Applicability of a microRNA-based dynamic risk score for type 1 diabetes

doi:10.21203/rs.3.rs-5053763/v1

Download PDF

Article

Applicability of a microRNA-based dynamic risk score for type 1 diabetes

https://doi.org/10.21203/rs.3.rs-5053763/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Identifying biomarkers of functional β-cell loss is critical in risk stratification for Type 1 Diabetes (T1D). We report a microRNA-based dynamic (responsive to environment) risk score developed using multi-center, multi-ethnic/country (“multi-context”) cohorts. Discovery (wet-lab and dry-lab) analysis identified 50 microRNAs that were measured across n=2,204 individuals from four contexts (4C=AUS/Australia, DNK/Denmark, HKG/Hong Kong SAR China, IND/India). A microRNA-based dynamic risk score (DRS) was generated (DRS4C), which effectively stratified individuals with/without T1D. Generative artificial intelligence (GAI) was used to create an enhanced (e)DRS4C that showed AUC >0.84 on an independent Validation set (n=313) from AUS, IND and NZL and predicted future exogenous-insulin requirement in islet transplantation recipients from Canada (CAN). In another T1D therapy, this microRNA signature stratified 1-year response to imatinib based on their profile at the study baseline. Utilizing machine learning and GAI, this study identified and validated a microRNA-based DRS for T1D stratification and treatment efficacy prediction.

Health sciences/Biomarkers/Predictive markers

Health sciences/Diseases/Endocrine system and metabolic diseases/Diabetes/Type 1 diabetes

Functional loss of islet β-cells is a common feature of progression to clinical diagnosis of type 1 diabetes (T1D) as well as after islet cell transplantation (Tx)¹. Islet cell death is known to commence many years before clinical diagnosis of T1D, which usually occurs when >60% of islet β-cells have been lost^2,3 to auto-reactive immune cells. Autoantibodies⁴ and human leucocyte antigen (HLA) genotypes^5-7 are well-recognized biomarkers in T1D. However, autoantibodies already mark the start of progression to T1D, reminding us of “the streetlight effect”⁸ and, although a strong genetic link to the development of T1D is known^9-12, 80% of individuals who develop T1D have no family history (first or second-degree relatives) of T1D¹³. Discordance for T1D between monozygotic twins^14,15 implies that non-genetic (i.e. variable or dynamic) factors influence the outcome of a genetic predisposition to T1D. The identification of new drugs^16-23, such as the anti-CD3 monoclonal antibody, teplizumab¹⁶, and other immunotherapies (including rituximab, abatacept, alefacept, anti-thymocyte globulin²², imatinib¹⁷, golimumab²³ and baricitinib²¹) demonstrate that the course of T1D progression can be changed, necessitating the development and validation of biomarkers that capture dynamic changes in T1D risk²⁴.

The GPPAD POInT study²⁵ demonstrated that children progressing to autoimmunity in T1D show a sharp and sustained rise in postprandial blood glucose at ~2 months before seroconversion. Recent studies have also underscored link between HbA1c and progression to autoimmunity in T1D^26-28. These studies reveal that dynamic functional biomarkers of β-cells may be important candidates for developing a risk score that captures T1D risk relative to a changing environment²⁴. MicroRNAs are functional^29-31, dynamic biomarkers^32,33 as well as regulators^34-38 of islet cell function. MicroRNAs are also demonstrated to be differentially expressed in immune cell subtypes^39-42 and between individuals with and without T1D. Mature microRNAs, which we report here, are ~22 nucleotide, single-stranded RNA molecules that do not encode for any protein. They act post-transcriptionally, mostly inhibiting protein formation by interfering with mRNA translation and/or stability³⁸. MicroRNAs are present in the circulation either within the extracellular vesicles or bound to protein/lipid moieties. These are released for inter-cellular communication, or as a result of stressed/ruptured cells. MicroRNAs are a valuable biomarker that can offer readouts of β-cell function over time^30,43, are resistant to degradation by endogenous nucleases and by repeated freeze-thawing, have a long half-life and can be reliably measured⁴⁴ from stored clinical samples.

In recent years, many novel predictive biomarkers have been proposed, measured, and tested for their potential to stratify individuals with and/or without T1D^24,45-47. Here we demonstrate the development of a microRNA-based dynamic risk score (DRS) for T1D stratification, its validation in an independent case-control, multi-ethnic/country cohort and the applicability of such a risk score/signature in the prediction of response to T1D therapies.

Study populations and strategy:

This study encompasses microRNA analysis of participant’s plasma/tissue samples from four continents across seven countries (Australia, Canada, Denmark, Hong Kong SAR China, India, New Zealand and The United States of America) that were similar in processing (<1 freeze-thaw). In addition to these wet-lab analyses, we also analysed microRNAs across 25 previously reported studies (until 2020) that demonstrated significant expression of wet-lab discovery microRNAs (Extended data Table 1). Details on available demographic characteristics of wet-lab discovery set participants, 4C (Train & Test) cohort, validation set, and T1D therapy participants are provided in Extended data Table 2

Our study participants (Extended data Table 2) are from diverse backgrounds (countries), influenced by factors such as ethnicity, geography, culture, beliefs, socioeconomic status, and urbanization. While understanding ethnic diversity is important for identifying differences in biomarkers, focusing at country-level offers a more practical and inclusive way to develop health policies relevant to a specific context. To address this, we define participants from a country as a “context,” using the three-letter country code to represent diversity in age, ethnicity and socioeconomic status.

Identification of a set of 50 microRNAs through discovery analyses

We carried out a stepwise discovery analysis by profiling 754 known and validated microRNAs in developing human pancreas (n=55)²⁹, in human islet vs non-pancreatic tissues (total, n=189)³⁴, and in plasma samples from recent onset T1D (n=5) and age-/sex-matching Controls (n=5) (Figure 1a). We identified islet-cell/pancreas enriched microRNAs that were significantly different between the islet and non-pancreatic tissues, during β-cell development, and in T1D plasma samples (Extended data Table 3).

To further strengthen our discovery analysis, we evaluated studies that reported microRNAs in T1D and Controls and identified 25 previous reports (Extended data table 1) containing microRNAs analysed from 1,757 T1D and 1,112 Controls. This confirmed that microRNAs within our wet-lab discovery set are also recognized across other studies to be either associated with HLA (genetic risk for T1D), Autoantibody expression, Case-controls or with early onset (<15 years of age). We also selected five microRNAs (Figure 1b) that did not reach statistical significance in our wet-lab analysis; but were significant in previous reports. We also included negative microRNA controls (undetectable across almost all samples), positive microRNA controls (known cellular/housekeeping microRNAs) and stage-specific spike-in controls for measurement in every plasma sample assessed. The panel of microRNAs, their source for inclusion and their associations with T1D risk (based on published reports) are compiled in Figure 1b.

Development of a microRNA-based DRS for T1D

We measured the abundance of these 56 microRNAs (together called as PREDICT T1D microRNAs) in plasma of n=2,204 study participants (non-T1D Controls n=752, T1D n=1,452) from four contexts (AUS, DNK, HKG and IND). All microRNAs except the QC and three negative controls (miR-558, miR-188-3p and miR-220c) were significantly different between T1D and T1D-free groups when compared within a single context or across all four contexts (4C) combined together (Extended data Figure 1). Some of these microRNAs also correlated (p<0.05, r>0.9) with each other within individual contexts (Extended data Figure 2). Genomic location mapping of PREDICT T1D microRNAs indicated that they are present on various chromosomes; with only one microRNA (miR-30a-5p) located on chromosome 6, albeit not within the HLA locus (Extended data Figure 2, Extended data Table 4). None of the PREDICT T1D microRNAs are located on the Y-chromosome, and our feature selector algorithm did not identify participant’s sex as a covariate in T1D prediction.

DRS4C was generated using machine learning (ML) Random Forest algorithm on 4C cohort microRNA expression data randomised to Train and Test sets (Figure 2a). DRS4C yielded a good ROC curve (Figure 2b) on the Test set (n=661). Model development assessed participant’s age at sampling (in years) and all 50 microRNAs, after excluding negative and spike-in controls. The model performance metrics are provided in Figure 2c whilst the key model variables are presented in Figure 2d.

Using Generative Artificial Intelligence (GAI) for model enhancement.

Although our DRS4C model training was carried out on a reasonably large dataset (n=1,543), the levels of expression for each variable follow a distribution leaving several probabilities of microRNA expression and/or age missing from these real-life samples (Figure 3a). Generative AI was used to create synthetic datasets (Figure 3b) that augment the original Train set data, ensuring comprehensive coverage of microRNA expression level and age probabilities across all contexts. A total of n=1000 (1K), n=10,000 (10K) and n=100,000 (100K) synthetic participants were created with synthetic data on all 50 miRNAs and age that followed the distribution of four-context Train sets (Figure 3b). These 1K-, 10K- and 100K-augmented datasets were used for generation of three eDRS4C models (1K, 10K and 100K) using Random Forest algorithm. The predictive potential of eDRS4C risk scores was assessed on an independent validation set (n=313 individuals) from AUS, IND and New Zealand (NZL). All the three eDRS4C models offered a ROC Area Under the Curve of >0.84 (Figure 3c). Intriguingly, model derived from 1K synthetic samples offered higher Precision (PPV) and Specificity (Figure 3c) whereas model developed using 100K synthetic samples offered a better Recall (sensitivity) and Negative Predictive Value (NPV). This allowed selecting a version of eDRS4C appropriate to populations with high or low risk of T1D. For example, in case of population screening where T1D prevalence is low, the eDRS4C (1K) model will minimise diagnosing someone incorrectly with T1D (false positives) or aid in prioritizing those that are truly at risk of T1D. The eDRS4C (100K) model can be applied to scenarios wherein missing a T1D diagnosis could delay identifying an individual at risk of life-threatening Diabetic Ketoacidosis (DKA). This model may also be more applicable in a high-risk of T1D setting (family history, geographical location, or pandemic exposure) where it may be more important to identify every possible case. Thus GAI-based model enhancement led to development of two versions of eDRS4C that can be applied in different settings.

Applicability of microRNA-based eDRS4C in prediction of future diabetes status

Since the eDRS4C (1K) offers higher precision and specificity, we assessed the application of this risk score in human clinical islet transplant settings (Figure 4a), which is a recognized T1D therapy in Canada (CAN), as well as other parts of the world including Europe and Australia. It is well known that when individuals with T1D are transplanted with cadaveric (donor) human islets, several of the transplanted islets die within the first 24h post-transplant¹. Some islet graft recipients can show none to very-low exogenous insulin requirement post-transplant. We calculated eDRS4C (1K) in plasma samples from n=15 individuals at 1h and 24h post-transplant (Figure 4b) and assessed if it could predict their insulin independence at 1-month post-transplantation (Figure 4c). This was the farthest time point available to us where post-transplant exogenous insulin requirement was recorded and none of these recipients had a second islet transplant. The model performance (Figure 4d) was exceptional for precision and specificity at 1h post-transplant indicating correct identification of progression to insulin requirement. The AUC of 0.815 at 1h indicated that the prediction was best at 1h post-transplant, when cell death is anticipated to be higher than at 24h post-transplant. The eDRS4C (100K) model, offered the same AUC of 0.815 at 1h post-transplant due to improved Recall (sensitivity) of 0.667, thereby offering an F1 score of 0.750.

Application of the microRNA signature in predicting treatment efficacy

In an emerging T1D drug therapy (imatinib therapy¹⁷), the 50 PREDICT T1D microRNAs were assessed across the intervention group of n=43 participants at the trial baseline and drug recipients were stratified into upper quartile (UQ; n=11) and the lower quartile (LQ; n=11) of response to therapy (c-peptide measurement) as identified at the end of 12 months (Figure 5a). Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction analyses segregated the UQ study participants from LQ based on microRNA expression at the study baseline (Figure 5b). Intriguingly, 17 of the 50 microRNAs were significantly higher in the UQ participants than the LQ participants (Figure 5c). The UMAP decision tree analysis indicated hsa-miR-27b-3p as the most important variable in stratifying UQ and LQ responders to drug intervention (Extended Data Figure 3a). Intriguingly, miR-27b-3p targets 36 different genes in the tyrosine kinase pathway, which were different from nine genes reportedly targeted by imatinib (Extended Data Figure 3b). Pathway analyses (Extended Data Table 5) indicated that several of these 17 microRNAs target the tyrosine kinase pathway. These data indicated that it was not just the higher expression of tyrosine kinase inhibitory microRNAs (Figure 5c) but a much broader targeting of the tyrosine kinase pathway (Extended Data Figure 3b, Extended Data Table 5). Thus, the best responders to imatinib appear to be those that had higher baseline expression of tyrosine kinase targeting microRNAs. These observations demonstrate a synergistic mechanism between microRNAs and imatinib leading to better drug responsiveness.

Our study involves large number of samples from T1D and non-T1D individuals belonging to different contexts spread over four continents of the world and assessed using robust methodologies utilizing optimized robotic workflows to address any variation in processing and assaying samples (including haemolysis). The analysis of re-identifiable study samples employing unbiased ML and GAI techniques facilitated the development of an eDRS4C that offers improved capacity in accurately stratifying individuals with T1D, based on the setting in which that score is applied. Generative AI enhanced the Specificity of the outcome when 1K synthetic samples were used to train the model, while the use of 100K synthetic samples significantly improved the Sensitivity of predictions. Insulin independence at one-month post-transplantation can be predicted with high accuracy using eDRS4C models, and more intriguingly, a subset of 17 of the PREDICT T1D microRNAs stratified the best and the worst responders to imatinib therapy, at the study baseline.

Since these risk scores are built using circulating microRNA biomarkers, which can change in response to environment over time (i.e. “dynamic”), we also include the age at sample collection, which was identified as an important variable, in the development of these models. It has been known that early onset of T1D is more aggressive⁴⁸ and that children with T1D are less likely to retain detectable residual C-peptide than adults with T1D⁴⁹. Several studies^50-52 demonstrate the role of age in clinical progression and severity of T1D. DRS4C and eDRS4C showed a significant contribution (weighting) of age at sample collection in the development of these risk scores.

PREDICT T1D microRNAs capture β-cell development and function, which is supported through various other studies (reviewed in^53,54). These microRNAs are also reported in T1D circulation, targeting different biological and functional pathways including apoptosis, insulin synthesis, inflammation, NF-KB and the Jak-STAT pathway^41,55. Interestingly, multiple microRNAs from the PREDICT T1D signature are known to regulate inflammasome components⁵⁶, immune cell development/function, and central/peripheral tolerance^40,57. However, these microRNAs along with age at sampling have not been previously reported as a single biomarker signature. Although peripheral circulation is a ‘mixed bag’ of molecules and it is difficult to precisely identify the cells of origin of these microRNAs, a robust multi-context validation approach through this study emphasises the importance of the PREDICT T1D signature in T1D stratification.

An advantage that eDRS4C offer over static biomarkers²⁴ is that the microRNA-based risk score is dynamic and therefore, capable of predicting future changes in diabetes status. This is timely as there are several therapies that can now change the course of T1D progression. Static risk scores (e.g. GRS or PRS) are hardwired to our genome (i.e. constant over lifetime) and are not, on their own, able to predict change in clinical progression of T1D. More importantly, the capacity of PREDICT T1D microRNAs to prioritize islet transplant recipients for a subsequent transplant, and to stratify responsiveness to T1D intervention at the study baseline implies an exclusive opportunity for T1D therapy/trial participants. This approach lays foundation for the usability of dynamic risk scores from biomarkers like the PREDICT T1D signature, in personalised medicine for T1D. Indeed, similar dynamic biomarkers, such as those assessed using the circulating proteome have been recently shown to outperform polygenic risk score and HbA1c for T2D prediction⁵⁸.

Previous report⁵⁹ indicated that autoantibodies provided an AUC of 0.75, while family history was reported with an AUC between 0.51-0.56. Although GRS2 alone⁵⁹ offered ROC AUC of 0.73, a combined risk score (CRS) of GRS2, autoantibodies and family history gave an AUC>0.92 in their study cohort⁵⁹. The additive values of other biomarkers to eDRS4C needs to be addressed in future studies.

To the best of our knowledge this is the first study of this scale with a large number of multi-centre heterogeneous samples (n=3,008 samples from n=2850 participants), capturing multi-context information, of cross-sectional as well as longitudinal cohorts of T1D stratification and therapy. The study is unique in utilising AI-driven predictive methods to develop models that are validated in separate case-control cohorts. We also provide a novel workflow for exploring GAI to enhance T1D stratification algorithms.

Our discovery analysis (Figure 1) was designed to include microRNAs that are associated with β-cell development, adult islet cell function, as well as those that are differentially expressed in plasma of new-onset T1D participants. We also included microRNAs that did not reach statistical significance across our analyses but were reported in multiple publications from the microRNA biomarker community (Extended Data Table 1). The microRNAs identified within our discovery set are also captured across 25 different studies involving a total of n=2,869 samples (T1D:1,757 and Control:1,112). The potential of microRNAs identified through the efforts put together by us and others (Extended Data Table 1) led to the creation of this PREDICT T1D microRNA signature. There are more studies in recent years that report some of the PREDICT T1D microRNAs to be important in other larger cohorts⁶⁰. However, our signature was based on articles until 2020, when analyses on this finalised list of microRNAs were initiated. The validation of our eDRS4C and its application in two T1D therapies underscores the utility of this signature in spite of any unintentional variation in sampling procedures. Our discovery microRNA validation in >2200 individuals, is the largest validation cohort for such microRNA-based studies.

Our study limitations include the lack of other contexts (e.g. Africa, South America, Middle East), which we plan to build in the coming years. However, since GAI enables capturing most to all probable expression levels and ages of participants, the eDRS4C signatures were applicable across other contexts assessed (NZL, CAN, USA, Figure 3c). Clinical data including HbA1c, family history, C-peptide levels and T1D duration were not available for all samples in our study. Notably, since most of our study participants had been diagnosed with (stage 3) T1D, autoantibody measurements were not included. Future studies can be carried to assess the value of autoantibody and/or family history in the subset (DNK) of our 4C cohort for generating CRS to stratify T1D. This study was not designed to profile T1D endotypes^61,62, which is an ongoing interest. Similarly, we also plan to assess the potential of eDRS4C in stratifying T1D and T2D, given the increasing recognition of presenting in adulthood and potential misdiagnosis as T2D⁶³.

Despite these limitations, the PREDICT T1D signature and the eDRS4C have a lot to offer in clinical management of T1D. A foreseeable application is the usability in monitoring the changing risk of participants for clinical studies aimed to retard or prevent T1D progression. The profiling of the PREDICT T1D microRNA signature at the study baseline would aid in the personalised stratification of treating those that benefit the most. Since the microRNAs in PREDICT T1D signature target several other pathways (e.g. JAK-STAT pathway), we aim to extend our findings from the imatinib trial to other T1D drug intervention studies. These studies will lead to identifying changes in microRNA following drug intervention and the combination therapies that could be designed for T1D. It would also be important to capture multi-omics dynamic biomarker data and assess the additive value of these biomarkers to GRS⁶⁴ from larger, multi-ancestry cohorts.

In our hands, Random Forest workflow was the most suitable approach; however other ML workflows or biostatistical methods, or a combination of these, may be effective for some clinical datasets⁶⁵. Techniques such as Recursive Feature Elimination (RFE) could also be used for variable minimization/selection. Feature reduction is an important next step, as we are working to develop a low-cost hands-on kit, which could measure microRNA levels within minutes.

For individuals living with type 1 diabetes, the discovery and validation of such microRNA-based DRS and its application in predicting diabetes status or treatment efficacy in T1D therapies can offer better management and personalisation of treatment options. The ability to identify drug efficacies or monitor changes in future T1D status implies that treatments can be tailored to individual needs. The monitoring offered through such DRS will empower clinicians and their patients by providing timely insights into how their condition is evolving and how they are responding to treatments. Access to personalised risk scores and the ability to track changes over time allows people with T1D to take an active role in managing their condition.

Although our studies include individuals at risk of T1D (DNK Siblings) we lacked individuals progressing to autoimmunity in T1D. Future studies in longitudinal cohorts (e.g. ENDIA Study), and screening programs e.g. EDENT1FI (European action for the Diagnosis of Early Non-clinical Type 1 diabetes For disease Interception) aimed at identifying individuals at risk for T1D, might also benefit from microRNA-based risk scores. In the future, we also aim to incorporate other dynamic biomarkers (eg. cfDNA, metabolites, proteins, cytokines, lipids), to systematically evaluate their contribution, paving way for personalised medicine in T1D.

Study population:

The study samples are stored or freshly collected plasma and tissue samples from seven countries, and in some cases, across various collaborative centres within a country. Methods for plasma collection were checked to be as uniform as possible, and all plasma samples were freeze-thawed <1 time prior to this analysis. Details of samples (number, age, sex) included in different sets of this study are provided in Extended Data Table 2.

The wet-lab discovery set of samples included developing human pancreatic tissues (n=55), adult human islets and non-pancreatic tissues (total, n=189), plasma from individuals within 96h of acute presentation with T1D (n=5) and age as well as sex-matched healthy Controls (n=5). The microRNAs identified from the literature are provided in Extended Data Table 1 and their comparison with microRNAs identified from the wet-lab discovery samples (Extended Data Table 3) is presented in Figure 1b.

The plasma samples for model development were from individuals without T1D (“Controls”), or siblings (genetic risk) of individuals with T1D (“Siblings”) or individuals with T1D (“T1D”) from four contexts: AUS (total n=728, T1D n=519, Controls n=209), DNK (total n=608, T1D n=316, Siblings n=292), HKG (total n=238, T1D n=120, Controls n=118), and IND (total n=630, T1D n=497, Controls n=133). These were randomised to a Train set and a Test set as presented in Figure 2a.

To test the performance of the risk score in a separate set of samples including those from a different context (New Zealand/NZL), a case-control validation set (total n=313, T1D n=150, non-T1D n=163) from AUS, IND and NZL was used.

The eDRS4C models (1K, 10K and 100K) developed through this study were measured in plasma samples from Canadian (CAN) islet transplant recipients (15 recipients @1h post-Tx and 24h post-Tx samples) to assess future T1D status. The PREDICT T1D microRNAs were also assessed for their capacity to stratify drug efficacy in the imatinib intervention study¹⁷, where individuals with recent onset T1D were treated with imatinib (n=43) or placebo (n=21) for 26 weeks¹⁷ and their outcomes assessed at the end of 1 year.

Ethical approval:

Each study population was recruited and consented by the lead collaborators at each site according to the Declaration of Helsinki guidelines and approved by their respective institutional human ethics committees. Ethical approvals for the use of developing human pancreas, adult islets and non-pancreatic tissues were obtained from human ethics committees of the National Centre for Cell Science (India) and Sydney Local Health District, as detailed in our earlier publications.^34,66,67 Profiling of circulating microRNAs was approved through the human ethics committees at St Vincent’s Hospital Melbourne, Sydney Local Health District, University of Sydney and Western Sydney University.

MicroRNA profiling:

Tissue/plasma samples were collected using standardized procedures (<1 freeze-thaw cycle), received coded and frozen in a re-identifiable manner, and stored frozen at -80°C until use. Plasma samples were completely thawed and a 100µL aliquot was used for RNA isolation using an automated QIAcube-HT platform (Qiagen, Hilden, Germany) as described earlier^33,68. Samples with extensive (visible) haemolysis were excluded. Initially, 500μL of TRIzol reagent (ThermoFisher Scientific, USA), 10g of glycogen, spike-in microRNA control (ath-miR-172a), and 100µL chloroform (all from Sigma-Aldrich, Hamburg, Germany) were added to each sample. After centrifugation, the aqueous phase was transferred to the RNeasy column (RNeasy-HT Kit, Qiagen, Hilden, Germany) for total RNA isolation. For solid tissue samples (islet, non-islet pancreas and developing pancreas) a TRIzol-based manual RNA isolation workflow was implemented^29,34,67. RNA was quantitated on NanoDrop™ spectrophotometer (ThermoFisher Scientific, USA) and 10ng or 100ng of RNA was used for reverse transcription using microRNA RT kit (ThermoFisher Scientific, USA), spike-in microRNA control (ath-miR-159a) and Megaplex RT primer pool A and B (discovery samples profiled for 754 microRNAs). Custom RT primers were used for all samples profiled for the signature of 50 selected microRNAs and 6 control microRNAs. Preamplification and TaqMan-based microRNA qPCR were performed using the low sample input (LSI) protocol recommended by the manufacturer and described earlier³³. Real-time qPCR was carried out using TaqMan OpenArray® Human microRNA inventoried (754 microRNAs) or custom panels (specifically designed for the 56 PREDICT T1D microRNA set) on the QuantStudio 12K Flex platform (ThermoFisher Scientific, USA). MicroRNAs for haemolysis and repeats were assessed independently. Raw data files obtained after the TaqMan OpenArray® real-time qPCR were checked for sample loading and uploaded to the ThermoConnect tool for global normalisation (for discovery set). MicroRNA data that did not pass QC (with amplification score <1.24 and cycle value (Cq) confidence interval <0.6) were converted to undetectable (Ct=39). Other samples assessed on custom panels (train-test, validation, T1D therapy) were normalised using the spike-in controls added during RNA isolation and reverse transcription. Normalised Ct-values were converted to Fold-Over-Detectable (FOD; or transcript abundance) values⁶⁹ and were used in further analyses.

Data analysis:

Discovery set:

The relative abundance of the microRNAs across two groups (islet vs non-islet; T1D vs Controls) in the discovery samples was calculated using 2^∆∆Ct (FOD difference) and significantly different microRNAs (p-value <0.05) were identified using Welch’s t-test.

Train-Test set:

i. Samples from four different contexts: MicroRNA abundance of the 50 PREDICT T1D microRNAs was compared between T1D and Controls (for AUS, HKG and IND), or T1D and Siblings (for DNK) and for all contexts together, using the Kruskal-Wallis test followed by uncorrected Dunn’s test comparing the mean ranks.

ii. Model training and testing: This analysis was performed on samples from the four different contexts using the 50 PREDICT T1D microRNA signature data along with the participant’s age at sample collection as continuous independent variables. Participant’s T1D status was used as a categorical dependent variable, where “Controls/Siblings” was set as the reference/Control group and “T1D” was set as a case group. Each context datasets were randomised to 70% train (n=1,543) and 30% test (n=661) sets using sample.split() command in caTools (version 1.18.2) R package⁷⁰. The samples selected in the train and test sets were kept the same throughout analyses in R as well as during the independent validation in Python (see below). Classification models were built on training sets of four contexts together using random forest ensemble decision forest machine learning workflow, using R packages randomForest (v4.7-1.1) and Caret (v6.0-94)⁷¹ in R (v4.2.2)⁷² and R-studio (v2023.12.1-402).⁷³ To settle the class imbalances in training datasets, four different sampling methods, i.e. up sampling, down sampling, SMOTE (Synthetic Minority Over-sampling Technique) and ROSE (Random Over-Sampling Examples), were assessed with three different cross-validations (5-, 10-, or 50-fold). Up to Fifty-fold cross-validation, using SMOTE sampling and 2000 decision trees, was selected as the optimally tuned workflow, based on accuracy, the area under the curve (AUC), sensitivity, specificity, and F1-score (scripts made available via github/isletbiology). Predicted probabilities were calculated for individuals in train datasets using the predict() command from the package stats in R.^74,75 The receiver operator characteristics (ROC) curve analyses, AUCs and confidence intervals were computed using roc() and ci.auc() commands from the pROC package in R^76,77 to obtain threshold values. The best threshold value for classification was determined by the Youden index method using coords() command from pROC package in R.⁷⁷ Individuals with predicted probability greater than the Youden index threshold were classified as T1D, whilst those with probabilities lower than the Youden index classified as non-T1D. Model performance was also independently validated in python. The PyCaret (https://pycaret.org/) within a Jupyter notebook environment was used for model training and evaluation. Random Forest Classifier was tuned within the PyCaret framework. The trained model was evaluated on the Test set using various performance metrics provided by PyCaret and scikit-learn (Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011) libraries. These metrics include Accuracy, Precision, Recall, F1-Score, and the Area Under the ROC Curve (AUC). AUC plots and detailed classification metrics were generated using the evaluate_model() function in PyCaret and custom scripts using scikit-learn. The validation was performed using StratifiedKFold cross-validation with 10-fold cross-validation. To interpret the feature importance and visualize the impact of each feature on the predictions made by the model, SHAP (SHapley Additive exPlanations) values were computed and visualized (https://dl.acm.org/doi/10.5555/3295222.3295230). A SHAP explainer object was created for the trained Random Forest model using the Explainer class of the shap package. SHAP values were visualized using the shap.plots.beeswarm function.

iii. Synthetic data model training and testing: T1D-free participants from the train sets of four contexts (AUS, DNK, HKG, IND) were used to generate three different synthetic datasets of age and microRNA expression profiles for 1000, 10,000 or 100,000 synthetic individuals using the SingleTableMetadata module from the Synthetic Data Vault (SDV) library. A GaussianCopulaSynthesizer model, a component of the SDV library, was used with the provided metadata. The model was then trained on the original control dataset to capture the underlying dependencies and distributions of the data. Synthetic data for the control dataset was generated using the trained model, ensuring compliance with the specified constraints, including age limitations between 2 and 64 years. The three negative internal control microRNAs were confirmed/set to be undetectable in the synthetic dataset. The suitability of synthetic data for downstream tasks such as model training and evaluation was confirmed with Table Evaluator. These synthetic datasets were pooled with the original real-life train data of the four contexts (AUS, DNK, HKG, IND) to create augmented training datasets for the generation of eDRS4C models. Random forest models were trained on these three augmented datasets using the same tuning parameters mentioned above.

Validation set:

The different models of eDRS4C (from 1K, 10K and 10K synthetic sample sets) were used to classify individuals as T1D/non-T1D in an independent dataset (n=313) referred to as the “Validation set”. Predicted probabilities and class for this dataset were calculated using the methodologies described above.

T1D Therapy set:

Islet transplant samples: The eDRS4C risk score was used to predict their T1D management (need for exogenous insulin) based on the eDRS4C (1K) at 1h and 24h post-Tx.

Imatinib trial: Baseline microRNA expression of the 50 PREDICT T1D microRNAs was used in a dimensionality reduction Uniform Manifold Approximation and Projection (UMAP) algorithm using the Entropy criterion and Minkowski metric for a best-fit method using BioVinci data visualization package (BioTuring Inc., San Diego, CA).

The genomic location of each of the 50 PREDICT T1D microRNA identified using the starting base pair position for each of the microRNA (based on assay IDs from ThermoFisher Scientific) and verified through the NCBI Gene database (Extended Data Table 4) was used to mark the start of the gene coding the corresponding microRNAs. Shiny Circos⁷⁸ tool (https://venyao.xyz/shinycircos/) was used to visualize genomic locations for the 50 PREDICT T1D microRNAs (Extended Data Figure 2). Links connecting these microRNAs represent their context-specific co-expression (Spearman correlation coefficient >0.9 and p<0.05), which is coloured separately for each context. The width of each connecting line is arbitrarily set for visual enhancement. Sankey plots were created using https://sankeymatic.com/.

Contributions:

AAH served as study chair and wrote the first draft of the manuscript. Study contributions can be recognized specifically under the following headers - Conceptualization/planning: AAH, MVJ; Lab analyses: MVJ, RJF, CLM, AAH; Clinical data: SK, AAA, KD, AJJ, MEC, AMJS, WH, CSY, RCWM, ED, TJ, SEG, FP; Data Analysis: MVJ, SK, WKMW, PSK, RAK, VT, IA, AP, AAH; Funding acquisition: AAH, MVJ, SEG; Project administration: AAH, MVJ, WKMW; Writing (core committee): AAH, MVJ, WKMW, PSK; Writing/critical review/section contribution: All authors. Consumer perspective: RH, RJF. All authors were involved at different steps in the conduct of the study, and the collection and review of study data. AAH, MVJ, WKMW, PSK, NHTP and MC verified the data. The writing group had full access to all of the data and made the decision to submit for publication. All authors reviewed and commented on various versions of the paper. Core writing group made appropriate revisions to submit this manuscript.

The Predict T1D Study Group:

Australia: Mugdha V Joglekar, Wilson KM Wong, Pooja S Kunte, Reshmi A Kulkarni, Ryan J Farr, Cody L Maynard, Maria Virginia Pereira E Cotta, Charlotte Dong, Fahmida K Ema, Hrishikesh Hardikar, Madilyn Coles, Riley Hayward, Sathya Perera, Nhan Ho Trong Pham, Nirupa Sachithanandan, Sarang Satoor, Caroline Taylor, Anandwardhan A. Hardikar (Diabetes & Islet Biology Group, Sydney), Gilles Guillemin, (currently IPB University, Bogor, Indonesia), Ikhlak Ahmed, Ammira Alshabeeb-Akil (currently Sidra Medical Research Institute, Doha, Qatar), Sonia Isaacs, Maria Craig (UNSW, Sydney), Thomas Loudovaris, Helen Thomas (St Vincent’s Institute, Melbourne), David Martin (Strathfield Private Hospital, Sydney), Jenny Gamble (Centenary Institute, Sydney), Yoon Hi Cho, Kim Donaghue (Westmead Childrens’ Hospital Sydney), Suzette Coat, William Hague (University of Adelaide), Martha Lappas (Melbourne University), Elham Hosseini Beheshti, Georges Grau, Andrzej Januszewski, Emma Scott, Alicia Jenkins (University of Sydney), Wayne Hawthorne (Westmead Institute of Medical Research, Westmead), Nirubasini Paramalingam, Chontiey Saxon, Elizabeth Davies, Tim Jones (Telethon Institute, University of Western Australia, Perth).

Canada: Indri Purwana, Saira Qureshi, Peter Senior, James Shapiro (University of Alberta, Edmonton).

Denmark: Simranjeet Kaur, Flemming Pociot (Steno Diabetes Center Copenhagen).

Hong Kong PRC China: Guozhi Jiang, Juliana Chan, Ronald Ma (The Chinese University of Hong Kong).

India: Dattatray Bhat, Aboli Bhalerao, Alma Baptist, Smita Dhadge, Vidya Gokhale, Kalpana Jog, Tejas Limaye, Neelima Thuse, Chittaranjan Yajnik (KEM Hospital, Pune), Rohan R Patil, Mahesh S Karandikar (Department of Physiology, DY Patil University, Pune), Vinod Thorat, Sharda Bapat, Aniruddha Pant (AlgoAnalytics, Pune), Sheela Joglekar, Vinay Joglekar (Shree Seva Medical Foundation, Shirwal).

New Zealand: Janet Rowan (University of Auckland).

USA: Noha Lim, Stephen Gitelman (University of California San Francisco).

Declaration of Interests:

The authors declare no competing financial interests related to the current manuscript. MVJ declares research funding from JDRF International (Advanced post-doctoral fellowship: 3-APF-2016-178-A-N and JDRF Transition Award: 1-FAC-2021-1063-A-N), Australia India Early and Midcareer Fellowship from Australian Academy of Science, FP has funding from INNODIA (IMI2-JU No 115797) and INNODIA HARVEST (IMI2-JU No 948268), SEG indicates funding from Juvenile Diabetes Research Foundation (Gleevec Study) and AAH declares funding from JDRF Australia (1-CDA-2021-1084-M-B), JDRF International (3-SRA-2022-1263-S-B), the Helmsley Charitable Trust (3-SRA-2019-694-M-B and 1-SRA-2021-1067-M-B), the NHMRC Ideas grant (2021/GNT2011557) and Western Sydney University infrastructure support through the Ainsworth Medical Research Funding.

Data sharing

Data collected for the study and presented herein will be made available to others for collaborative research, but deidentified for Participant level data. Related study documents, including Python and R codes, lab SOPs, and ethics documents will also be available. All study methods and scripts will be shared via github or through the Study PI (AAH). Requests should be sent by direct email to AAH ([email protected]).

Declaration of generative AI and AI-assisted technologies in the writing process

Generative AI was not used for creating the textual paragraphs of this submission.

Acknowledgements:

Authors acknowledge the contribution from cadaveric tissue donors, study participants, their families and all support staff across multiple study sites in Australia, Canada, Denmark, Hong Kong PRC China, India, United States of America and New Zealand. Authors acknowledge approvals through HMSC (KEMHRC ID:1510), Government of India, for study samples from IND, the Physiology Department, DY Patil University, Pune and the KEM Hospital & Research Centre, Pune, for their infrastructure through Hinduja Foundation, Mumbai, Mukul Madhav Foundation, Pune and Nityasha Foundation, Pune. AAH also acknowledges infrastructure support through University of Sydney and The Western Sydney University School of Medicine, the Ainsworth Medical Research Funding support and from all of the study team/collaborators on this project.

Mathis, D., Vence, L. & Benoist, C. beta-Cell death during progression to diabetes. Nature 414, 792-798 (2001).
Borchers, A.T., Uibo, R. & Gershwin, M.E. The geoepidemiology of type 1 diabetes. Autoimmun Rev 9, A355-365 (2010).
Atkinson, M.A. & Gianani, R. The pancreas in human type 1 diabetes: providing new answers to age-old questions. Curr Opin Endocrinol Diabetes Obes 16, 279-285 (2009).
Ziegler, A.G., et al. Seroconversion to multiple islet autoantibodies and risk of progression to diabetes in children. JAMA 309, 2473-2479 (2013).
Barker, J.M., et al. Two single nucleotide polymorphisms identify the highest-risk diabetes HLA genotype: potential for rapid screening. Diabetes 57, 3152-3155 (2008).
Oram, R.A., et al. A Type 1 Diabetes Genetic Risk Score Can Aid Discrimination Between Type 1 and Type 2 Diabetes in Young Adults. Diabetes Care 39, 337-344 (2016).
Sharp, S.A., et al. Development and Standardization of an Improved Type 1 Diabetes Genetic Risk Score for Use in Newborn Screening and Incident Diagnosis. Diabetes Care 42, 200-207 (2019).
Battaglia, M. & Atkinson, M.A. The streetlight effect in type 1 diabetes. Diabetes 64, 1081-1090 (2015).
Bonifacio, E. Predicting type 1 diabetes using biomarkers. Diabetes Care 38, 989-996 (2015).
Pociot, F. & Lernmark, A. Genetic risk factors for type 1 diabetes. Lancet 387, 2331-2339 (2016).
Ferrat, L.A., et al. A combined risk score enhances prediction of type 1 diabetes among susceptible children. Nat Med 26, 1247-1255 (2020).
Elding Larsson, H., et al. Children followed in the TEDDY study are diagnosed with type 1 diabetes at an early stage of disease. Pediatr Diabetes 15, 118-126 (2014).
Steck, A.K. & Rewers, M.J. Genetics of type 1 diabetes. Clin Chem 57, 176-185 (2011).
Redondo, M.J., et al. Genetic determination of islet cell autoimmunity in monozygotic twin, dizygotic twin, and non-twin siblings of patients with type 1 diabetes: prospective twin study. BMJ 318, 698-702 (1999).
Hyttinen, V., Kaprio, J., Kinnunen, L., Koskenvuo, M. & Tuomilehto, J. Genetic liability of type 1 diabetes and the onset age among 22,650 young Finnish twin pairs: a nationwide follow-up study. Diabetes 52, 1052-1055 (2003).
Herold, K.C., et al. An Anti-CD3 Antibody, Teplizumab, in Relatives at Risk for Type 1 Diabetes. N Engl J Med 381, 603-613 (2019).
Gitelman, S.E., et al. Imatinib therapy for patients with recent-onset type 1 diabetes: a multicentre, randomised, double-blind, placebo-controlled, phase 2 trial. Lancet Diabetes Endocrinol 9, 502-514 (2021).
Quattrin, T., et al. Golimumab and Beta-Cell Function in Youth with New-Onset Type 1 Diabetes. N Engl J Med 383, 2007-2017 (2020).
Rigby, M.R., et al. Alefacept provides sustained clinical and immunological effects in new-onset type 1 diabetes patients. J Clin Invest 125, 3285-3296 (2015).
Russell, W.E., et al. Abatacept for Delay of Type 1 Diabetes Progression in Stage 1 Relatives at Risk: A Randomized, Double-Masked, Controlled Trial. Diabetes Care 46, 1005-1013 (2023).
Waibel, M., et al. Baricitinib and beta-Cell Function in Patients with New-Onset Type 1 Diabetes. N Engl J Med 389, 2140-2150 (2023).
Nakayasu, E.S., et al. The role of proteomics in assessing beta-cell dysfunction and death in type 1 diabetes. Expert Rev Proteomics 16, 569-582 (2019).
Quattrin, T., et al. 3-LB: Golimumab (GLM) Preserves ß-Cell Function and Reduces Insulin Use and Hypoglycemia in Children and Young Adults with Recently Diagnosed Type 1 Diabetes (T1D): The Phase 2 T1GER Study. in American Diabetes Association (2020).
Joglekar, M.V., Kaur, S., Pociot, F. & Hardikar, A.A. Predicting future type 1 diabetes using dynamic biomarkers from the circulation. Lancet Diabetes Endocrinol in press(2024).
Warncke, K., et al. Elevations in blood glucose before and after the appearance of islet autoantibodies in children. J Clin Invest 132(2022).
Salami, F., et al. HbA1c as a time predictive biomarker for an additional islet autoantibody and type 1 diabetes in seroconverted TEDDY children. Pediatr Diabetes 23, 1586-1593 (2022).
Helminen, O., et al. HbA1c Predicts Time to Diagnosis of Type 1 Diabetes in Children at Risk. Diabetes 64, 1719-1727 (2015).
Helminen, O., et al. OGTT and random plasma glucose in the prediction of type 1 diabetes and time to diagnosis. Diabetologia 58, 1787-1796 (2015).
Joglekar, M.V., Joglekar, V.M. & Hardikar, A.A. Expression of islet-specific microRNAs during human pancreatic development. Gene Expr Patterns 9, 109-113 (2009).
Poy, M.N., et al. A pancreatic islet-specific microRNA regulates insulin secretion. Nature 432, 226-230 (2004).
Samandari, N., et al. Circulating microRNA levels predict residual beta cell function and glycaemic control in children with type 1 diabetes mellitus. Diabetologia 60, 354-363 (2017).
Wei, R., et al. Dynamic expression of microRNAs during the differentiation of human embryonic stem cells into insulin-producing cells. Gene 518, 246-255 (2013).
Joglekar, M.V., et al. Circulating microRNAs from early childhood and adolescence are associated with pre-diabetes at 18 years of age in women from the PMNS cohort. J Dev Orig Health Dis 13, 806-811 (2022).
Wong, W.K.M., et al. Machine learning workflows identify a microRNA signature of insulin transcription in human tissues. iScience 24, 102379 (2021).
Yu, D., et al. Roquin represses autoimmunity by limiting inducible T-cell co-stimulator messenger RNA. Nature 450, 299-303 (2007).
Asirvatham, A.J., Gregorie, C.J., Hu, Z., Magner, W.J. & Tomasi, T.B. MicroRNA targets in immune genes and the Dicer/Argonaute and ARE machinery components. Mol Immunol 45, 1995-2006 (2008).
Voinnet, O. Induction and suppression of RNA silencing: insights from viral infections. Nat Rev Genet 6, 206-220 (2005).
Joglekar, M.V., Parekh, V.S. & Hardikar, A.A. New pancreas from old: microregulators of pancreas regeneration. Trends Endocrinol Metab 18, 393-400 (2007).
Snowhite, I.V., et al. Association of serum microRNAs with islet autoimmunity, disease progression and metabolic impairment in relatives at risk of type 1 diabetes. Diabetologia 60, 1409-1422 (2017).
Scherm, M.G., Serr, I., Kaestner, K.H. & Daniel, C. The role of T cell miRNAs for regulatory T cell induction in islet autoimmunity. Mol Metab 27S, S122-S128 (2019).
Dotta, F., Ventriglia, G., Snowhite, I.V. & Pugliese, A. MicroRNAs: markers of beta-cell stress and autoimmunity. Curr Opin Endocrinol Diabetes Obes 25, 237-245 (2018).
Syed, F. & Evans-Molina, C. Nucleic acid biomarkers of beta cell stress and death in type 1 diabetes. Curr Opin Endocrinol Diabetes Obes 23, 312-317 (2016).
Liu, R., et al. MicroRNA-21 promotes pancreatic beta cell function through modulating glucose uptake. Nat Commun 13, 3545 (2022).
Wong, W., Farr, R., Joglekar, M., Januszewski, A. & Hardikar, A. Probe-based Real-time PCR Approaches for Quantitative Measurement of microRNAs. J Vis Exp (2015).
Yi, L., Swensen, A.C. & Qian, W.J. Serum biomarkers for diagnosis and prediction of type 1 diabetes. Transl Res 201, 13-25 (2018).
Sims, E.K., Evans-Molina, C., Tersey, S.A., Eizirik, D.L. & Mirmira, R.G. Biomarkers of islet beta cell stress and death in type 1 diabetes. Diabetologia 61, 2259-2265 (2018).
Brenu, E.W., Harris, M. & Hamilton-Williams, E.E. Circulating biomarkers during progression to type 1 diabetes: A systematic review. Front Endocrinol (Lausanne) 14, 1117076 (2023).
Rawshani, A., et al. Excess mortality and cardiovascular disease in young adults with type 1 diabetes in relation to age at onset: a nationwide, register-based cohort study. Lancet 392, 477-486 (2018).
Januszewski, A.S., et al. Insulin micro-secretion in Type 1 diabetes and related microRNA profiles. Sci Rep 11, 11727 (2021).
So, M., et al. Characterising the age-dependent effects of risk factors on type 1 diabetes progression. Diabetologia 65, 684-694 (2022).
Leete, P., et al. The Effect of Age on the Progression and Severity of Type 1 Diabetes: Potential Effects on Disease Mechanisms. Curr Diab Rep 18, 115 (2018).
Marcovecchio, M.L., et al. The INNODIA Type 1 Diabetes Natural History Study: a European cohort of newly diagnosed children, adolescents and adults. Diabetologia 67, 995-1008 (2024).
Osmai, M., et al. MicroRNAs as regulators of beta-cell function and dysfunction. Diabetes Metab Res Rev 32, 334-349 (2016).
Grieco, G.E., et al. The Landscape of microRNAs in betaCell: Between Phenotype Maintenance and Protection. Int J Mol Sci 22(2021).
Assmann, T.S., Recamonde-Mendoza, M., De Souza, B.M. & Crispim, D. MicroRNA expression profiles and type 1 diabetes mellitus: systematic review and bioinformatic analysis. Endocr Connect 6, 773-790 (2017).
Boxberger, N., Hecker, M. & Zettl, U.K. Dysregulation of Inflammasome Priming and Activation by MicroRNAs in Human Immune-Mediated Diseases. J Immunol 202, 2177-2187 (2019).
Simpson, L.J. & Ansel, K.M. MicroRNA regulation of lymphocyte tolerance and autoimmunity. J Clin Invest 125, 2242-2249 (2015).
Gadd, D.A., et al. Blood protein assessment of leading incident diseases and mortality in the UK Biobank. Nat Aging 4, 939-948 (2024).
Ferrat, L.A., et al. A combined risk score enhances prediction of type 1 diabetes among susceptible children. Nat Med 26, 1247-1255 (2020).
Sebastiani, G., et al. A set of circulating microRNAs belonging to the 14q32 chromosome locus identifies two subgroups of individuals with recent-onset type 1 diabetes. Cell Rep Med 5, 101591 (2024).
Redondo, M.J. & Morgan, N.G. Heterogeneity and endotypes in type 1 diabetes mellitus. Nat Rev Endocrinol 19, 542-554 (2023).
Keenan, H.A., et al. Residual insulin production and pancreatic ss-cell turnover after 50 years of diabetes: Joslin Medalist Study. Diabetes 59, 2846-2853 (2010).
Harding, J.L., et al. The Incidence of Adult-Onset Type 1 Diabetes: A Systematic Review From 32 Countries and Regions. Diabetes Care 45, 994-1006 (2022).
Redondo, M.J., et al. Type 1 diabetes in diverse ancestries and the use of genetic risk scores. Lancet Diabetes Endocrinol 10, 597-608 (2022).
Lynam, A.L., et al. Logistic regression has similar performance to optimised machine learning algorithms in a clinical setting: application to the discrimination between type 1 and type 2 diabetes in young adults. Diagn Progn Res 4, 6 (2020).
Joglekar, M.V., et al. A Pro-Endocrine Pancreatic Islet Transcriptional Program Established During Development Is Retained in Human Gallbladder Epithelial Cells. Cell Mol Gastroenterol Hepatol 13, 1530-1553 e1534 (2022).
Wong, W.K., et al. The long noncoding RNA MALAT1 predicts human pancreatic islet isolation quality. JCI Insight 5(2019).
Joglekar, M.V., et al. Postpartum circulating microRNA enhances prediction of future type 2 diabetes in women with previous gestational diabetes. Diabetologia 64, 1516-1526 (2021).
Hardikar, A.A., Farr, R.J. & Joglekar, M.V. Circulating microRNAs: understanding the limits for quantitative measurement by real-time PCR. J Am Heart Assoc 3, e000792 (2014).
Tuszynski, J. caTools: Tools: Moving Window Statistics, GIF, Base64, ROC, AUC, etc. (2022).
Kuhn, M. Building Predictive Models in R Using the caret Package. Journal of Statistical Software 28, 1 - 26 (2008).
Team, R.C. R: A Language and Environment for Statistical Computing, (2022).
Team, R. RStudio Team (2020). RStudio: Integrated Development for R. RStudio. (RStudio, PBC., 2020).
Everitt, B. Book reviews : Chambers JM, Hastie TJ eds 1992: Statisti cal models in S. California: Wadsworth and Brooks/Cole. ISBN 0 534 16765-9. Statistical Methods in Medical Research 1, 220-221 (1992).
Richard A. Becker, J.M.C., Allan Reeve Wilks. The New S Language, (Wadsworth & Brooks/Cole, 1988).
Fawcett, T. An introduction to ROC analysis. Pattern Recognition Letters 27, 861-874 (2006).
Robin, X., et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 12, 77 (2011).
Wang, Y., et al. shinyCircos-V2.0: Leveraging the creation of Circos plot with enhanced usability and advanced features. iMeta 2, e109 (2023).

There is NO Competing Interest. The authors declare no competing financial interests related to the current manuscript.

ExtendedDataTable5.xlsx
Extended Data Table 5
Extendeddataset070924.docx
Extended Data Tables and Figures

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Applicability of a microRNA-based dynamic risk score for type 1 diabetes

Status:

Version 1

Abstract

Figures

Introduction

Results

Discussion

Online Methods

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1