Integration of Infant Metabolite, Genetic, and Islet Autoimmunity Signatures to Predict Type 1 Diabetes by Age 6 Years

Abstract

Context

Biomarkers that can accurately predict risk of type 1 diabetes (T1D) in genetically predisposed children can facilitate interventions to delay or prevent the disease.

Objective

This work aimed to determine if a combination of genetic, immunologic, and metabolic features, measured at infancy, can be used to predict the likelihood that a child will develop T1D by age 6 years.

Methods

Newborns with human leukocyte antigen (HLA) typing were enrolled in the prospective birth cohort of The Environmental Determinants of Diabetes in the Young (TEDDY). TEDDY ascertained children in Finland, Germany, Sweden, and the United States. TEDDY children were either from the general population or from families with T1D with an HLA genotype associated with T1D specific to TEDDY eligibility criteria. From the TEDDY cohort there were 702 children will all data sources measured at ages 3, 6, and 9 months, 11.4% of whom progressed to T1D by age 6 years. The main outcome measure was a diagnosis of T1D as diagnosed by American Diabetes Association criteria.

Results

Machine learning–based feature selection yielded classifiers based on disparate demographic, immunologic, genetic, and metabolite features. The accuracy of the model using all available data evaluated by the area under a receiver operating characteristic curve is 0.84. Reducing to only 3- and 9-month measurements did not reduce the area under the curve significantly. Metabolomics had the largest value when evaluating the accuracy at a low false-positive rate.

Conclusion

The metabolite features identified as important for progression to T1D by age 6 years point to altered sugar metabolism in infancy. Integrating this information with classic risk factors improves prediction of the progression to T1D in early childhood.

The development of type 1 diabetes (T1D) is driven by an interaction between genetic and environmental factors. The relationships and roles of human leukocyte antigen (HLA) and other genes as they affect development of islet autoimmunity and subsequent progression to T1D continues to be refined (1-4). Environmental and biomarker discovery research, as well as examination of the interplay between potential risk factors and gene variants, has provided insights into T1D risk and potential pathogenic mechanisms (5-11).

The Environmental Determinants of Diabetes in the Young (TEDDY) study has followed thousands of children who are at increased genetic risk of T1D and has collected diverse data, such as infant characteristics, family history, diet, genetics, islet autoantibodies (IAAbs), and metabolomics. Previous analyses have generated a considerable list of risk factors associated with T1D, such as high-risk genotypes, genetic risk scores (GRS) computed from T1D-associated single-nucleotide variations (SNVs; formerly single-nucleotide polymorphisms [SNPs]), prebiotic or probiotic exposure, and timing of gluten exposure. A key goal of TEDDY, as well as other large cohort studies, is to develop models to predict onset of T1D. Diagnosing children as early as possible has the potential to reduce the risk of diabetic ketoacidosis at onset as well as reduce the risk of subsequent complications by blunting initial hyperglycemia and reducing subsequent glucose excursions through improved glycemic control (12-15). One approach to address the challenge of prediction is machine learning, which can build mathematical models to discriminate individuals into groups based on multiple risk factors and observed data for defined outcomes of interest (16-22).

Machine learning predictive models are at the center of precision medicine because they can predict the future outcome of an individual as a probability estimate based on the current input data for that person. Within individual binary risk factors, such as HLA genotype, all patients are essentially given a binary probability, the same probability for all individuals if they have the at-risk HLA genotype. By extending this concept to quantitative variables, such as GRS (23), more refined probabilities can be assigned using classical statistical machine learning methods, such as logistic regression (4). Adding additional risk factors to the machine learning model can continue to refine these individual-based predictions and, depending on the discriminatory power of the feature, can either increase or decrease the overall accuracy of the model. Thus, these models can be interrogated to identify the specific risk factors that work well together, as a multivariate panel, to enable segregation of the class of interest. In the context of T1D research, predictive modeling, to date, has mostly focused on the separation of controls from those with T1D based on genetics (4, 23). However, more recently, prognostic evaluations using other measures have become common as they have clear applicability to screening in high-risk children (6).

Prior work in T1D birth cohorts have demonstrated associations between genetic background and specific environmental exposures with T1D-related outcomes of interest, specifically for children younger than 6 years, based on genetic screening and the prospective follow-up information collected on the TEDDY enrolled children beginning at birth (6). Herein, we use machine learning to evaluate the potential of multiple sources of information on each TEDDY participant collected during infancy, specifically before age 9 months, to predict the likelihood they will develop T1D by age 6 years. To make the predictions we explore a combination of demographic data, such as sex, family history, and dietary information, in combination with genetic and untargeted plasma metabolomic profiles. Although the statistical association of the risk factors used have been evaluated via multiple studies, this is the first attempt to integrate all these factors with metabolomics into a single machine learning model from which the probability of development of T1D by age 6 years can be used to evaluate accuracy. We then evaluate prediction models at each time point to determine the benefit of screening at 1, 2, or 3 time points, as well as the age at which the screening is performed. The data-driven feature selection approach allows a collection of specific demographics, dietary, genetic, and metabolomic features to be identified and used to accurately classify children at age 9 months into their 6-year T1D risk outcome.

Materials and Methods

Study Design and Measurements

TEDDY is a prospective cohort study following children recruited from the general population at birth, based on having increased risk of T1D, identified through T1D-associated high-risk HLA genotypes or a relative with T1D. TEDDY participants were recruited before age 4.5 months at 6 study sites in 4 countries: the United States, Germany, Sweden, and Finland (24). The ethics committee or institutional review board approved the TEDDY study as applicable to each country. Written informed consents were obtained from a parent or primary caregiver for all participants for genetic screening and the prospective follow-up separately. Participants were evaluated for the development of islet autoimmunity every 3 months thereafter until either the development of T1D or age 4 years. After age 4 years, those with autoantibody seroconversion continued visits every 3 months and the remainder were evaluated every 6 months. We used features that were collected before or at age 9 months, including those participants with clinic visits at approximately age 3, 6, and 9 months. As seen in Fig. 1, there are 8676 children enrolled in TEDDY. This analysis focused on the 1843 distinct children selected for the 1:3 matched, nested, case-control study used for omics analyses (25). Of these 1843 children, 655 had complete data available from the TEDDY Data Coordinating Center, including demographic data (eg, infant diet, family history), birth measurements, GRS, HLA genotypes, IAAb status at age 9 months, and metabolomics data at age 3, 6, and 9 months. T1D was diagnosed using American Diabetes Association criteria (25, 26).

Figure 1.

Flowchart of selection of children from The Environmental Determinants of Diabetes in the Young (TEDDY) cohort.

General infant attributes

There were 24 factors that do not require an assay-based test that are considered as the general infant attributes (GIA) data set. This included an initial set of 4 infant birth data (sex, gestational age, birth length, birth weight), 12 growth measures (height and weight at each of the 3 time points, as well as the change from ages 3 to 6, 3 to 9, and 6 to 9 months), 3 family history data points (any first-degree relative with T1D [FDR-T1D], father with T1D, mother with T1D), and 5 dietary variables (exposure to formula with cow milk by either age 28 days or 6 months, exposure to prebiotics or probiotics by either age 28 days or 6 months, gluten exposure by age 6 months). We tested if the distributions of samples from the 6 clinical centers (Colorado, Georgia, and Washington, USA; and Finland, Germany, and Sweden) were different between our progressors and nonprogressors to T1D by age 6 years. There was not a significant difference (P value = ~0.58; χ² test of independence) between the fraction from each center that progressed to T1D by age 6 years and those that did not.

Islet autoantibody measurements

Radiobinding assays in 2 laboratories (Barbara Davis Center, Aurora, Colorado, USA, and the University of Bristol Laboratory, Bristol, UK) were used to measure islet autoantibodies; insulin autoantibody (IAA), glutamic acid decarboxylase autoantibody (GADA), and insulinoma-associated antigen 2 autoantibody (IA-2A) as previously described (27, 28). In the 2020 Islet Autoantibody Standardization Workshop, the sensitivities were at minimum 99% for all assays and the specificities were 62%, 78%, and 72% for IAA, GADA, and IA-2A, respectively (29). Children who had 2 or more consecutive confirmed positive samples were defined as persistently IAAb positive unless it was determined to be due to maternal transfer or they developed T1D before the next sample collection. This data set is defined with a categorical variable as either IAAb positive or negative as defined earlier for IAA, GADA, or IA-2A at their 9-month visit. The 3- and 6-month time points were excluded because of small counts (< 5 participants).

Genetic risk score and human leukocyte antigen

When the children were aged 9 to 12 months, the HLA-DR and HLA-DQ genotypes were confirmed by reverse blot hybridization at the central HLA Reference Laboratory at Roche Molecular Systems as previously described (30). For this study these genotypes were translated to 5 categorical variables (DR3/4 [HLA-DR3/HLA-DR4], DR4/4, DR3/3, DR4/8, or other) as previously described (5). The T1D GRS as a single quantitative variable for each TEDDY participant was computed as a weighted sum across 41 T1D-associated non-HLA region SNVs (effect size as weights for the SNV) as previously described (23).

Metabolomics

Untargeted plasma metabolomics were carried out across multiple time points for TEDDY participants based on an existing nested, case-control design (25). This analysis profiled 10 522 plasma samples in which primary metabolites were quantified from citrated plasma using gas chromatography–time-of-flight mass spectrometry (GC-TOF MS) at the NIH West Coast Metabolomics Center at the University of California, Davis. The GC-TOF MS metabolomics data acquisition followed previously described protocols (31) followed by data processing and compound identification using the BinBase algorithm (32) and normalization using Random Forest normalization (33). From these data we specifically extracted the children who had complete quantified metabolomics data collected at the 3-, 6-, and 9-month visits, which yielded a total of 139 known metabolites. All metabolomics data were analyzed as relative abundance on the log₂ scale.

Statistical Analysis

Statistical analyses were performed for initial feature filtering (34, 35). For features that were normally distributed, which included the T1D GRS, a standard 2-sample t test was performed. For all categorical features, a Fisher exact test was performed. For those features that did not have a normal distribution (gestational age, growth data, log₂-transformed metabolomics data,), a Wilcoxon rank sum test was performed. All statistics were completed using MatLab (R2019a) software.

Machine Learning

Machine learning was performed with a naive Bayes classifier in MatLab (R2019a). Results for each participant were evaluated in the context of the predicted probabilities from 5-fold cross-validation (CV). Models were assessed based on a receiver operating characteristic (ROC) curve, for which the area under the ROC curve (AUC) serves as an overall metric of performance at all possible classification thresholds. A random classifier returns an AUC of 0.5 and a perfect classifier yields an AUC of 1.0. Feature-importance matrices were generated using Repeated Optimization for Feature Interpretation (ROFI) (20, 36). ROFI uses statistical sampling and optimization to identify subsets of features that have a multivariate representation that increases the ability to predict the TEDDY participants who develop T1D from those that do not by age 6 years. The repeated analyses component of ROFI yields importance metrics for each feature that describes the likelihood that a sample will be included as important to the machine learning model. ROFI was implemented using 20 repetitions of 5-fold CV of naive Bayes within a simulated annealing optimization routine (v = 1E-3), convergence criteria of 1E-5, AUC as the optimization metric, and 100 repetitions of the model output. The final subset of features was selected based on frequency of inclusion in the optimized model as previously described with code available in the peppuR package to implement ROFI (https://github.com/pmartR/peppuR) (22).

Results

Summary cohort characteristics of the 31 nonmetabolite features for the 655 TEDDY participants included in this study are given in Table 1. Of the 655 children, 73 (11.2%) developed T1D by age 6 years. Before the machine learning analysis, the 139 metabolites at each of the 3 time points and the 31 features in Table 1 were subjected to a statistics filter to remove features that do not show any statistical association with diagnosis of T1D by age 6 years (34, 35). Since the goal was to evaluate the utility of multiple risk factors in the context of machine learning, an uncorrected and liberal threshold (P < .1) was used for filtering to retain those that may have a weak statistical association but may be important in the context of a multivariate signature. There were 23 time-based metabolite abundances also selected at this threshold; 8 metabolites at age 3 months, 11 metabolites at age 6 months, and 4 metabolites at age 9 months. Five of the 24 GIA features were retained: FDR-T1D, FDR-T1D (father), gestational age, weight at age 9 months, and exposure to formula with cow milk by age 6 months. IAAb status at age 9 months was positively associated with progression to T1D by age 6 years, as well as the T1D GRS and 2 of the HLA genotypes (DR3/4 and DR4/4). The largest Pearson correlation between any 2 quantitative features was approximately 0.5, suggesting that duplicative features should not dramatically affect model generation and feature selection. Within the qualitative features, only the family history metrics (FDR-T1D and FDR-T1D [father]) had high similarity as FDR-T1D (father) is a subset of FDR-T1D.

Table 1.

Characteristics of The Environmental Determinants of Diabetes in the Young (TEDDY) participants categorized for machine learning based on type 1 diabetes outcome at age 6 years

Data	Feature	Positive T1D at age 6 y	Negative T1D at age 6 y	P
	No. of participants	73	582
GIA	Female	40 (54.8%)	263 (45.2%)	.136
GIA	T1D first-degree relative	24 (32.9%)	124 (21.3%)	.037
GIA	T1D first-degree relative is mother	5 (6.8%)	41 (7.0%)	≥ .999
GIA	T1D first-degree relative is father	14 (19.2%)	64 (11.0%)	.054
GIA	Gestational Age, wk	39.57	40.00	.006
GIA	Length at birth, cm	50.80	51.00	.991
GIA	Length at age 3 mo	62.60	62.50	.641
GIA	Length at age 6 mo	68.60	68.40	.500
GIA	Length at age 9 mo	72.60	72.50	.458
GIA	Growth age 3 to 6 mo	5.60	50.50	.876
GIA	Growth age 3 to 9 mo	10.0	10.0	.861
GIA	Growth age 6 to 9 mo	4.50	4.40	.861
GIA	Weight at birth, kg	3.60	3.53	.214
GIA	Weight at age 3 mo	6.57	6.51	.228
GIA	Weight at age 6 mo	8.13	8.03	.128
GIA	Weight at age 9 mo	9.24	9.16	.084
GIA	Weight gain at age 3 to 6 mo	1.52	1.48	.246
GIA	Weight gain at age 3 to 9 mo	2.64	2.52	.250
GIA	Weight gain at age 6 to 9 mo	1.06	1.05	.477
GIA	Formula (cow milk) before 28 d	33 (42.2%)	272 (46.7%)	.901
GIA	Formula (cow milk) before 6 mo	48 (65.8%)	451 (77.5%)	.040
GIA	Formula (prebiotic or probiotic) before age 28 d	16 (21.9%)	120 (20.6%)	.761
GIA	Formula (prebiotic or probiotic) before age 6 mo	25 (34.2%)	203 (34.9%)	≥ .999
GIA	Gluten before age 6 mo	30 (41.1%)	265 (45.5%)	.533
GRS	GRS	10.5	10.2	.002
HLA	DR3/4	43 (58.9%)	217 (37.3%)	4.37E-16
HLA	DR4/4	8 (11.0%)	117 (20.1%)	.081
HLA	DR3/3	8 (11.0%)	90 (15.5%)	.385
HLA	DR4/8	8 (11.0%)	106 (18.2%)	.142
HLA	Other	6 (8.2%)	52 (8.9%)	≥ .999
IAAb	Persistent islet autoantibody positive at age 9 mo	23 (31.5%)	11 (1.9%)	.001

Numbers for qualitative features are percentages as evaluated by Fisher exact test and for qualitative features are the median as evaluated by Wilcoxon rank sum test with the exception of the GRS, which is the mean and evaluated via 2-sample t test. Features in bold italics are used in machine learning.

Abbreviations: GIA, general infant attributes; GRS, genetic risk score; HLA, human leukocyte antigen; T1D, type 1 diabetes.

Feature Selection

We used ROFI to acquire feature-importance metrics for each of the 32 features included in the initial modeling. Fig. 2 gives the importance metric of each feature, which is the likelihood of inclusion in the model when optimizing AUC within a naive Bayes classifier. A common threshold used in this approach is 50%, which did provide a higher overall accuracy with respect to the AUC (inset to Fig. 2) in this setting. At a ROFI feature-importance threshold of 50%, there were 16 total features. As seen in the inset to Fig. 2, the AUC did not increase when including additional features after these 16 features (red triangle). The final set of 16 features reduced the feature set to 3 genetic and the immunologic features: DR3/4, T1D GRS, and IAAb positivity at age 9 months, 1 infant characteristic (gestational age), 1 dietary marker (cow milk formula < age 6 months), and 11 time-specific metabolite measurements.

Figure 2.

Importance of each feature to the predictive model for all 32 features where brown indicates above the 50% threshold (16 features). Of the 16 features there were 5 nonmetabolite features (islet autoantibody [IAAb] at age 9 months, DR3/4, genetic risk score [GRS], gestational age, and exposure to cow’s milk before age 6 months), 3 metabolites measured at age 3 months, 5 metabolites measured at age 6 months, and 3 metabolites measured at age 9 months. Inset is the area under the curve (AUC) value achieved by feature reduction using ROFI-based thresholds where the x-axis on the left indicates the Repeated Optimization for Feature Interpretation (ROFI) feature importance score and the y-axis on the right indicates the average AUC of the model based on the threshold. The ROFI-based threshold of 50%, brown square, is a common threshold in this is case also relates to a near optimal AUC.

Screening Age and Sampling Time for Metabolite Measurements

The feature selection in Fig. 2 is a combination of markers across 3 time points, which in practice would require 3 sequential blood samples from a child. To evaluate how well the prediction can be made with less sampling, ROFI was used to evaluate the models based on screening age, as well as the time point(s) for which samples would be drawn as constraints. Each model was optimized in the same fashion as for the full data set, and the final evaluation was based on a 50% feature-importance threshold capturing variability in the CV process through the 100 repetitions of the optimization process. This resulted in 14 distinct models, each with 100 estimates of the AUC, based on the screening age and time points at which blood samples would be available at that screening age. As seen in Fig. 3, if no samples are collected screening was based only on the demographics and yielded a low overall AUC. Adding in metabolomics, HLA and GRS increased the AUC at ages 3 and 6 months; however, there was a dramatic increase in accuracy adding the 9 months as both the sample collection and the primary screening age. This is due to the importance of IAAb as a feature measured at age 9 months in terms of prediction. In adding more blood draws, the increase remained highest when the 9-month sampling was included for the same reason regarding the strong predictive power of the IAAb data set. A Kruskal-Wallis test with a Tukey post hoc adjustment indicated that screening at age 9 months and including all 3 time points does not yield a statistically larger AUC than screening at age 9 months and including only metabolites from ages 3 and 9 months (P = .997). The model including only metabolomics from ages 3 and 9 months returns 13 features, of which the top 16 overlap completely for those specific time points with the addition of glycerol-alpha-phosphate. If only a single time point for a blood draw is selected, then the 9-month time point AUC is significantly larger than all the remaining sampling and screening age combinations. It includes 7 features, which completely overlaps with the top 16 in the full model excluding the 3-and 6-month metabolites.

Figure 3.

Box plot of the area under the curve (AUC) values from 100 repetitions of 5-fold cross-validation based on the number and timing of blood samples drawn from children, where each box represents the month(s) at which a blood draw is taken. Statistical comparisons between AUC values for each sampling/screening combination was performed with a Kruskal-Wallis test with a Tukey post hoc adjustment.

Fig. 4A gives representative global ROC curves associated with the largest AUC at no, 1, 2, or 3 sampling time points assuming a screening age of 9 months. Based on the prior Kruskal-Wallis analysis associated with Fig. 3, the ROC curves visually show the same pattern in which the 3- and 9-month sampling was not statistically different from all 3 time points. The single time point of age 9 months was significantly smaller overall; however, when focusing on a small range of predictions within a false-positive rate (FPR) of less than 0.05, the 3 were more similar, although the 2-sample model of ages 3 and 9 months is highest at the lowest FPRs (Fig. 4B). Specifically, if the percentage of false positives was set to 5% (~29 of the 582 negative participants), then the true-positive rate (TPR) dramatically increased from approximately 7.6% (~5 of the 73 T1D participants) with no sampling to approximately 38% (~28 of the 73 T1D participants) by adding the single time point of age 9 months and adding the second sampling increased the TPR by another 2%.

Figure 4.

Representative receiver operating characteristic curve based on data available at different sampling time point(s) when screening at age 9 months evaluating A, the entire range of false positive, and B, truncated range to 5% false positives. The dotted line represents no blood draws (ie, prediction is based solely on demographic data), the dash-dot line represents a single blood draw at age 9 months, the dashed line blood draws at ages 3 and 9 months, and the solid line includes data from blood draws at all 3 ages.

Assay Evaluation

An advantage of machine learning is the inclusion of the multivariate nature in which the various assay types work together in combination with infant measurement, demographic, and dietary information to make a prediction. The full evaluations in Figs. 2 to 4 assume that all 4 distinct assays (IAAb, HLA, GRS, and metabolomics) were included, which may not be necessary in practice. Selecting the 9-month sampling and screening time point as the best single evaluation point, each assay was evaluated individually in addition to the infant measurement, demographic, and dietary information and both the AUC and TPR at a set 5% FPR were used for evaluation (Fig. 5). The addition of the IAAb measurement yielded the largest increase both in the AUC and TPR. When evaluating 2 data sets, adding GRS to the IAAb data set yielded the largest AUC; however, adding the metabolomics to the IAAb data set was the largest gain for the TPR. With 3 data sets, adding the GRS or metabolomics to the IAAb with the HLA data had nearly the same AUC as each other and using all data sets. The maximum TPR with 3 data sets was with IAAb, HLA, and metabolomics. The AUC and TPR values for the 14 models were compared with a Kruskal-Wallis test with a Tukey post hoc test adjustment, which found that using only a 2-data set model of IAAb and GRS returned AUC values that were not statistically different from the full model. Based on the TPR, the best 2-data set model was IAAb and metabolomics, which again was not statistically different from the full model.

Figure 5.

Change in A, area under the curve (AUC), and B, true positive rate (TPR), as collections of assays are included in the model under the constraint that data is collected and evaluated at age 9 months. Statistical comparisons between AUC and TPR values for each data set combination was performed with a Kruskal-Wallis test with a Tukey post hoc adjustment.

Discussion

Here we demonstrated that an overall good prediction of T1D outcomes at age 6 years can be achieved by evaluating children for a small profile of features at age 9 months and integrating this information with their HLA typing, a T1D GRS, and IAAb status at age 9 months. Improvements in prediction can be achieved by adding an additional metabolite screening at age 3 months. It is important to recognize that features are selected as a multivariate model and, thus, univariate interpretations have limited utility as applied to complex biological networks. The machine learning framework allows assignment of individualized probabilities of progressing to T1D before age 6 years in children at a moderately high genetic risk at a 9-month screening, providing opportunities for monitoring, prevention of diabetic ketoacidosis at onset, or enrollment in immune intervention trials.

The top nonmetabolite features identified has having high feature importance (see Fig. 2) are established risk factors, such as IAAb, HLA, GRS, and gestational age (3, 6, 10, 23, 37). We further identified the 9-month age to be the optimal single screening and sampling time point; adding the 3-month as a secondary blood sampling time point increased the AUC from 0.77 to 0.81. We further evaluated the individual assays in the context of the 9-month screening age, noting IAAb as the most predictive data set, and adding either the GRS or metabolomics data to the IAAb gave the largest increase in accuracy, dependent on the metric of evaluation. Likely for clinical utility, a machine learning model would be tailored to a predefined TPR or FPR, and in the context of this study the most robust data types are IAAb and metabolomics, the combination of these two data types having both high AUC and TPR (see Fig. 5).

A limitation of this study is the size and imbalance of the cohort available for the machine learning: 73 children who progressed to T1D vs 582 who were not diagnosed with T1D at age 6 years. In addition, TEDDY is focused on high-risk children and as such the initial case-control design of the metabolomics study has a large proportion of children who are autoantibody positive by age 6 years (130), which likely makes these results specific to children of highest risk. Further, some features, such as gestational age, have a small change on the order of days, which may limit its utility in practice. A second limitation to this study is the use of global metabolomics measurements, and thus it is essential these findings be further validated through targeted metabolomic assay development and evaluation on independent studies, especially since metabolite measurements can change over time.

Our analysis of TEDDY plasma, untargeted metabolomic data, found that approximately 62% of the top metabolites in terms of feature importance, greater than 50, overlap with identified significant longitudinal metabolome profiles associated with the appearance of a first IAAb of multiple autoantibodies (38). It is also observed that several metabolites may be dietary based, such as piperidone and ascorbic acid (39). We note that the observed metabolites in the machine learning model suggest that children who develop T1D before age 6 years may have different sugar metabolite profiles as infants, as the selection of fructose, levoglucosan, glycerol-alpha-phosphate, and xylulose suggest that the carbohydrate metabolism is altered. In addition to altered carbohydrate metabolism, we observed the pentose phosphate pathway and purine degradation metabolism as potential targets of interest in the pathophysiology of T1D. Uridine and inosine are intermediates of the purine degradation pathway, including adenosine 5′-triphosphate, that has uric acid as the end product. Uric acid has been shown to be associated with insulin resistance and type 2 diabetes (40) but its association with T1D is not known. There are also metabolites of the pentose phosphate pathway, such as xylulose, observed in samples taken months preceding the onset of T1D. The alteration of carbohydrate levels, pentose phosphate pathway, and adenosine 5′-triphosphate could be due to a dysfunction of β cells and improper insulin secretion, as these processes are regulated by insulin (41, 42). In fact, β-cell dysfunction has been observed up to 5 years before the onset of T1D (43). Our results suggest alterations in sugar metabolism may start or develop early in infancy in individuals who develop T1D as younger children, but further focused studies are needed to evaluate these hypotheses.

Glossary

Abbreviations

AUC

area under the receiver operating characteristic curve

CV

cross-validation

FDR-T1D

first-degree relative with T1D

FPR

false-positive rate

GADA

glutamic acid decarboxylase antibody

GC-TOF MS

gas chromatography–time-of-flight mass spectrometry

GIA

general infant attributes

GRS

genetic risk scores

HLA

human leukocyte antigen

IA-2A

insulinoma-associated antigen 2 autoantibody

IAA

insulin autoantibody

IAAb

islet autoantibody

ROC

receiver operating characteristic curve

ROFI

Repeated Optimization for Feature Interpretation

SNV

single-nucleotide variation

T1D

type 1 diabetes

TEDDY

The Environmental Determinants of Diabetes in the Young

TPR

true-positive rate

Financial Support

This research was performed by an external analytic partner under the auspices of the TEDDY Study Group, which was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), National Institute of Allergy and Infectious Diseases (NIAID), Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD), National Institute of Environmental Health Sciences (NIEHS), Centers for Disease Control and Prevention (CDC), and JDRF (grant Nos. U01 DK63829, U01 DK63861, U01 DK63821, U01 DK63865, U01 DK63863, U01 DK63836, U01 DK63790, UC4 DK63829, UC4 DK63861, UC4 DK63821, UC4 DK63865, UC4 DK63863, UC4 DK63836, UC4 DK95300, UC4 DK100238, UC4 DK106955, UC4 DK112243, UC4 DK117483, U01 DK124166, and contract No. HHSN267200700014C). TEDDY is supported in part by the National Institutes of Health/National Center for Advancing Translational Sciences Clinical and Translational Science Awards to the University of Florida (grant No. UL1 TR000064) and the University of Colorado (grant No. UL1 TR002535).

Disclosures

The authors have nothing to disclose.

Data Availability

The data sets generated and analyzed during the present study will be made available in the NIDDK Central Repository at https://repository.niddk.nih.gov/studies/teddy.