Stratification of Type 1 Diabetes Risk on the Basis of Islet Autoantibody Characteristics

Sera of all first-degree relatives of patients with type 1 diabetes from the Bart’s Oxford (BOX) and the Munich family studies (13,17) were tested for IAA, GADA, and IA-2A. A total of 180 nondiabetic relatives (76 from BOX study and 104 from Munich family study) were selected on the basis of positivity for at least one of these antibodies on two or more occasions and whether a sufficient volume of the first positive sample was available for complete testing according to the study protocol. The study cohort consisted of 65 offspring (38 sons, 27 daughters), 68 siblings (32 brothers, 36 sisters), and 47 parents (21 fathers, 26 mothers) of the diabetic proband. The first autoantibody-positive serum sample from these relatives was used for all measurements performed in this study. The median age of relatives at the time of collection of the first available antibody-positive sample was 14.5 years (interquartile range 8.3–30.3 years). Subjects were prospectively monitored for the development of type 1 diabetes over a median follow-up period of 5.9 years (interquartile range 3.8–10.7 years) for a total of 1,248 subject years. Of the 180 autoantibody-positive relatives, 59 developed type 1 diabetes during follow-up (median time to diabetes 3.6 years; interquartile range 1.3–6.1 year). Diabetes was diagnosed using World Health Organization/American Diabetes Association criteria (18). Another 78 relatives with confirmed positive antibodies were not included due to insufficient serum sample. Twenty of these developed diabetes. An additional 136 relatives had islet autoantibodies in one sample but were either negative on follow-up samples (n = 70; none developed diabetes) or had no follow-up sample (n = 66; 13 developed diabetes) and were not included in the current study. Seven of 5,782 relatives who were islet autoantibody negative at screening developed diabetes during follow-up. The respective local ethical committees approved BOX and Munich family studies.

IAA, GADA, and IA-2A were measured by protein A/G radiobinding assays as previously described (7,19) using ¹²⁵I-labeled insulin and [³⁵S]methionine-labeled in vitro-translated recombinant human GAD65 and IA-2, respectively. Samples with antibody titers above the discriminatory range of the assays were titrated until they fell within this part of the standard curve and the units multiplied by the appropriate dilution factor. The thresholds for positivity in each assay corresponded to the 99th percentile of control subjects. These assays had sensitivities and specificities of 80 and 94% (GADA), 58 and 100% (IA-2A), and 30 and 98% (IAA) in the First DASP Assay Proficiency Evaluation (20).

GADA and IA-2A epitope specificity was determined by radiobinding assay of GAD65/67 chimeric proteins and IA-2/IA-2β fragments or chimeric proteins as previously described (21,22). Thresholds for positivity were defined as the upper limit of 50 control subject sera. GAD antibody epitope specificities were classified as GAD65-NH₂-terminal (residues 1–100), GAD65-MID (residues 235–442), GAD65-COOH-terminal (residues 436–585), and/or GAD67. IA-2/IA-2β antibody epitopes were classified as IA-2-JM (residues 605–682), IA-2-PTP-specific (unique to the IA-2 PTP domain and not shared with IA-2β PTP as determined by competition), IA-2/IA-2β-PTP-crossreactive (shared between IA-2 and IA-2β as determined by competition), and IA-2β-PTP-specific (unique to the IA-2β PTP domain and not shared with IA-2 PTP as determined by competition).

IgG subclasses and isotypes of IAA, GADA, and IA-2A were determined by radiobinding assays as previously described (23) using IgG subclass or isotype specific biotin-labeled mouse-anti-human monoclonal antibodies (Becton Dickinson, San Diego, CA) bound on Sepharose 4B streptavidin beads (Zymed, San Francisco, CA). The antibodies used were mouse monoclonal antibodies against human IgG1 (clone G17-1), IgG2 (clone G18-21), IgG3 (clone G18-3), IgG4 (clone JDC-14), IgM (clone G20-127), IgA (clone G20-359), and IgE (clone G7-26). Nonspecific binding was determined for each serum using anti-rat IgM monoclonal antibody (clone G53-238)-coated beads. The absence of cross-reactivity between subclasses was confirmed using high-titer human IgG1 monoclonal GADA and IA-2A antibodies. Results for IAA subclasses were expressed as nanounit of insulin bound per milliliter after subtraction of binding with the anti-rat IgM-coated beads. The cutoff for positivity for each IAA IgG subclass was 150 nU/ml (mean plus 3 SD of IAA-negative control subjects). GADA and IA-2A subclasses were expressed as difference (Δ) in counts per minute (IgG subclass or isotype specific cpm − anti-rat IgM cpm) and converted to an SD score (SDS). The cutoff for positivity for each GADA and IA-2A IgG subclass or isotype was 3 SDS, respectively.

Associations of variables with antibody titers were analyzed using the χ² test for trend. Time-to-event methods (life-table analysis and Cox proportional hazards model) were used to compare outcome (diabetes status) for participants with different covariate categories. The time between first antibody-positive sample and diagnosis of diabetes was defined as the time to event in relatives developing diabetes. Analyses considered censoring in relatives with diabetes-free status at the end of the follow-up period defined as date of last contact or date of entry into an intervention trial. Significance between groups was determined using the log rank test. Hazard ratios (HRs) were determined using Cox’s proportional hazards model. The proportional hazards assumption in the Cox model was tested by including in the model the covariate in question and the interaction between time and that covariate. For all analyses, a two-tailed P value of 0.05 was considered significant. All statistical analyses were performed using the Statistical Package for Social Science (SPSS 11.0, Chicago, IL).

GADA ≥99th centile were detected in 149 of the 180 relatives (83%), IA-2A in 62 (34%), and IAA in 72 (40%). Single autoantibodies were detected in 100 relatives (56%), two islet autoantibodies in 57 (32%), and three islet autoantibodies in 23 (13%) (Table 1). IgG1 was the most frequently detected subclass for each of the autoantibodies (Fig. 1A–C). IgG3 islet autoantibodies were rare, except for IAA where its prevalence was similar to that of IgG2 and IgG4. IgG subclasses other than IgG1 were found in 66 (44%) GADA-positive relatives, 29 (47%) IA-2A–positive relatives, and 39 (54%) IAA-positive relatives (Table 2). IgA and IgM autoantibodies were detected in 12–17% of relatives for each autoantibody. IgE autoantibodies were rare (<2%) in all groups.

Of the GADA-positive relatives, 115 (77%) had autoantibodies to multiple epitopes of the antigen, including 113 with antibodies to both GAD65–MID and GAD65–COOH-terminal epitopes (Fig. 1D). For IA-2A, 34 (55%) relatives had antibodies to multiple epitopes, and 30 (48%) relatives had antibodies that bound epitopes shared with IA-2β (Fig. 1E).

For each islet autoantibody, the detection of autoantibody IgG subclasses other than IgG1 were associated with high-titer antibodies, and for GADA and IA-2A, antibodies to multiple epitopes were associated with high-titer antibodies (Fig. 2), suggesting that multiple subclass usage and antibody epitope spreading were markers of pronounced high-titer humoral responses.

The overall 5- and 10-year cumulative risks for disease were 21% (95% CI 14.8–27.2) and 39% (30.6–47.4), respectively. By univariate analysis, diabetes risk was significantly associated with the number and combinations of autoantibodies (Table 1), IA-2A titer and IAA titer (Table 2, Fig. 3), IA-2A IgG subclasses and IAA IgG subclasses (Tables 2 and 3), and IA-2A epitope reactivity (Tables 2 and 4).

Diabetes risk was significantly higher in relatives with two or more autoantibodies (61% by 10 years; 95% CI 46–76%) than in relatives with one autoantibody (25%; 15–35%; P < 0.0001) and, among single autoantibody-positive relatives, risk was significantly higher in relatives with IA-2A alone (47% by 10 years; 13–81%) than in those with GADA alone (22%; 11–33%; P = 0.01; Table 1).

IA-2A-positive relatives with titers in the upper three quartiles (>27 units) had significantly higher diabetes risk (79% by 10 years; 95% CI 64–94%) than IA-2A–positive relatives with IA-2A titers in the lowest quartile (20%; 1–45%; P = 0.002), and IAA-positive relatives with IAA titers in the upper quartile (>9 units) had significantly higher risk (77% by 10 years; 53–99%) than IAA-positive relatives with IAA titers in the lower three quartiles (37%; 30–54%; P = 0.002; Table 2, Fig. 3). No relationship was found between GADA titers and risk of diabetes.

In the GADA-positive relatives, no relationship was found between GADA IgG subclasses and the risk of diabetes (Tables 2 and 3). In the IA-2A–positive relatives, risk was higher in those with IgG2, IgG3, or IgG4 IA-2A (100% by 10 years; 95% CI 85–100%) than those without these IgG subclasses (37%; 18–48%; P = 0.0007; Tables 2 and 3). In the IAA-positive relatives, risk was higher in those with IgG2, IgG3, or IgG4 IAA (68% by 10 years; 47–89%) than those without these IgG subclasses (28%; 10–46%; P = 0.007; Tables 2 and 3).

No association was found between GADA epitope reactivity and the risk of diabetes (Tables 2 and 4). In the IA-2A–positive relatives, type 1 diabetes risk was higher in those with autoantibodies that bound IA-2β (86% by 10 years; 95% CI 70–99%) than IA-2β antibody-negative relatives (38%; 18–58%; P = 0.008; Tables 2 and 4), and in relatives with antibodies binding multiple epitopes (75% by 10 years; 57–93%) than single epitope reactivity (50%; 26–74%; P < 0.05; Table 2).

Covariates found to be significantly associated with type 1 diabetes in univariate analysis, namely autoantibody number, IA-2A titer, IAA titer, IA-2A subclass, IAA subclass, multiple IA-2 epitopes, and IA-2β positivity, were included in the Cox proportional hazards model (Table 5). IA-2A titer >25th centile of positives (adjusted HR 5.4; 95% CI 1–29), IA-2A IgG2, IgG3, or IgG4 subclass positive (3.3; 1.4–8.1), and IAA IgG2, IgG3, or IgG4 subclass positive (4.6; 1.5–14) significantly contributed to the proportional hazards model. The adjusted HR for the presence of multiple autoantibodies, IAA titer, multiple IA-2A epitopes, and IA-2β autoantibody status did not reach statistical significance.

Since testing for IgG subclasses is currently expensive and not in general use, the multivariate analysis was also performed without the IgG subclass covariates. In this model, IA-2A titer >25th centile of positives (adjusted HR 5.1; 95% CI 1.1–24.2) and IAA titer >75th centile of positives (2.4; 95% CI) significantly contributed to diabetes risk. Removal of the IgG subclass covariates from the model, however, significantly decreased the fit of the proportional hazards model (χ², 18.8; P < 0.0001).

Based on the univariate and multivariate analyses, four models were selected, and their ability to stratify diabetes risk compared (Fig. 4, Table 6). Model 1 was based on antibody number with classification of relatives into one islet autoantibody (A), any two autoantibodies (B), and all three autoantibodies (C). Model 2 included all antibody characteristics that contributed significantly to the proportional hazards model. Relatives were classified into those having none (A), one (B), two (C), or all three (D) of the significant diabetes risk covariates (IA-2A titer >25th centile of positives; IgG2 or IgG4 IA-2A subclass positive; and IgG2, IgG3, or IgG4 IAA subclass positive). Model 3 was based on the multivariate analysis that excluded IgG subclasses with classification of relatives into those with none (A), one (B), or both (C) of the significant diabetes risk covariates (IA-2A titer >25th centile of positives and IAA titer >75th centile of positives). Model 4 was based on IA-2A and IA-2β autoantibody measurement with relatives classified into IA-2A negative (A), IA-2A positive and IA-2β autoantibody negative (B), and IA-2β autoantibody positive (C). Each of these models was able to stratify diabetes risk in the entire cohort (all P < 0.0001) (Fig. 4A). Models 2, 3, and 4 significantly stratified diabetes risk both within the relatives who had single islet autoantibodies (Fig. 4B) and the relatives who had multiple islet autoantibodies (Fig. 4C). Addition of models 2, 3, or 4 to model 1 in Cox’s proportional hazards model significantly improved diabetes risk stratification over that provided by antibody number alone (model 2, P < 0.0001; model 3, P < 0.0001; model 4, P < 0.0001). Models 2, 3, and 4 could identify autoantibody-positive relatives with a 5-year diabetes risk that was >50%. Model 2 provided the widest risk discrimination (5- and 10-year risks: 7 and 20% in category A, 19 and 47% in category B, 54 and 85% in category C, and 89 and 100% in category D) (Table 6).

Autoantibodies to islet autoantigens are the most widely used markers of pre-type 1 diabetes, and risk of diabetes in relatives can be defined on the basis of the number of islet autoantibodies detected, together with first-phase insulin response to intravenous glucose (24). We have shown that diabetes risk can be further stratified if other islet autoantibody characteristics including antibody titer, IgG subclass, and/or positivity against IA-2β are also taken into consideration. Risk of type 1 diabetes was increased in relatives with titers of IA-2A and IAA above a threshold that was substantially higher than the 99th centile of control subjects, but risk was not related to GADA titer. The breadth of the humoral response as measured by IgG subclass usage and epitope reactivity was directly related to antibody titer, and for IA-2A and IAA these parameters also correlated with risk. IA-2A were particularly strong predictors of type 1 diabetes in this cohort and indicated high risk even in the absence of antibodies to GAD or insulin. On the basis of these findings we have developed a strategy that, using autoantibody testing alone, is able to identify a subgroup of relatives with an 89% risk of diabetes within 5 years and a subgroup of relatives with islet autoantibodies whose risk was <10%.

The particular strengths of this study are the relatively large number of islet autoantibody-positive relatives followed, the duration of follow-up, and the extensive autoantibody characteristics examined. The number and combinations of islet autoantibodies found in the study cohort were similar to those in biochemical islet autoantibody-positive relatives identified in the Diabetes Prevention Trial 1 (DPT-1) clinical trial (9), suggesting that the findings are unlikely to be biased by inappropriate selection. Limitations of the study include the lack of metabolic data, which means that we have not been able to compare the models based on autoantibodies alone with others that include both autoantibody and metabolic testing, and the complexity of the analyses, which could result in false- positive or false-negative associations. Previous studies have found that some family members with all four islet autoantibodies (ICA, GADA, IA-2A, and IAA) appear relatively protected from developing diabetes (6,13), thereby limiting the use of autoantibody testing alone for accurate risk assessment. We also found that up to 30% of relatives positive for three autoantibodies are likely to be diabetes-free for at least 10 years. Although confidence intervals were wide, the study has been able to identify a subset of relatives with extremely high risk (all developed type 1 diabetes within 6 years of follow-up). Metabolic testing is unlikely to have improved risk stratification in this group, though we cannot exclude the possibility that it would have been useful in the remaining relatives at lower risk.

A general inverse relationship between antibody level and potentially β-cell-destructive T-cell responses, particularly to GAD, has been suggested (25). We have, however, previously shown that ICA titer was directly related to risk of type 1 diabetes (12) and have now found that high titers of IA-2A and IAA, though not GADA, are also associated with increased risk. Together with the observation that titer is related to IgG subclass usage and epitope reactivity, these findings do not support the hypothesis that a strong humoral response to autoantigens marks a protective nondestructive Th2 response. IA-2A were particularly strong predictors of diabetes in this cohort, and as in other studies (26,27), were highly specific indicators of risk even in the absence of autoantibodies to GAD or insulin. In contrast, GADA and IAA are found in other diseases and, in the absence of IA-2A, are associated with a relatively low risk or slow progression to diabetes (4,28). IA-2A may develop later than IAA and GADA in the preclinical disease (7,29). A marked rise in IA-2A titer is seen in some relatives close to diabetes onset (P.A., unpublished observation) and, in the model of islet transplantation where alloislets are placed into patients with long-standing type 1 diabetes, IA-2A are activated only when there is clear evidence of alloimmunity (30), whereas GADA can be immediately activated upon exposure to islet mass without evidence of alloreactivity. These observations support the hypothesis that IA-2A are markers of active β-cell destruction and suggest that autoimmunity against IA-2 participates in advancing β-cell destruction to clinical disease or has characteristics that promote the process.

The associations between diabetes risk and autoantibody characteristics such as subclass and epitope reactivity have varied between studies (15,16,23,31–39). Most have, however, examined these characteristics only for single antibodies without considering the other islet autoantibodies present or antibody titer. In this cohort, IA-2A and IAA IgG subclass and IA-2A epitope reactivity were found to modify risk. Although other diabetes-associated characteristics were strongly associated with high-titer antibodies, multivariate analysis indicated that IgG subclass status significantly improved risk estimation based on IA-2A and was superior to titer for IAA. The reason for this additional benefit is unclear. These characteristics may reflect specific mechanisms underlying the pathogenesis of type 1 diabetes or may simply improve the accuracy of measurement of antibody titer, which may appear artifactually low if antibodies bind to multiple epitopes. We favor the hypothesis that titer is the primary marker of diabetes risk and that multiple IgG subclasses and IA-2β positivity act as confirmatory markers of high titer responses.

The size of our study has allowed us to consider a number of models combining different antibody characteristics. The most effective model included quantification of IA-2A and measurement of IA-2A and IAA IgG subclasses. The feasibility of applying such a model to clinical trials depends on practical considerations such as the ability to measure these characteristics reproducibly. Preliminary studies suggest that quantification of IA-2A is relatively concordant between laboratories (20) and initial assessment of GADA IgG subclass measurement indicated that some, though not all, laboratories have sensitive, specific, and concordant assays (E.B., unpublished observation). The strength of the associations between type 1 diabetes risk and IA-2A titer, IA-2A IgG subclass, and IAA IgG subclass suggest that efforts at standardized and reliable measurement are worthwhile. An alternative, less effective, but relatively simple model was based on only IA-2A and IA-2β antibody testing, and even a single measurement of IA-2β antibodies was an effective test to identify those at highest risk. The number of islet autoantibodies did not significantly improve the Cox models based on antibody titer, subclasses, and/or epitopes (data not shown). Nevertheless, the risk assessment models were effective in stratifying risk both in relatives who were positive for a single islet autoantibody marker and in relatives who had multiple islet autoantibodies, suggesting that other variations of these models that also include the number of islet autoantibodies might be effective in stratifying diabetes risk.

We have shown that the combination of autoantibody titer, subclass, and/or epitope reactivity may improve type 1 diabetes risk stratification, which could be effectively stratified on the basis of these characteristics in a single sample. These observations need to be validated in other large cohorts but, if replicated, they have the potential to simplify screening and recruitment for clinical trials.

This work was supported by a grant from Deutsche Forschungsgemeinschaft (ZI 310/12-5). The BOX study is supported by Diabetes U.K. P.A. received support from Deutsche Diabetes-Gesellschaft (Projektförderung 2002).

The authors thank Annette Knopff, Kerstin Koczwara, and Alastair Norcross for technical support and Markus Walter and Michael Hummel for clinical assistance. We would like to thank all study subjects for their participation.

This study forms parts of the dissertations of K.W. and J.R.