Plasma miRNAs and Treatment Failure in Participants in the Treatment Options for Type 2 Diabetes in Adolescents and Youth (TODAY) Study

In the Treatment Options for Type 2 Diabetes in Adolescents and Youth (TODAY) study, more than half of the participants developed treatment failure over <4 years of follow-up (1). miRNAs are short noncoding RNAs that regulate gene expression in the pancreas, liver, adipose tissue, and skeletal muscle (2–4) and are actively secreted into the circulation (5), where they are markers of gene expression regulation in tissue and mediate interorgan cross talk (5). We conducted a secondary analysis in TODAY to examine the potential of circulating miRNAs as biomarkers of treatment failure in youth with type 2 diabetes (T2D).

Participants were selected from TODAY, a randomized trial that tested three treatment approaches (1,000 mg metformin twice daily, 4 mg metformin plus rosiglitazone twice daily, or metformin plus a lifestyle intervention) in youth ages 10–17 years with overweight or obesity diagnosed with T2D according to American Diabetes Association criteria in the prior 2 years (1). For this secondary analysis, deidentified data and samples were provided by the Central Repository at the National Institute of Diabetes and Digestive and Kidney Disease. Adequate fasting plasma was available from the baseline visit for 383 participants. We randomly selected 210 samples, all of which but one passed initial quality control tests, leaving a final analytic sample of 209. The parent study was approved by the institutional review board at each participating institution. Parents provided written informed consent, and pediatric participants provided assent. This project was determined not to involve human subjects by the institutional review board at the University of Washington. HbA_1c testing was performed every 2 months in the first year and quarterly thereafter. Assays were performed at the TODAY central laboratory at the Northwest Lipid Research Laboratory, University of Washington (Seattle, WA) (6).

Details of high-throughput miRNA sequencing are provided in the Supplementary Methods.

We summarized descriptive statistics and inspected principal components analysis plots to evaluate overall data quality and identify problematic samples. Participants were randomly divided into training (n = 125) and test data sets (n = 84) using a 60%/40% split, ensuring equal proportions (45%) of treatment failures in each data set. In the training data set, we fit regularized Cox models, stratified by treatment arm, using miRNA log counts/million counts in the R glmnet package (7,8). To examine whether miRNAs could predict treatment failure at 365 days, we used glmnet to fit a logistic regression model with the same training/test sets as for the survival analysis and used the training set model to make predictions using the test set.

Next, we examined whether individual miRNAs were associated with treatment failure independently of factors known to be prospectively associated with treatment failure (1,9–12): sex, race/ethnicity, BMI, Tanner score, baseline A1C level, maternal history of diabetes, oral disposition index at baseline, and treatment arm. For the 44 miRNAs with some evidence of association with treatment failure in the glmnet analysis (i.e., they were associated with treatment failure in the test data set at a false discovery rate of <0.1), we fit individual Cox proportional hazards models and used likelihood ratio tests to determine if miRNAs significantly improved model fit. Ingenuity pathway analysis software (Qiagen IPA; Qiagen, Redwood City, CA) was used to identify predicted mRNA targets (13). Additional details are provided in the Supplementary Methods.

Participants were age 14.5 years at baseline, and 62% were female. Youth with treatment failure were more likely to report Black or Latinx race/ethnicity and had a higher HbA_1c (Table 1). Principal components analysis plots showed clustering by read sequencing depth, suggesting that this contributed a large amount of variability (data not shown). Therefore, we generated several surrogate variables using the Bioconductor sva package, a common strategy for addressing variation in high-throughput experiments (14).

Neither the glmnet survival nor logistic models identified a panel of miRNAs that predicted treatment failure. In the glmnet Cox model using the training data set, participants with treatment failure had a higher predicted relative risk of treatment failure based on our miRNA model than those who did not experience treatment failure; however, predicted risk of treatment failure based on the miRNA model was not different in the test data set (Fig. 1). Therefore, the prediction model was not internally validated. In the glmnet logistic regression models in the training data set, no miRNAs were useful for classifying participants by treatment failure.

We identified 44 miRNAs that were associated with treatment failure in the glmnet survival models in the test data set (false discovery rate <0.1). Of these, two were associated or marginally associated with failure over all follow-up time in separate multivariate models. For each doubling, miR-4306 was associated with a 12% decrease (P = 0.040) and miR-483-3p was marginally associated with a 12% increase (P = 0.080) in failure. The addition of both miRNAs improved model fit (log likelihood without vs. with miRNAs −360.3 vs. −363.5; P = 0.040) in a multivariate Cox model, although neither miRNA was statistically significantly associated with failure in this model. Independently of these miRNAs, higher HbA_1c was associated with a higher risk of treatment failure, whereas White race was associated with a lower risk of failure compared with Black race, as was higher log-transformed C-peptide oral disposition index (Table 2). Identified miRNAs were predicted to have signaling roles for the retinoic acid receptor, cardiac hypertrophy, the serotonin receptor, and insulin secretion (Fig. 2).

In this secondary analysis of a subset of data and samples from a randomized trial of youth with recent-onset T2D, we found that plasma miRNAs alone did not predict treatment failure over an average 962 days of follow-up. However, in separate Cox models, higher levels of miR-483-3p and lower levels of miR-4306 were associated or marginally associated with failure independently of clinical factors that have previously been related to treatment failure in T2D. The addition of miR-483-3p and miR-4306 significantly improved the fit of a multivariate model of treatment failure. These miRNAs were predicted to have roles in signaling for the retinoic acid receptor, cardiac hypertrophy, the serotonin receptor, and insulin secretion.

To our knowledge, this study is the first to identify circulating miRNAs prospectively related to treatment failure in youth with T2D. miRNAs were prospectively related to treatment response in adults with prediabetes enrolled in a clinical trial of yoga versus stretching (N = 83), where a panel of 14 miRNAs was associated with fasting blood glucose at 12 months (15). In adults with T2D (N = 26) treated with metformin with or without liraglutide or dulaglutide, a higher proportion of individuals classified as having high overall levels of eight prespecified miRNAs met a glycemic target (A1C <7%) at 12 months that defined treatment success (16). Neither miR-483-3p nor miR-4306 was included in either analysis.

In the current study, miR-483-3p and miR-4306 were associated or marginally associated with failure but did not predict its occurrence. Predictive models may fail for several reasons. First, participants in the training and test data sets may be dissimilar. There were no differences for any prespecified confounders or failure rates between the data sets, suggesting that this issue may be less important here. Second, overfitting may occur when a machine learning model learns the training data set so well that it performs poorly on a test data set (17). Third, noise resulting from measurement error or other factors may limit generalizability.

miR-483 may have a functional role in T2D. It is highly expressed in β-cells (18). It is upregulated in adipose tissue from adult rat offspring of mothers fed a suboptimal (low-protein) diet compared with controls and directly represses expression of growth/differentiation factor-3 (19). This finding is generally consistent with the direction of the association seen in the current study, in which higher levels of miR-483-3p were associated with a higher risk of failure. Conversely, however, miR483^−/− mice fed a high-fat diet had hyperglycemia compared with controls, although β-cell mass was maintained (20), suggesting that miR-483 may help protect β-cell function. The mechanisms by which miR-483 might contribute to differential T2D treatment response therefore remain unclear. In a sensitivity analysis in which we did not adjust for oral disposition index, miR-483 was associated with a statistically significant 19% higher risk of failure; however, this association was attenuated in our primary models. This might suggest that circulating miR-483 levels are related to β-cell function. miR-4306 is less well studied in diabetes.

Strengths of the study include the use of data and samples from a racially and ethnically diverse and rigorously conducted randomized trial, a relatively large sample size, and next-generation miRNA sequencing. There are also limitations, most importantly lack of external validation. Additionally, because only a small number of participants experienced treatment failure at 365 days, this analysis may have been underpowered.

In conclusion, a panel of plasma miRNAs alone did not successfully classify treatment failure in TODAY participants. However, higher levels of miR-483-3p and lower levels of miR-4306 were independently associated or marginally associated with treatment failure. Future studies should replicate these findings, examine contributing mechanisms, and contrast differences in patterns of miRNAs that are related to T2D outcomes in youth and adults.

Funding. This project was funded by grants from the the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), National Institutes of Health (K08DK103945 and R03DK122100). The TODAY study was conducted by the TODAY study investigators and supported by the NIDDK. The data and biospecimens from the TODAY study reported here were supplied by the NIDDK Central Repository.

The manuscript was not prepared in collaboration with investigators of the TODAY study and does not necessarily reflect the opinions or views of the TODAY study, NIDDK Central Repository, or NIDDK.

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

Author Contributions. P.L.W. conceived the project, obtained funding, and wrote the manuscript. T.K.B. analyzed the data and reviewed/edited the manuscript. J.W.M. analyzed the data and reviewed/edited the manuscript. S.S. analyzed the samples and reviewed/edited the manuscript. E.J.B. contributed to the discussion and reviewed/edited the manuscript. D.A.E. contributed to the design/interpretation of the analyses and reviewed/edited the manuscript. P.L.W. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.