R.P., J.S., S.M.C., and A.P. analyzed and interpreted data and participated in the preparation of the manuscript and approved its final version. R.P., J.S., S.M.C., and A.P conceived the study or parts of the study. S.M.C. was responsible for statistical design and the analysis plan. S.M.C. and A.P. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.