Metabolic zones were developed to characterize heterogeneity of individuals with islet autoantibodies.
Baseline 2-h oral glucose tolerance test data from 6,620 TrialNet Pathway to Prevention Study (TNPTP) autoantibody-positive participants (relatives of individuals with type 1 diabetes) were used to form 25 zones from five area under the curve glucose (AUCGLU) rows and five area under the curve C-peptide (AUCPEP) columns. Zone phenotypes were developed from demographic, metabolic, autoantibody, HLA, and risk data.
As AUCGLU increased, changes of glucose and C-peptide response curves (from mean glucose and mean C-peptide values at 30, 60, 90, and 120 min) were similar within the five AUCPEP columns. Among the zones, 5-year risk for type 1 diabetes was highly correlated with islet antigen 2 antibody prevalence (r = 0.96, P < 0.001). Disease risk decreased markedly in the highest AUCGLU row as AUCPEP increased (0.88–0.41; P < 0.001 from lowest AUCPEP column to highest AUCPEP column). AUCGLU correlated appreciably less with Index60 (an indicator of insulin secretion) in the highest AUCPEP column (r = 0.33) than in other columns (r ≥ 0.78). AUCGLU was positively related to “fasting glucose × fasting insulin” and to “fasting glucose × fasting C-peptide” (indicators of insulin resistance) before and after adjustments for Index60 (P < 0.001).
Phenotypes of 25 zones formed from AUCGLU and AUCPEP were used to gain insights into type 1 diabetes heterogeneity. Zones were used to examine GCRC changes with increasing AUCGLU, associations between risk and autoantibody prevalence, the dependence of glucose as a predictor of risk according to C-peptide, and glucose heterogeneity from contributions of insulin secretion and insulin resistance.
Introduction
The use of a two-dimensional grid (2dgrid) with glucose and C-peptide as respective y and x axes has provided a new approach for the assessment of metabolic changes before and after the diagnosis of clinical type 1 diabetes (i.e., stage 3). Thus far, 2dgrids have been used to graphically study changes in metabolic function pertaining to the natural history of type 1 diabetes (1,2), the efficacy of experimental treatments designed to delay the onset of type 1 diabetes (3,4), and heterogeneity of clinical features at diagnosis (5).
In this study, we explored the utility of 2dgrids for defining phenotypes from baseline data of autoantibody-positive individuals (all relatives of people with type 1 diabetes) who were screened with 2-h oral glucose tolerance tests (OGTTs) in the TrialNet Pathway to Prevention (TNPTP) study. We demarcated 25 zones on a 2dgrid, defined by area under the curve glucose (AUCGLU) and area under the curve C-peptide (AUCPEP), with the purpose of analyzing demographic, metabolic, immunologic, and genetic heterogeneity in those individuals.
We examined the capacity of zones to characterize the metabolic development of type 1 diabetes, determined the ability of zones to predict type 1 diabetes, identified correlations between risk for type 1 diabetes and phenotypic attributes, and gauged the impact of insulin resistance upon glycemia. Such information could increase precision for selection of target populations in type 1 diabetes prevention trials.
Research Design and Methods
Participants
Data from TNPTP participants were analyzed. All were relatives of individuals with type 1 diabetes (ages 1–50 years) who screened positive for type 1 diabetes-associated autoantibodies. There were 7,160 participants who had at least one positive autoantibody from testing for microinsulin (mIAA), insulinoma-associated antigen 2 (IA-2A), and glutamic acid decarboxylase (GADA) autoantibodies, after which testing was performed secondarily for the presence of islet cell cytoplasmic autoantibodies. The presence of zinc transporter 8 autoantibodies (ZnT8A) was also tested secondarily, after it was introduced into the TNPTP several years following the study’s initiation. All individuals analyzed had 2-h oral glucose tolerance tests (OGTTs) (1.75 g/kg to a maximum of 75 g), of whom 6,620 were in the nondiabetic range at baseline with complete OGTT and autoantibody data. Of that number, 101 values were missing for age and 718 for BMI (mostly due to absence of a height measurement). Also, 147 Index60 values >3.0 or <−3.0, which corresponded to approximately ±2.5 SD, were not included in the analyses. The study was approved by institutional review boards at all sites, and all participants provided informed consent.
Procedures
Participants were followed with OGTTs at 6-month or yearly intervals based upon their degree of risk for type 1 diabetes. Measurements of autoantibodies, BMI, and HbA1c were obtained together with the OGTTs. Index60 (6) was derived from OGTTs using proportional hazards regression. Its calculation is shown by the following equation: Index60 = 0.0165 × (60-min glucose) − 0.3644 × (60-min C-peptide) + 0.36945 × ln(fasting C-peptide). The units are mg/dL for glucose and ng/mL for C-peptide.
If participants initially screened positive for one autoantibody, a confirmatory measurement was required to qualify for an OGTT. Those who initially screened positive for two autoantibodies were not required to have confirmatory measurements to qualify for the study. HLA genotyping was conducted for all subjects. If an OGTT met American Diabetes Association criteria for a diagnosis (7) of type 1 diabetes (fasting glucose ≥126 mg/dL and/or 2-h glucose ≥200 mg/dL), a confirmatory OGTT was repeated, unless there was additional clinical evidence clearly indicative of diabetes. BMI percentiles of those <18 years were based on sex- and age-specific Centers for Disease Control and Prevention growth charts. BMI percentiles of those ≥18 years were based on sex-specific BMI distributions from the 2015–2016 National Health and Nutrition Examination Survey (NHANES). A BMI value ≥85th percentile for age was considered overweight (including those with obesity). Glucose was measured by glucose oxidase. Insulin was measured initially by radioimmunoassay, which was then transitioned to the Tosoh assay. A previously reported conversion was used (8). C-peptide was determined by the Tosoh assay. Assays for mIAA, GADA, IA-2A, and ZnT8A were performed at the Barbara Davis Center, the TrialNet reference laboratory (9,10). Islet cell cytoplasmic autoantibody assays (11,12) were performed at the University of Florida.
Analysis
Zones Defined by AUCGLU and AUCPEP Values
Data from baseline OGTTs were analyzed based upon 25 zones on a 2dgrid (Supplementary Fig. 1) from five AUCPEP columns (labeled A–E in order of increasing values) and from five AUCGLU rows (labeled 1–5 in order of increasing values). For example, the zone with the lowest AUCPEP and highest AUCGLU values was labeled A5, while the zone with the highest AUCPEP and lowest AUCGLU values was labeled E1. The AUCGLU values used to divide the five rows and the AUCPEP values used to divide the five columns, the basis for the 25 zones, were selected to provide enough data for analyses within zones (range of individuals within zones, 138–487) and among zones. The values used to divide AUCGLU rows were <115 mg/dL, 115 to <130 mg/dL, 130 to <145 mg/dL, 145 to <160 mg/dL, and ≥160 mg/dL; values used to divide AUCPEP columns were <3.5 ng/mL, 3.5 to <5.0 ng/mL, 5.0 to <6.5 ng/mL, 6.5 to <8.0 ng/mL, and ≥8.0 ng/mL. Each zone was characterized by demographic, autoantibody, genetic, and metabolic measures pertinent to type 1 diabetes, along with the 5-year risk of type 1 diabetes (Supplementary Table 1).
Glucose and C-peptide Response Curves
Two-dimensional glucose and C-peptide response curves (GCRCs) have been used in prior studies (1–5). They are constructed from mean glucose and mean C-peptide OGTT values at 30-, 60-, 90-, and 120-min time points. GCRCs provide a visual representation of an OGTT. After a virtual line was drawn between the 30-min and 120-min time points, a polygon was formed. A central point (centroid) for the polygon can be calculated with a standard formula (4). Slopes of the sides of GCRC shapes (based on a horizontal axis) were used to quantify differences between GCRC shapes.
Glucose and C-peptide were studied in two dimensions using a grid based on AUCGLU values on the y axis and AUCPEP values on the x axis to form the 25 zones. Glucose and C-peptide were also studied in two dimensions within the zones, with glucose values on the y axis and C-peptide values on the x axis from OGTT time points forming GCRCs. Thus, whereas the AUCGLU and AUCPEP values were used to define the zones, glucose and C-peptide values at OGTT time points were used to form a GCRC within each zone.
Insulin Resistance
There are no validated measures of insulin resistance in autoantibody-positive populations. We chose an approach that utilized the product term fasting glucose × fasting insulin (glu*ins), which is the same metabolic term as that used to calculate HOMA of insulin resistance (HOMA-IR) (13) and therefore correlates highly with the latter. The actual HOMA-IR formula was not used to avoid any suggestion that our analysis formally validates the use of HOMA-IR in autoantibody-positive populations; the analysis was intended to be exploratory. We also used another product term, fasting glucose × fasting C-peptide (glu∗pep). Adjustments for Index60, a measure reflecting β-cell function, and age were made with covariance analyses.
Statistical Assessments
To assess the consistency of the zone phenotypes, the cohort was randomly, equally, and fully divided into two samples for certain analyses; they are referred to as split samples (split sample 1 and split sample 2). The number of individuals within the 25 zones of samples 1 and 2 ranged from 63 to 265 and 56 to 226, respectively. t tests and χ2 tests were used for comparisons of variables between the lowest and highest AUCGLU row in each AUCPEP column and between the lowest and highest AUCPEP column in each AUCGLU row. Cumulative incidence was described with Kaplan-Meier curves. Cox proportional hazards regression was used to assess disease prediction. A two-tailed P value of <0.05 was considered statistically significant. Bonferroni adjustments were made for multiple comparisons as appropriate. SAS version 9.4 was used.
Results
Demographic Heterogeneity
Individuals in the five zones within column A (lowest AUCPEP) were substantially younger and less frequently overweight (P < 0.001) than those in the five zones within column E (highest AUCPEP). The mean ± SD age ranged from 6.3 ± 4.4 to 9.1 ± 9.2 years among zones of column A in comparison with 21.3 ± 11.3 to 26.9 ± 14.1 years among zones of column E. Those with overweight ranged from 7% to 16% among zones of column A compared with 35% to 38% among zones of column E. Values of baseline demographic variables are shown in Supplementary Table 2 according to split sample 1 and split sample 2. There were no significant differences between the samples except for Index60 and AUCGLU. However, the differences in the values were small. The correlation coefficients of the measures between the zones of samples 1 and 2 were r ≥ 0.94, except for GADA (r = 0.57) and DR3/DR4 (r = 0.65).
Comparisons for age and percent overweight were statistically significant at each row between the zones of column A and column E (P < 0.001). Zones of column A had appreciably higher percentages of male individuals than zones of column E for rows 1–4 (all P < 0.01) and, to a lesser extent, for row 5 (P = 0.045). The differences between the glucose rows were much smaller without distinct patterns.
Metabolic Heterogeneity
We constructed a GCRC from mean glucose and mean C-peptide values at the 30-, 60-, 90-, and 120-min time points for each of the 25 zones. Figure 1 shows GCRCs and their centroids (i.e., central points) from the 25 zones. GCRC shapes changed incrementally from rows 1 to 5. Since the change was most apparent when row 5 was compared with row 1 in column E, we color-coded the GCRCs for E1 and E5: blue for 30–60 min, green for 60–90 min, and gold for 90–120 min. The arrows show the directions of change from the preceding time point to the next. There were significant differences in the slopes from 30 to 60 min and from 60 to 90 min between row 1 and row 5 in column E for each comparison (P < 0.001 for both slopes). Corresponding to these differences in slopes, E1 and E5 differed in glucose changes from 30 to 60 min (decreasing in E1 and increasing in E5) and differed in C-peptide changes from 60 to 90 min (decreasing in E1 and increasing in E5). The differences between E1 and E5 suggest pathologic progression, with the persistent increase of C-peptide in E5 perhaps representing a compensatory response to the increasing glucose from deficient earlier insulin secretion.
GCRCs, which are phenotypic expressions of mean glucose and mean C-peptide values at 30-, 60-, 90-, and 120-min OGTT time points, are shown for each of the 25 zones apportioned from the full cohort. The arrows represent the combined glucose and C-peptide changes from the preceding time point to the next time point. The GCRC centroids (their central points) are also shown. There are incremental changes in GCRC shape as AUCGLU increases, resulting in marked differences from the lowest to highest AUCGLU. The changes are similar within each column but are most evident in column E, for which the GCRC sides in E1 and E5 are color-coded. The slopes from 30 to 60 min and 60 to 90 min are indicative of the incremental changes, with significant differences in the 30- to 60-min and 60- to 90-min slopes between E1 and E5 (P < 0.001 for both slopes). Note how the angle at the 60-min time point changes from being acute and facing leftward in E1 to being obtuse and facing downward in E5. This represents differences between E1 and E5: from 30 to 60 min, glucose decreases in E1 and increases in E5; from 60 to 90 min, C-peptide decreases in E1 and increases in E5. The increasing C-peptide in E5 suggests an attempt to compensate for increasing glucose.
GCRCs, which are phenotypic expressions of mean glucose and mean C-peptide values at 30-, 60-, 90-, and 120-min OGTT time points, are shown for each of the 25 zones apportioned from the full cohort. The arrows represent the combined glucose and C-peptide changes from the preceding time point to the next time point. The GCRC centroids (their central points) are also shown. There are incremental changes in GCRC shape as AUCGLU increases, resulting in marked differences from the lowest to highest AUCGLU. The changes are similar within each column but are most evident in column E, for which the GCRC sides in E1 and E5 are color-coded. The slopes from 30 to 60 min and 60 to 90 min are indicative of the incremental changes, with significant differences in the 30- to 60-min and 60- to 90-min slopes between E1 and E5 (P < 0.001 for both slopes). Note how the angle at the 60-min time point changes from being acute and facing leftward in E1 to being obtuse and facing downward in E5. This represents differences between E1 and E5: from 30 to 60 min, glucose decreases in E1 and increases in E5; from 60 to 90 min, C-peptide decreases in E1 and increases in E5. The increasing C-peptide in E5 suggests an attempt to compensate for increasing glucose.
The consistency of the GCRCs between the zones of sample 1 and sample 2 is shown in Supplementary Fig. 2.
Islet Autoantibody, DR3/DR4, and Risk Heterogeneity
Figure 2 shows marked heterogeneity in the prevalence of mIAA, IA-2A, and DR3/DR4 and the 5-year risk among zones. In addition to the frequencies shown in the figure, heat maps of the zones are shown for 5-year risk, IA-2A, mIAA, and DR3/DR4. The heat maps provide depictions of overall frequency distributions of those measures. Frequencies tended to be highest (most intense heat map) in the area of A5 and lowest (least intense) in the area of E5. A close inspection shows an especially marked similarity between 5-year risk and IA-2A. This corresponds to the very high correlation between the 5-year risk and IA-2A among the 25 zones (r = 0.96). Correlations of 5-year risk with other measures also tended to be high but not to the same extent (mIAA, r = 0.74; GADA, r = 0.74; DR3/DR4, r = 0.89; all P < 0.001). Supplementary Fig. 3 is a scatterplot that shows the strength of correlation between risk and IA-2A prevalence and its consistency between sample 1 and sample 2. In Cox regression analyses, there were significant associations of type 1 diabetes with each autoantibody and with DR3/DR4 (P < 0.001 for all); the strongest association was again with IA-2A (χ2 values with age as a covariate were IA-2A = 535, mIAA = 111, GADA = 41, and DR3/DR4 = 67). Autoantibody titer was not included in the analysis.
The 5-year risk of type 1 diabetes and the prevalence of mIAA, IA-2A, and DR3/DR4 are shown for each zone. The number of subjects per zone is indicated in parentheses. Also shown are heat maps that provide an overall view of the 5-year risk and the prevalence of the other measures in the zones. Note that the 5-year risk and the prevalence of the other measures tend to be highest in the area of A5 and lowest in the area of E1. Note also that the heat maps for 5-year risk and IA-2A are very much alike. T1D, type 1 diabetes; yr, year.
The 5-year risk of type 1 diabetes and the prevalence of mIAA, IA-2A, and DR3/DR4 are shown for each zone. The number of subjects per zone is indicated in parentheses. Also shown are heat maps that provide an overall view of the 5-year risk and the prevalence of the other measures in the zones. Note that the 5-year risk and the prevalence of the other measures tend to be highest in the area of A5 and lowest in the area of E1. Note also that the heat maps for 5-year risk and IA-2A are very much alike. T1D, type 1 diabetes; yr, year.
In Supplementary Table 3, heterogeneity is evident in both the AUCGLU and AUCPEP dimensions for all the measures. The 5-year risk, each autoantibody measure, and DR3/DR4 were significantly greater in row 5 than in row 1 (P < 0.001 for all), and except for GADA, each measure was significantly greater in column A than in column E (P < 0.001). The differences remained significant with Bonferroni adjustments for multiple comparisons (P = 0.008 was the threshold for significance).
Figure 2 underscores the importance of C-peptide’s impact on risk prediction. For example, despite lower glucose levels (151.4 ± 4.4 vs. 175.1 ± 13.4 mg/dL), those in A4 were at greater risk than those in E5 for type 1 diabetes (0.64 vs. 0.41, P < 0.001).
The importance of C-peptide for risk prediction is also evident in the cumulative incidence curves for diabetes that were derived from the five zones in row 5, the highest AUCGLU category (Fig. 3). Despite zones having similarly high glucose levels, curves declined at each AUCPEP increment, resulting in a marked overall decline (P < 0.001) of 5-year risk from A5 (0.88) to E5 (0.41). There was less risk at higher C-peptide levels regardless of the glucose level.
Cumulative incidence curves for the five zones in the highest AUCGLU row are shown. The curves decrease as AUCPEP progressively increases from zone A5 to zone E5, resulting in a large difference between them. This shows the importance of C-peptide as a modifier of type 1 diabetes risk predicted by glucose. T1D, type 1 diabetes.
Cumulative incidence curves for the five zones in the highest AUCGLU row are shown. The curves decrease as AUCPEP progressively increases from zone A5 to zone E5, resulting in a large difference between them. This shows the importance of C-peptide as a modifier of type 1 diabetes risk predicted by glucose. T1D, type 1 diabetes.
Glucose Heterogeneity: Combined Impact of β-Cell Deficiency and Insulin Resistance
Since insulin resistance and type 2 diabetes are so prevalent in the general population, we assessed whether, in addition to a deficiency in insulin secretion, insulin resistance could be a contributor to glucose levels in autoantibody-positive populations. The analysis was focused on those in column E, since they were the most likely to be the most insulin resistant; they were the oldest individuals with the highest BMI percentile.
Two findings suggested that β-cell loss was less responsible for higher glucose levels in column E than in the other columns. In one finding, fasting C-peptide decreased from the lowest to the highest AUCGLU row in columns A–D (P < 0.001 for columns A–C and P < 0.034 for column D), whereas fasting C-peptide increased from the lowest to highest AUCGLU row in column E (2.65 ± 0.85 ng/mL in zone E1 to 3.08 ± 1.25 ng/mL in zone E5 [P < 0.001]). In the other finding, the correlation between AUCGLU and Index60 (as a measure of β-cell secretion) in column E was much smaller than the correlations in the other columns (r = 0.33 vs. r ≥ 0.78). Based on these findings, we assessed whether insulin resistance could have contributed to glycemia in column E. For this purpose, linear regression analyses were performed, with AUCGLU as the dependent variable in the models and with glu*ins or glu*pep (see Research Design and Methods) as the independent variable of interest, along with Index60 and age as covariates that were contingent upon the model.
Table 1 shows linear regression indices for columns A and E, along with t values as an indicator of degree of association. Within column E, there was a positive association of AUCGLU with glu*ins (P < 0.001) that became stronger when either Index60 or both Index60 and age were added to models. We also examined similar models in the table with glu*pep replacing glu*ins. Associations of AUCGLU with glu*pep in the models again were positive (P < 0.001) and somewhat stronger. In contrast to column E, the associations of AUCGLU with glu*ins in column A were nonsignificant in all models, and the associations of AUCGLU with glu*pep in column A were inverse. Patterns of overall associations in the full cohort of AUCGLU with glu*ins or with glu*pep were like those in column E (P < 0.001). The associations of AUCGLU with glu*ins and with glu*pep adjusted for Index60 were still considerable with adjustments for age.
Linear regression values for AUCGLU
Variables . | Column A (n = 1,127) . | Column E (n = 1,330) . | All (n = 6,510) . | ||||||
---|---|---|---|---|---|---|---|---|---|
Slope ± SE . | t value . | P value . | Slope ± SE . | t value . | P value . | Slope ± SE . | t value . | P value . | |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.4 ± 0.5 | 12.9 | <0.001 | 6.5 ± 0.2 | 27.7 | <0.001 |
glu*ins | −0.001 ± 0.000 | −1.6 | 0.106 | 0.000 ± 0.000 | 3. | <0.001 | 0.001 ± 0.000 | 8.4 | <0.001 |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.8 ± 0.5 | 13.8 | <0.001 | 8.5 ± 0.2 | 34.5 | <0.001 |
glu*ins | −0.000 ± 0.000 | −1.5 | 0.145 | 0.001 ± 0.000 | 11.9 | <0.001 | 0.001 ± 0.000 | 21.6 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.2 | <0.001 | 6.9 ± 0.5 | 13.9 | <0.001 | 9.2 ± 0.3 | 35.3 | <0.001 |
glu*ins | −0.000 ± 0.000 | −1.4 | 0.174 | 0.000 ± 0.000 | 6.2 | <0.001 | 0.001 ± 0.000 | 21.8 | <0.001 |
Age | −0.09 ± 0.06 | −1.6 | 0.103 | 0.19 ± 0.05 | 4.1 | <0.001 | 0.18 ± 0.02 | 8.0 | <0.001 |
— | — | — | — | — | — | — | — | — | — |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.4 ± 0.5 | 12.9 | <0.001 | 6.5 ± 0.2 | 27.7 | <0.001 |
glu*pep | −0.07 ± 0.02 | −3.3 | 0.003 | 0.04 ± 0.01 | 7.6 | <0.001 | 0.003 ± 0.003 | 11.0 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.7 | <0.001 | 7.6 ± 0.5 | 15.7 | <0.001 | 11.5 ± 0.3 | 44.5 | <0.001 |
glu*pep | −0.04 ± 0.01 | −3.1 | <0.002 | 0.06 ± 0.01 | 11.4 | <0.001 | 0.12 ± 0.003 | 35.3 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.3 | <0.001 | 7.6 ± 0.5 | 13.9 | <0.001 | 11.4 ± 0.3 | 43.2 | <0.001 |
glu*pep | −0.03 ± 0.01 | −2.7 | 0.008 | 0.06 ± 0.01 | 10.9 | <0.001 | 0.12 ± 0.003 | 33.8 | <0.001 |
Age | −0.05 ± 0.06 | −0.8 | 0.402 | 0.06 ± 0.05 | 1.4 | <0.173 | −0.05 ± 0.02 | −2.15 | 0.032 |
Variables . | Column A (n = 1,127) . | Column E (n = 1,330) . | All (n = 6,510) . | ||||||
---|---|---|---|---|---|---|---|---|---|
Slope ± SE . | t value . | P value . | Slope ± SE . | t value . | P value . | Slope ± SE . | t value . | P value . | |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.4 ± 0.5 | 12.9 | <0.001 | 6.5 ± 0.2 | 27.7 | <0.001 |
glu*ins | −0.001 ± 0.000 | −1.6 | 0.106 | 0.000 ± 0.000 | 3. | <0.001 | 0.001 ± 0.000 | 8.4 | <0.001 |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.8 ± 0.5 | 13.8 | <0.001 | 8.5 ± 0.2 | 34.5 | <0.001 |
glu*ins | −0.000 ± 0.000 | −1.5 | 0.145 | 0.001 ± 0.000 | 11.9 | <0.001 | 0.001 ± 0.000 | 21.6 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.2 | <0.001 | 6.9 ± 0.5 | 13.9 | <0.001 | 9.2 ± 0.3 | 35.3 | <0.001 |
glu*ins | −0.000 ± 0.000 | −1.4 | 0.174 | 0.000 ± 0.000 | 6.2 | <0.001 | 0.001 ± 0.000 | 21.8 | <0.001 |
Age | −0.09 ± 0.06 | −1.6 | 0.103 | 0.19 ± 0.05 | 4.1 | <0.001 | 0.18 ± 0.02 | 8.0 | <0.001 |
— | — | — | — | — | — | — | — | — | — |
Index60 | 33.6 ± 0.6 | 52.7 | <0.001 | 6.4 ± 0.5 | 12.9 | <0.001 | 6.5 ± 0.2 | 27.7 | <0.001 |
glu*pep | −0.07 ± 0.02 | −3.3 | 0.003 | 0.04 ± 0.01 | 7.6 | <0.001 | 0.003 ± 0.003 | 11.0 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.7 | <0.001 | 7.6 ± 0.5 | 15.7 | <0.001 | 11.5 ± 0.3 | 44.5 | <0.001 |
glu*pep | −0.04 ± 0.01 | −3.1 | <0.002 | 0.06 ± 0.01 | 11.4 | <0.001 | 0.12 ± 0.003 | 35.3 | <0.001 |
Index60 | 33.5 ± 0.6 | 52.3 | <0.001 | 7.6 ± 0.5 | 13.9 | <0.001 | 11.4 ± 0.3 | 43.2 | <0.001 |
glu*pep | −0.03 ± 0.01 | −2.7 | 0.008 | 0.06 ± 0.01 | 10.9 | <0.001 | 0.12 ± 0.003 | 33.8 | <0.001 |
Age | −0.05 ± 0.06 | −0.8 | 0.402 | 0.06 ± 0.05 | 1.4 | <0.173 | −0.05 ± 0.02 | −2.15 | 0.032 |
Variables of interest are glu*ins above the dashed line and glu*pep below the dashed line. Black lines separate models, and white lines separate variables within a model; glu, ins, and pep represent fasting values of glucose, insulin, and C-peptide, respectively.
The models in Table 1 indicated that the adjustment for Index60 substantially increased the associations of AUCGLU with glu*ins and glu*pep in the full cohort, whereas the adjustment for age had little impact on the associations. Thus, another analysis was performed to further explore the impact of the adjustments for Index60 and age in the full cohort. In a regression model (Supplementary Table 4) with AUCGLU as the dependent variable and AUCPEP <8.0 ng/mL (i.e., columns A–D) vs. ≥8.0 ng/mL (i.e., column E) as the independent variable, AUCGLU was significantly greater for those in column E than those in columns A–D (mean difference 10.6 ± 0.8 mg/dL). When Index60 was added as a covariate to the model, the difference in AUCGLU increased substantially (36.6 ± 0.7 mg/dL), whereas adding age as a covariate to the model had little effect on the difference (12.0 ± 0.8 mg/dL). Although the P values were all <0.001 in the table, they were appreciably more significant for the difference in the model with Index60 as a covariate than with age as a covariate. The greater impact of Index60 on the difference in AUCGLU was evident from its much higher t values.
Conclusions
The methods of analysis used for this study were unique in that they focused on groups of autoantibody-positive relatives of individuals with type 1 diabetes, defined by metabolic zones for which mean values and prevalence values of characteristics were obtained to form phenotypes. We chose to use zones demarcated by AUCGLU and AUCPEP for characterizing phenotypes based on the following considerations: 1) metabolic indices have been useful for demonstrating heterogeneity of autoantibody-positive individuals (5,14–16); 2) AUCGLU and AUCPEP represent the two key metabolic components of type 1 diabetes pathophysiology; and 3) AUCGLU and AUCPEP together on a grid provide a two-dimensional approach for assessing diversity among groups.
Five main findings, each novel, resulted from analyzing zones defined by AUCGLU and AUCPEP on the 2dgrid: 1) the two-dimensional pattern of heterogeneity according to zones defined by AUCGLU and AUCPEP; 2) the visual and quantitative changes (i.e., slopes) in GCRC shapes from lower to higher glucose levels, and the relation of those changes to the decline of β-cell function; 3) the tight correlation of type 1 diabetes risk with IA-2A prevalence among groups; 4) the demonstration that C-peptide is a critical factor for predicting risk when glucose levels are high; and 5) evidence suggesting that insulin resistance can appreciably increase glucose levels in autoantibody-positive individuals, especially at high AUCPEP.
As AUCGLU increased, there were marked changes in GCRC shape (Fig. 2), the result of changes in glucose and C-peptide relative to each other. From an overall perspective, with increasing glucose, the changes in the orientation of the three sides open the shape of the GCRC from E1 to E5. Evidence indicates that when glucose levels eventually become diagnostic, the GCRC shape will be almost linear from increased glucose at the 30- to 60-min and the 60- to 90-min intervals, with the 90- to 120-min interval being the last to become positive. The line also becomes more vertical as insulin responsiveness decreases (5). These evolving shapes of GCRCs are phenotypic expressions of deteriorating responsiveness to the OGTT glucose challenge.
The analyses were cross-sectional. Thus, although the changes of GCRC shape with increasing glucose might suggest temporal changes during the progression to type 1 diabetes, this can only be definitively determined from longitudinal analyses. However, the granular analysis of the large number of individuals undertaken in this study would be difficult to perform longitudinally.
No studies have examined autoantibody or DR3/DR4 prevalence according to glucose and C-peptide using a two-dimensional approach. The 2dgrid revealed an increase of autoantibody positivity prevalence and DR3/DR4 prevalence in two directions, toward higher glucose and toward lower C-peptide. The two-dimensional pattern is such that the autoantibodies and DR3/DR4 are most prominent in the area of A5 and least prominent in the area of E1. This diagonal pattern of increasing pathology on the 2dgrid from the lower right corner to the upper left corner is indicative of the two-dimensionality of deteriorating β-cell function and thus the need to consider both glucose and C-peptide in assessments of the pathogenesis and risk of type 1 diabetes.
It is known that IA-2A is a strong predictor of type 1 diabetes (17) and that its titer increases during the progression to type 1 diabetes (18). However, the strength of correlation of risk with IA-2A and its consistency between sample 1 and sample 2 suggest that there is an even tighter linkage between β-cell decline and IA-2A than previously recognized. Correlation analyses of type 1 diabetes risk with IA-2A prevalence have not been reported, since they can only be performed among groups and not individuals. The Cox regression analysis, showing that IA-2A was the strongest predictor of type 1 diabetes, was consistent with the correlation analyses. Our findings regarding GADA are similar to previous findings that show progression to type 1 diabetes is slower among older individuals with GADA (16).
In addition to the discussion points above, other findings further emphasize the importance of assessing associations of type 1 diabetes risk with glucose in the context of insulin or C-peptide. Specifically, we found 1) individuals with lower glucose and lower C-peptide can have substantially more risk than individuals with higher glucose and higher C-peptide, and 2) individuals with higher glucose and lower C-peptide values can be at greater risk than individuals with similar glucose but higher C-peptide. These findings have both clinical and research implications.
Our findings suggested that insulin resistance contributed to glucose levels in column E. Of note was the much weaker correlation between AUCGLU and Index60 in column E than in the other columns. This suggests that at least some of the unexplained variance of AUCGLU in that column is attributable to insulin resistance. In Table 1, the positive association of AUCGLU with glu*ins and with glu*pep in column E, and the lack of a positive association in column A, provided more evidence that insulin resistance contributed to glucose levels. Interestingly, the positive association tended to be stronger when glu*pep replaced glu*ins in the models. The basis for this is unclear.
The addition of Index60 to the models in Table 1 appreciably enhanced the positive associations of AUCGLU with both glu*ins and glu*pep in the full cohort, whereas age appeared to have little impact. Another analysis of the full cohort (Supplementary Table 4) indicated that the addition of Index60 to regression models sizably increased AUCGLU values in column E relative to other columns; however, age again had little influence. The impact of Index60 in the regression models could be explained by an adjustment for insulin secretion when Index60 is introduced into a model. This would tend to remove the influence of insulin secretion on AUCGLU and thus provide a clearer picture of the effect of insulin resistance on AUCGLU.
The findings indicate that individuals with autoantibodies who have AUCPEP values ≥8.0 ng/mL differ phenotypically from others with autoantibodies. Although this statistical differentiation does not in itself indicate that those individuals with high C-peptide represent an endotype defined by genetics or mechanisms (19), the AUCPEP threshold of ≥8.0 ng/mL could be used to indicate the likelihood of appreciable insulin resistance and the possibility that it is a meaningful determinant of glucose.
The novel analytic approach used in this study provided several unique insights regarding heterogeneity. Another strength was the large TNPTP cohort of autoantibody-positive individuals with comprehensive information. Limitations included the absence of more detailed genetic information beyond the DR3/DR4 genotype characterization. Values of the variables studied were consistent between the two samples except for DR3/DR4 and GADA. However, there was not an external validation. It should be emphasized that the heterogeneity studied in the analyses was predicated upon a metabolic scaffolding of AUCGLU and AUCPEP.
A contribution of insulin resistance to glycemia is potentially an important consideration in selecting target populations for prevention trials. The inclusion of individuals whose insulin resistance contributes appreciably to glycemia could cloud determinations of efficacy of mechanistically based treatments. If not excluded, those individuals might be considered a prespecified subgroup in the analyses.
We have introduced the novel concept of characterizing autoantibody-positive populations according to zones, which are defined on a 2dgrid by AUCGLU and AUCPEP. The findings indicate that these zones enhance the study of heterogeneity in autoantibody-positive populations and, in doing so, provide new insights into the development of the disorder. Importantly, it appears that zones can improve the precision of subject selection for clinical trials that assess potential preventive treatments of type 1 diabetes and other studies of the disorder. Thus, the use of zones could offer a valuable platform for advancing type 1 diabetes research.
This article contains supplementary material online at https://doi.org/10.2337/figshare.22190029.
A complete list of members of the Type 1 Diabetes TrialNet Study Group and TrialNet Affiliate Sites is available in the supplementary material.
Article Information
Acknowledgments. The authors acknowledge the support of the Type 1 Diabetes TrialNet Study Group, which identified study participants and provided samples and follow-up data for this study.
Funding. The Type 1 Diabetes TrialNet Study Group is a clinical trials network funded by the National Institutes of Health (NIH) through the National Institute of Diabetes and Digestive and Kidney Diseases, the National Institute of Allergy and Infectious Diseases, and the Eunice Kennedy Shriver National Institute of Child Health and Human Development, through cooperative agreements U01 DK061010, U01 DK061034, U01 DK061042, U01 DK061058, U01 DK085453, U01 DK085461, U01 DK085465, U01 DK085466, U01 DK085476, U01 DK085499, U01 DK085504, U01 DK085509, U01 DK103180, U01 DK103153, U01 DK103266, U01 DK103282, U01 DK106984, U01 DK106994, U01 DK107013, U01 DK107014, UC4 DK106993, UC4 DK11700901, and U01 DK 106693-02, and by JDRF. This work was also made possible with support from grants KL2TR002530 (A. Carroll, principal investigator) and UL1TR002529 (A. Shekhar, principal investigator) from the NIH, Clinical and Translational Sciences Award, National Center for Advancing Translational Sciences, and NIH grants R01 DK121843 (to M.J.R.) and R01 DK124395 (to M.J.R.).
The contents of this article are solely the responsibility of the authors and do not necessarily represent the official views of the NIH or JDRF.
Duality of Interest. No potential conflicts of interest relevant to this article were reported.
Author Contributions. J.M.S. conceptualized the study, analyzed and interpreted data, and wrote the manuscript. D.C. provided statistical expertise, performed analytic procedures, and interpreted data. E.K.S., H.M.I., B.M.N., L.M.J., M.A.K., C.E.-M., K.C.H., and J.S.S. contributed to data interpretation and manuscript refinement. M.J.R. contributed to analysis, data interpretation, and manuscript refinement. J.S. and D.C. are the guarantors of this work and, as such, had full access to all study data and take responsibility for the integrity of the data and the accuracy of the data analysis.