Data-Driven Phenotyping of Presymptomatic Type 1 Diabetes Using Longitudinal Autoantibody Profiles

Clinical diagnosis of type 1 diabetes is preceded by asymptomatic islet autoimmunity (IA) with the presence of islet autoantibodies (IAbs) such as insulin autoantibody (IAA), antibody against GAD (GADA), and islet antigen 2 antibody (IA-2A). IA predicts development of type 1 diabetes, but heterogeneous islet autoantibody (IAb) profiles and dynamics before diagnosis require further investigation. The rate of progression from the initiation of IA to type 1 diabetes is highly variable (1–6). Several studies have investigated characteristics of different subgroups of patients who progress differently to type 1 diabetes, which ultimately could be useful for identifying children at risk of rapid or slow progression to the disease. For example, individuals have been categorized by family history of type 1 diabetes, genetic risk, age at development of IA, or type of first appearing IAb. In young children, IAA is the most sensitive and IA-2A is the most specific predictor for development of diabetes, and children who develop multiple IAbs early in life progress faster to diabetes than those who develop multiple IAbs at a later age (7–10). The TEDDY study showed that the presence of IA-2A predicts fast progression to clinical diabetes in young children (11). However, the development of IAbs over time can be heterogeneous; therefore, further investigation is required to define the IAb patterns associated with rapid or slow progression and understand the role of phenomena such as reversion of individual IAbs (12). In our collaboration, the T1DI (Type 1 Diabetes Intelligence) study group, we combined and harmonized data from five prospective studies of type 1 diabetes (13) for such analyses.

In most previous studies, subgroups of IAb profiles have been defined based on clinically predefined hypotheses and not on data-driven analyses. Interestingly, a study from BABYDIAB (14) reported a clustering algorithm for a cohort of 88 children with multiple positive IAbs using their longitudinal autoantibody profiles. However, the clustering algorithm did not account for the age at which the IAb was detected, and it did not differentiate between different IAb types (IAA, IA-2A, GADA), although the specificity of IAbs may influence the rate of progression. In addition, no principled approach to determine the number of clusters was used. In a more recent report, an improved clustering algorithm was presented to take into account the timing of changes in IAb development, and the algorithm was applied to 370 children who developed multiple autoantibody types in the evenly and frequently sampled TEDDY study (15). However, that algorithm may not be applicable to data with irregular measurements, such as the data in the T1DI study or in clinical practice. In the current report, we present a novel clustering algorithm that addresses these limitations. We applied it to a large cohort of 1,845 IAb-positive children from the T1DI study cohort to identify putative disease subtypes in children with different IAb profiles. This data-driven approach revealed clusters of children with distinct characteristics. In addition, we explored in the youngest children developing IA whether the type of IAb in the first positive sample was associated with progression to type 1 diabetes.

We analyzed data from the T1DI study cohort, which combined five prospective longitudinal cohorts: the DIPP (Type 1 Diabetes Prediction and Prevention) study from Finland (4), the DiPiS (Diabetes Prediction in Skåne Study) from Sweden (16), the DAISY (Diabetes Autoimmunity Study in the Young) from the U.S. (17), the DEW-IT (Diabetes Evaluation in Washington) study from the U.S. (18), and the BABYDIAB study from Germany (8). The analyzed study cohort consisted of 1,845 individuals (male 822, female 1,023) who developed at least one IAb and had at least three visits during follow-up. Of these participants, 498 developed type 1 diabetes during follow-up. The median follow-up time was 12.3 years (interquartile range 8.3, 15.4), and the median age at first IAb test was 3.5 months (interquartile range 2.6, 8.9). There were 1,034 individuals from DIPP, 201 from DiPiS, 270 from DAISY, 183 from DEW-IT, and 157 from BABYDIAB. Individuals were categorized into four HLA DR-DQ risk groups based on published odds ratios for type 1 diabetes (A very high risk, B high risk, C slightly elevated risk, and D average to low risk) (13).

IAbs were measured from serum or plasma samples by using standard methods, as described in each original study (13). IAb profile here refers to positivity at a single visit. For example, as shown in Supplementary Fig. 1, the IAb profile of participant A at age 3.2 years indicates that GADA was positive but IAA and IA-2A were negative. IAb dynamics refers to the progression of the IAb profile over time (i.e., how the IAb profile changes over time). First appearance of IAb refers to the time of the first measurement where the IAb was observed to be positive. Seroconversion refers to the positivity of the same IAb in two consecutive visits. Seroconversion age is the age at the first of these two measurements. Persistent IAb refers to the IAb becoming stable positive in two or more consecutive measurements and not reverting to negative. Transient IAb indicates that the IAb was positive once but then reverted to negativity and became stable negative (e.g., pos-neg-neg-neg). Fluctuating IAb indicates that the IAb was switching between positivity and negativity and was unstable (e.g., pos-neg-pos-neg). Reverting IAb indicates that the IAb was found to be stable positive and then reverted to stable negative (e.g., pos-pos-pos-pos-neg-neg-neg).

This novel algorithm aligns two individuals based on their IAb dynamics. Essentially, it computes and assigns a similarity score based on how closely the sequences of the two IAb profiles match each other. If the IAb profiles in the two individuals progress similarly, a higher similarity score (close to 1) is assigned, indicating that the two individuals have similar IAb dynamics. We extended the clustering algorithm described by Endesfelder (14) and present a novel approach that takes into account 1) the age at which the IAbs were measured, 2) the type of IAbs at that age, and 3) whether the individuals match either positive or negative measurements. Detailed development of the algorithm is described in the Supplementary Material. In addition, we used the proportion of ambiguously clustered pairs (PAC) method to determine the number of clusters that best describe the data (Supplementary Fig. 3) (19). The novel approach consists of the following steps:

Similarity was computed between two IAb profiles by considering which type of IAb was positive and when (what age) it was detected (20). We used individual autoantibody profiles across time to compute the similarity score (Supplementary Fig. 1). A similarity score of 0 indicated that the two individuals were dissimilar with respect to their IAb sequences, whereas a score of 1 indicated that the two sequences of IAb were the same. First, each individual was encoded in a binary matrix where the columns were IAb types, the rows were visits, and the entries indicated whether each IAb was positive or negative. Because a positive IAb match between two individuals was rarer than a negative IAb match, we assigned a weight to an IAb match based on its prevalence in the data. For example, if the IAb was rarely observed as positive, a higher weight was assigned to emphasize the positive match for that IAb between two profiles, because it was unlikely to happen otherwise. In addition, two IAb profiles that were closer in terms of time (i.e., the age of the two participants) were considered more similar than two profiles that occurred farther apart, given that the profiles were otherwise similar. To address this, the similarity between two profiles was weighted based on the time gap between the profiles and is represented by the time-aware profile distance score.

Two IAb profiles were aligned by considering their temporal dynamics using the similarity score between a profile pair of two individuals and the computed time-aware profile distance scores (21).

Individuals were clustered by using the computed and aligned similarity scores for all pairs of individuals. The hierarchical/agglomerative clustering algorithm was applied.

The number of stable clusters in the study cohort was determined using the PAC algorithm (19), which uses bootstrapping for stability analysis to identify robust clusters.

The characteristics of individuals in each cluster were analyzed separately. Kaplan-Meier survival analysis was used to examine the rate of progression from the first positive autoantibody to diagnosis of stage 3 type 1 diabetes. The log-rank test was used to compute the statistical significance between the progression rates, and a P value <0.05 was considered significant.

The data supporting these findings are available from the authors on reasonable request. The data are not publicly available because of privacy regulations.

When applying the novel data-driven time-aware clustering algorithm to the 1,845 IAb-positive individuals, five large clusters of individuals with different IAb profiles and dynamics were discovered: 5C1, 5C2, 5C3, 5C4, and 5C5 (Fig. 1A–E). The characteristics of children in these clusters, including the 5- and 10-year risks of type 1 diabetes, are presented in Table 1. The cumulative incidences of type 1 diabetes for the five clusters are shown in Fig. 1F. The children in cluster 5C2 (n = 89) had a high progression rate to diabetes (5-year risk of 69.9% and 10-year risk of 89.9%). All three IAbs were persistently positive in cluster 5C2; however, on average, IAA appeared first and was quickly followed by GADA and IA-2A, and diabetes was diagnosed at a median age of 4.3 years (Fig. 1B and Table 1). The individuals in clusters 5C3 (n = 464) and 5C4 (n = 49) also had relatively high progression rates to diabetes (5-year risks of 30.9% and 39.1%, respectively), but their median age at diagnosis was older than that in cluster 5C2 (9.2 and 8.2 years, respectively). Clusters 5C3 and 5C4 were characterized by two persistent IAbs: 5C3 by persistent IA2A and GADA but less frequent and reverting IAA, and 5C4 by persistent IAA and GADA but rare and transient IA-2A (Fig. 1C and D and Table 1). Most individuals in cluster 5C5 (n = 168) (Fig. 1E and Table 1) developed late persistent GADA. During follow-up, IAA was positive in 64.3% of the individuals in 5C5, but it often reverted, and only 7.2% were IAA positive in their last positive sample. IA-2A was positive in 34.5% of the individuals in 5C5, rarely in the first positive sample and infrequently also in the last positive sample. The 5-year risk of type 1 diabetes in cluster 5C5 was 10.5%. Finally, cluster 5C1 (n = 1,075) (Fig. 1A and Table 1) consisted of individuals who developed mainly single and transient IAbs, most often IAA or GADA, and had an estimated 5-year risk of type 1 diabetes of only 1.6%. The proportions of the four class II HLA genotype groups did not differ between the five main clusters (Table 1).

The PAC analysis also revealed 18 smaller but still stable clusters of individuals with typical progression patterns of IA. These 18 clusters were subclusters for the five main clusters (Fig. 2A). The cumulative type 1 diabetes incidence in the children in the 18 subclusters is shown in Fig. 2B. Subclusters 18C1, 18C2, and 18C4 originated from 5C1 and were characterized with positivity for a single, often reverting, IAb (IA-2A, IAA, or GADA, respectively) and low 5- and 10-year diabetes risks (Supplementary Fig. 4A–C and Supplementary Table 2). Eight of the remaining 15 subclusters (18C7, 18C10, 18C12, 18C13, 18C14, 18C16, 18C17, and 18C18) had at least 10 participants per cluster, and their IAb dynamics and clinical characteristics are also shown in Supplementary Fig. 4D–K and Supplementary Table 2).

We also validated the hypothesis that the long-term risk of type 1 diabetes is different based on the first appearing IAb in early life. We observed that children who seroconverted before age 2 years and developed two or more IAbs (n = 254) represented two groups based on IAA status in the first positive sample. In group A (n = 194), IAA was positive with or without other IAbs in the first positive sample, whereas in group B (n = 60), IAbs other than IAA were present in the first positive sample. Progression rates to diabetes in these two groups were significantly different (log-rank test P = 0.0002) (Fig. 3). Children who were IAA positive in their first positive sample had a 10-year risk of 71.4% (95% CI 64.5–77.9), 15-year risk of 82.7% (95% CI 76.1–88.3), and 20-year risk of 92.1% (95% CI 81.3–97.9), whereas children without IAA in their first positive sample had a 10-year risk of 47.6% (95% CI 35.2–61.8), 15-year risk of 53.9% (95% CI 40.2–68.7), and 20-year risk of 61.6% (95% CI 44.0–79.3).

Identification of IAb temporal patterns that are associated with progression rates to type 1 diabetes is crucial to improve prediction, understand various disease pathways, and design targeted prevention trials. For example, young children with IAb profiles and dynamics associated with a high 5-year progression rate need early interventions specific for their disease subtype. In addition, information on distinct IAb profiles and expected risk of progression to clinical diabetes will be valuable when providing counseling on diabetes risk to individuals with IAb positivity. Here we present a novel clustering algorithm that matches similar children based on their IAb dynamics and accounts for the various types of IAbs and the age when an IAb first appeared. Moreover, we show that this algorithm can be applied to individuals with different follow-up protocols. We applied the algorithm to a cohort of 1,845 IAb-positive children and discovered five main groups of children with different IAb profiles and dynamics, which were strongly associated with different progression rates to stage 3 type 1 diabetes. Two of the discovered clusters had similar patterns of IAb development and diabetes risk reported in earlier studies, but three of the clusters represented less well-described novel patterns of IA.

Earlier efforts to identify IAb temporal patterns had some limitations and were applied in much smaller cohorts than our T1DI cohort. We developed a novel clustering algorithm that addresses several limitations of the earlier algorithms and applied it to a large cohort of individuals with one or more positive IAb (IAA, GADA, and/or IA-2A). The T1DI cohort is the largest cohort of IAb-positive children currently available and represents individuals from five prospective studies from Europe and the U.S. This also imposed some challenges in using the combined data, such as how to cluster children with different sampling frequencies (0.6–2.0 years in the five original cohorts). Our novel time-aware clustering algorithm addresses the variation in sampling interval and matches children based on types and combinations of IAbs over time and by the ages at which the IAbs were observed. In addition, the novel algorithm accounts for the overall variation in positivity for the various autoantibody types in the cohort. The prevalence of IAA in the entire T1DI cohort was 5%, and that of GADA was 8%, showing that IAA was less often observed than GADA (Supplementary Table 1). Having two measurements with positive IAA should therefore be weighted more heavily than two measurements with positive GADA. We addressed this issue by weighting the positivity match differently for different IAb types. Furthermore, we used a principled way to determine the number of stable clusters in the cohort.

We discovered five distinct clusters, each with a typical IAb pattern and timing. In addition, the novel algorithm distinguished a total of 18 smaller subclusters that were stable and represented subgroups of children from the five main clusters. Cluster 5C2 and subcluster 18C7 included children with early initiation of IA, with persistent IAA, GADA, and IA-2A, and diagnosis of type 1 diabetes at age ∼4 years. This group represents a specific and aggressive subtype of type 1 diabetes that has also been recognized in earlier studies (1,10,15,22). Interestingly, the 464 children in cluster 5C3 were also characterized by persistent GADA, IA-2A, and IAA at an early age, but they frequently lost IAA during follow-up and were diagnosed with diabetes later, at age ∼9 years. This pattern in cluster 5C3 has been less described in the literature, with inconsistent conclusions. The pattern of being positive for the three IAbs and then losing IAA was also described by Endesfelder et al. (14), who reported a lower 10-year risk of 23% in children losing IAA compared with 76% in those who maintained IAA positivity. In contrast, the TEDDY study reported that in multipositive children, IAA reversion had little effect on the risk of type 1 diabetes (12). Cluster 5C4 included children with persistent positivity for IAA and GADA from age 3 years, but only few of them were ever positive for IA-2A. This is also a novel group, with a high 10-year risk of type 1 diabetes of 73.8%, and demonstrates that although the presence of IA-2A has been shown to be an important predictor of diabetes in young children (11,22), it is not necessary for disease development. The third relatively novel cluster included children in cluster 5C5 (and subcluster 18C17) who typically developed persistent GADA at age ∼6 years, less frequently had positivity for IAA or IA-2A, and progressed to type 1 diabetes rather late, at age ∼12 years. Single positivity for GADA is commonly observed in adults diagnosed with type 1 diabetes but is less described in children. The pattern in cluster 5C5 is in line with observations from the TEDDY study, reporting a significantly higher 5-year risk of diabetes in children with single but stable positivity for IAA than in those with single and stable GADA positivity (15). The largest cluster, 5C1, represented individuals with mainly single and transient IAbs associated with a low risk of progression to diabetes. This group is well known from earlier studies but can only be identified by observing IAbs in consecutive samples after first detection.

It is tempting to hypothesize that the initiatory factors of early and aggressively progressing disease subtypes are different from those triggering a different IAb pattern later in childhood. Therefore, for example, the children in clusters 5C2 and 5C5 should be studied in greater detail to identify possible specific initiators of the disease, such as whether they were environmental or genetic or a combination of both. The effective preventive interventions may be different for children in these groups.

It has been hypothesized that there are different subtypes of type 1 diabetes based on the order of IAb appearance (23,24). IAA as the first appearing IAb has been associated with fast progression to type 1 diabetes, whereas GADA as the first appearing IAb has been associated with a more moderate progression rate. This was also apparent in the clusters discovered by our novel algorithm. In our validation analysis of individuals with early seroconversion before age 2 years, positivity for IAA alone or in combination with other IAbs in the first positive sample was associated with a much higher progression rate to diabetes compared with IAA negativity in the first positive sample.

The strengths of this study include a large study cohort of individuals with IAb positivity, a long median follow-up time of 12.3 years, and measurements of IAA, GADA, and IA-2A from more than 260,000 individual samples. The T1DI cohort also includes a variety of class II HLA genotypes (13), and therefore, the T1DI study population resembles the general population more than the cohorts included in previous clustering analyses. There are also some limitations in our study. We did not include ZnT8A data, because the measurements were not systematically performed in the individual study cohorts. ZnT8A measurements may further improve risk stratification in the future. Current knowledge suggests that ZnT8A is not often seen at the initiation of IA but appears later during follow-up. If ZnT8A had been available for our analyses, it is likely that we would have seen even more heterogeneity in IAb patterns compared with our current results. No external validation cohort was available to repeat our results, and this will require further work in the future, before potential implementation of the algorithm for risk stratification. The clustering algorithm could potentially be developed to include additional factors, such as IAb levels and comprehensive genetic risk profiles of the participants for similarity matching. However, our results clearly demonstrate that IA in children, although heterogeneous, can still be used to group by disease subtype. It is important to learn more about disease subtypes in the future. Large data sets and data-driven analyses, as exemplified here, are excellent tools to clarify distinct pathogenetic pathways of type 1 diabetes and inform future research and practice.

In conclusion, we developed a clustering algorithm to define temporal patterns of IA, applied it in a large cohort of IAb-positive children, and discovered distinct patterns of IA that were strongly associated with varying progression rates to type 1 diabetes. These findings have implications for type 1 diabetes prediction and future prevention trials. Furthermore, in ongoing screening programs of IA, these findings may guide the interpretation of individual test results.

Acknowledgments. The authors thank the participants of the DAISY, DiPiS, DIPP, DEW-IT, and BABYDIAB studies and the dedicated study personnel at the study sites.

Funding. This work was supported by funding from JDRF (IBM: 1-RSC-2017-368-I-X, 1-IND-2019-717-I-X; DAISY: 1-SRA-2019-722-I-X, 1-RSC-2017-517-I-X, 5-ECR-2017-388-A-N; DiPiS: 1-SRA-2019-720-I-X, 1-RSC-2017-526-I-X; DIPP: 1-RSC-2018-555-I-X, 1-SRA-2019-721-I-X; DEW-IT: 1-SRA-2019-719-I-X, 1-RSC-2017-516-I-X) as well as from the National Institutes of Health (DAISY: DK032493, DK032083, DK104351; DiPiS: DK26190). DIPP was funded by JDRF (1-SRA-2016-342-M-R, 1-SRA-2019-732-M-B), the European Union (BMH4-CT98-3314), the Novo Nordisk Foundation, the Academy of Finland (decision no. 292538) and the Centre of Excellence in Molecular Systems Immunology and Physiology Research 2012-2017 (decision no. 250114), special research funds for University Hospitals in Finland, the Diabetes Research Foundation (Finland), and the Sigrid Juselius Foundation (Finland). BABYDIAB was supported by funds from the German Federal Ministry of Education and Research to the German Center for Diabetes Research. DiPiS was funded by the Swedish Research Council (14064), the Swedish Childhood Diabetes Foundation, the Swedish Diabetes Association, the Nordisk Insulin Fund, SUS funds, the Lions Club International (district 101-S), the Royal Physiographic Society, the Skåne County Council Foundation for Research and Development, LUDC-IRC/EXODIAB funding from the Swedish Foundation for Strategic Research (decision no. IRC15-0067), and the Swedish Research Council (decision no. 2009-1039). DEW-IT was funded by the Centers for Disease Control and Prevention (UR6/CCU017247), with additional support from the University of Washington Diabetes Research Center (P30 DK017047), the Hussman Foundation, and the Washington State Life Science Discovery Fund.

Duality of Interest. M.G., V.A., and K.N. are employees of IBM. J.L.D. performed this work as an employee of JDRF and is now an employee of Sanofi. O.L. is an employee of JDRF. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. M.G. conceptualized and developed the time-aware clustering algorithm, was responsible for all data analyses, and drafted and revised the manuscript. V.A. contributed to the interpretation of data, revised the manuscript, and is the guarantor of data integration, as a representative of IBM. K.N. contributed to the interpretation of data and revised the methods section of the manuscript. J.L.D., O.L., M.L., W.A.H., M.R., and A.G.Z. contributed to the interpretation of data and revision of the manuscript. M.L., W.A.H., M.R., A.G.Z., and R.V., as principal investigators of the individual studies, verify the underlying data from original study sites. R.V. contributed to the analysis plan and interpretation of data and drafted and revised the manuscript. All authors contributed to the discussion and/or revision of the final draft of the manuscript. V.A. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Handling Editors. The journal editors responsible for overseeing the review of the manuscript were Steven E. Kahn and Thomas Danne.