Toward an Improved Classification of Type 2 Diabetes: Lessons From Research into the Heterogeneity of a Complex Disease

Abstract

Context

Accumulating evidence indicates that type 2 diabetes (T2D) is phenotypically heterogeneous. Defining and classifying variant forms of T2D are priorities to better understand its pathophysiology and usher clinical practice into an era of “precision diabetes.”

Evidence Acquisition and Methods

We reviewed literature related to heterogeneity of T2D over the past 5 decades and identified a range of phenotypic variants of T2D. Their descriptions expose inadequacies in current classification systems. We attempt to link phenotypically diverse forms to pathophysiology, explore investigative methods that have characterized “atypical” forms of T2D on an etiological basis, and review conceptual frameworks for an improved taxonomy. Finally, we propose future directions to achieve the goal of an etiological classification of T2D.

Evidence Synthesis

Differences among ethnic and racial groups were early observations of phenotypic heterogeneity. Investigations that uncover complex interactions of pathophysiologic pathways leading to T2D are supported by epidemiological and clinical differences between the sexes and between adult and youth-onset T2D. Approaches to an etiological classification are illustrated by investigations of atypical forms of T2D, such as monogenic diabetes and syndromes of ketosis-prone diabetes. Conceptual frameworks that accommodate heterogeneity in T2D include an overlap between known diabetes types, a “palette” model integrated with a “threshold hypothesis,” and a spectrum model of atypical diabetes.

Conclusion

The heterogeneity of T2D demands an improved, etiological classification scheme. Excellent phenotypic descriptions of emerging syndromes in different populations, continued clinical and molecular investigations of atypical forms of diabetes, and useful conceptual models can be utilized to achieve this important goal.

Current definitions of type 2 diabetes (T2D) are based on broad pathophysiological concepts that provide little practical guidance to clinicians grappling with an expanding array of phenotypic presentations that elude precise diagnosis, and therapeutic options guided by consensus rather than targeted recommendations (1). Intensive research over decades has revealed that the pathophysiology of T2D is complex, variable, and multifactorial (2-5). It is now becoming increasingly clear that the phenotypes of patients diagnosed with “T2D” are equally complex and varied, likely reflecting heterogeneity of epigenetic, genetic, and acquired defects in intersecting pathways and multiple tissues leading to the final common pathway of hyperglycemia (4). Definitions of T2D as “[a condition]... due to a progressive loss of adequate β-cell insulin secretion frequently on the background of insulin resistance” (1), or “[a pathophysiologic process that]... ranges from predominantly insulin resistance with relative insulin deficiency to a predominantly secretory defect with or without insulin resistance” (6) could enfold virtually any type of diabetes and do not help to specify the varieties of patients around the world diagnosed with “T2D.” An etiological classification of T2D would facilitate precision medicine in diabetes, a highly desirable goal.

Here, we examine evidence for the emerging clinical heterogeneity of T2D, attempt to link its phenotypically diverse forms to pathophysiologic processes, review investigative methods that have proven successful in characterizing variant or atypical forms of diabetes from an etiological perspective, present conceptual frameworks to understand the heterogeneity of T2D, and suggest future directions to achieve the goal of an etiological classification of T2D.

The Phenotypic Heterogeneity of Diabetes Exposes Inadequacies of the Current Classification System

The heterogeneity and complexity of T2D have long been recognized (5, 6). Early observations of phenotypic heterogeneity and corresponding pathophysiologic complexity were made through comparisons of different ethnic/racial groups in diverse geographic regions (4). Beginning with case series in the 1960s through comparative cohort analyses in the 1970s and 1980s, clinical and phenotypic differences were reported in populations with varying risk factors and prevalence rates for T2D (4). Studies then sought to link these variations to differences in genetic susceptibility (7-9). For example, high rates of T2D were reported in “lean” East Asian and South Asian populations, suggesting a phenotypic contrast with obesity-associated T2D in Europid populations. The biological components measured in these studies included body mass index (BMI), waist circumference, dyslipidemia, hypertension, family history of diabetes, insulin sensitivity, and beta cell function, and the analyses of the study data demonstrated that these factors differentially affect population risk of T2D development and rate of progression to chronic complications (7-9). Comparative genetic analyses of the geographic or ethnic groups demonstrated variation in T2D risk loci based on sex (10), body fat distribution (11), and race or ethnicity (12-14). Guided by these findings, physiological investigations subsequently revealed the differences were linked to quantitative differences in insulin sensitivity of different adipose tissue depots (15, 16) or in the mass or function of beta cells (17, 18) between persons belonging to these different groups.

Such comparative studies continue to illuminate mechanisms responsible for phenotypic group differences and inform more complex gene-environment analyses. Minority ethnic groups in the United States have been shown to have higher rates of developing T2D (https://www.cdc.gov/diabetes/pdfs/data/statistics/national-diabetes-statistics-report.pdf) as the result of complex interactions between genetics, environment, and lifestyle, with large influences derived from social determinants of health (19). The observations include a propensity to greater degrees of insulin resistance in African Americans across BMI strata, and glucose metabolic abnormalities at lower levels of BMI in Chinese and South Asian individuals (20-22). Recently, the Mediators of Atherosclerosis if South Asians Living in America (MASALA) and the Multi-Ethnic Study of Atherosclerosis (MESA) studies demonstrated that race and ethnicity are markers of heterogeneity and differential risk for complications in T2D (23). South Asian individuals with T2D have higher glycated hemoglobin A1c (HbA1c), lower beta cell function (20, 21), younger age of onset, and higher incidence of coronary artery disease compared with non-Hispanic White, Chinese, African American, and Hispanic groups. Besides biological considerations, social disparities that influence diet, lifestyle, and access to care are important contributors to the phenotypic heterogeneity of the disease.

Beyond genetic susceptibility, a host of epigenetic influences related to prenatal exposures (24) and pregnancy “programming” (25), lifestyle factors and environmental exposures that affect metabolism both directly and through the mediation of the intestinal microbiome (26, 27), variations in lifestyle and socioeconomic status, as well as “life course” evolution of many of these factors (4, 28) are known to modulate the pathophysiology of T2D, and hence the phenotypic presentations of patients.

There are differences in T2D prevalence among the sexes, with women at younger ages having a higher predisposition to diabetes than men (29) (in contrast with type 1 diabetes [T1D], wherein there is a higher overall male predominance among youth). Importantly, from the standpoint of a useful etiological classification, the pathophysiology itself may be different on the basis of sex. For example, although the metabolic syndrome is considered almost invariably associated with T2D, inasmuch as insulin resistance is a foundation of both conditions, a recent report suggests that a substantial proportion (45%) of women with T2D have no generally accepted metabolic syndrome components (other than dysglycemia) (30). Males and females clearly have different pathophysiologic interactions of biological pathways leading to T2D.

T2D was once synonymous with “adult-onset” diabetes, but this concept was rendered obsolete in the late twentieth century by a rising epidemic of the condition in pediatric populations. The incidence of childhood-onset T2D has risen by 4.8% over the past decade (31-33), associated with the increasing prevalence of childhood obesity. Compared to adults with T2D, children with T2D lose insulin secretory capacity earlier in the course of the disease (34-36) and have more pronounced decrease in insulin sensitivity, higher rates of failure to respond to metformin alone or with rosiglitazone (37), and overall worse health outcomes due to a range of social and genetic factors (38-40). There is emerging heterogeneity within pediatric T2D as well. The small but distinct fraction of children who develop a T2D phenotype prior to or early in puberty (41) have unique features and pathophysiologic drivers distinct from the rise of pubertal hormones (such as growth hormone, insulin-like growth factor-1, and sex hormones) often implicated as a mechanism underlying T2D in older children and adolescents (41-43). Productive investigation toward etiologic definition of pre-pubertal T2D may involve the roles of exposures in the prenatal period (eg, due to maternal diabetes) or in early infancy (24, 44). Defining variations in pediatric T2D phenotype and course due to sex, ancestry, and geography also remain to be explored. Some children with T2D present with marked ketosis or diabetic ketoacidosis (DKA) even absent circulating islet autoantibodies, while others do not (45). Do the former simply represent a more rapid deterioration of beta cell function (and if so, why?), or as with adult A−ß+ ketosis-prone diabetes (see below), do they have novel defects in intermediary metabolism that accelerate ketone production or inhibit its oxidation?

As in adult T2D, pediatric natural history studies, as well as clinical trials coupled with population data analyses, such as the Search for Diabetes in Youth (SEARCH) (46), Treatment Options for type 2 Diabetes in Adolescents and Youth (TODAY) (47), or Restoring Insulin Secretion (RISE) (48) studies have contributed much to understanding pediatric T2D heterogeneity, and collaborations between the study groups have been useful in this regard (49). Progression to insulin dependence is variable in children with T2D (47), and one relevant factor is islet autoimmunity manifested by autoantibodies in some T2D children. These children have a higher likelihood of rapid progression to insulin dependence (50) due to early beta cell function loss (51). Terms such as diabetes 1.5, double diabetes (52), or latent autoimmune diabetes in youth (50) have sprouted to indicate these variants, but we should avoid the easy or eponymous attachment of labels that bring us no closer to an etiological classification than type 1 or type 2 diabetes.

Approach to an Etiological Classification Based on Understanding Atypical Forms of T2D

The process of defining the currently known monogenic syndromes of diabetes serves as a paradigm for an etiological classification of T2D diabetes based on careful linkage of “atypical” phenotypes with specific genetic causes. The clinical utility of this classification has been demonstrated in the manner in which it permits precise diagnosis and targeted therapy (53-56). Historically, the recognition of monogenic diabetes began with careful delineation of phenotypic characteristics such as early onset or autosomal dominant transmission of the trait and their initial classification as a separate entity denoted maturity-onset diabetes of youth (MODY). Aggregation of cases and families together with prospective population studies by dedicated expert groups (eg, at the Universities of Exeter and Chicago) have led to accurate screening modalities such as the phenotypic MODY risk calculator (57), and a genetic risk score that can differentiate adults and young adults with T1D from those with potentially monogenic diabetes (58). Following such screening or upon clinical suspicion, one can readily perform “MODY testing” using commercially available gene panels or rapid exome sequencing. In nononcologic adult clinical practice, this is one of the early examples of precision diagnosis made feasible by an etiological classification, and its benefits are clear whenever an apparently insulin-dependent patient is successfully transitioned to sulfonylurea (eg, persons with pathogenic variants in the hepatocyte nuclear factor-4α [HNF4-alpha] or hepatocyte nuclear factor-1α [HNF1-alpha] genes), or a person diagnosed with “diabetes” may dispense with diabetes medications altogether (eg, persons with variants in the glucokinase [GCK] gene), or a family member at risk is identified at a preclinical stage.

However, the rarity of known monogenic diabetes (under 4% in selected patient cohorts (59), thus even lower in the general diabetes population) makes this model less useful as a guide to classifying other atypical T2D phenotypes that may arise from more complex genetics or epigenetics. Longitudinal characterization of patients with such phenotypes and assessment of their natural histories have proven productive in delineating subgroups among these T2D variants, anticipating the development of “omics” tools for mechanistic analysis of their multifactorial etiologies. A good example lies in the evolution of our understanding of syndromes of ketosis-prone diabetes (KPD) (60).

KPD is defined by the striking presentation with DKA of patients lacking clinical evidence of autoimmune T1D. It is an emerging cluster of conditions recognized in larger numbers over the last 3 decades, among Asian, Afro-Caribbean, African American, US Hispanic, and South American populations (61-76). Based on longitudinal descriptions and multivariable analysis of a large database, an algorithm utilizing 2 biomarkers (islet autoantibodies and a measure of beta cell functional reserve, each present or absent by validated cutoffs) defines different forms of KPD with distinct pathophysiologic mechanisms (64). Of the 4 subgroups defined by this 2 × 2 factorial approach, the 2 autoantibody-negative (A−) groups represent complex atypical T2D syndromes. A−ß+ KPD (autoantibody absent, beta cell function present) comprises ~50% of patients presenting with DKA in a US urban multiethnic population. These patients can discontinue insulin therapy following the index DKA episode and about half can maintain excellent glycemic control on oral medications for many years (64). The latter subgroup of patients develop unprovoked DKA and are characterized by late onset, male predominance, obesity, DKA at initial diagnosis of diabetes, and lack of HLA class II T1D susceptibility alleles or T-cell reactivity to islet autoantigens (77, 78). Plasma metabolomics, validated by kinetic studies utilizing stable isotopes and mass spectrometry, has revealed that these patients have variable defects in branch-chain amino acid and arginine metabolism that may lead to abnormal ketogenesis but defective ketone oxidation (79, 80), and inability to maintain sufficient insulin secretion during hyperglycemic crises (81).

The clinical phenotype of A−β− (autoantibody absent, beta cell function absent) KPD resembles lean, late-onset T1D but the patients lack evidence for islet autoimmunity (absent islet autoantibodies and low frequency of HLA risk alleles for T1D) (64). In addition to identifying as patients who require life-long insulin therapy following the index DKA, they represent a valuable reservoir for discovery of novel monogenic or oligogenic mechanisms of severe beta cell failure using physiologic insulin secretion studies and more detailed analysis of patient-specific inducible pluripotent stem cells differentiated along the pancreatic islet lineage. Preliminary evidence indicates approximately a quarter have potentially pathogenic variants in the HNF-1α and pancreas-duodenum homeobox-1 (PDX-1) genes, encoding critical transcription factors for beta cell development (82).

The potential utility of the Aß classification has been demonstrated more broadly in application to pediatric diabetes (83), permitting identification of 4 groups with distinct characteristics at diagnosis and clinical course. The SEARCH investigators have used a similar 2 × 2 factorial analysis (presence/absence of islet autoimmunity and presence/absence of insulin resistance) to characterize different forms of diabetes, including T2D, in their childhood and adolescent population. The SEARCH classification of pediatric diabetes has genetic associations (84) and has proven useful for epidemiological discovery (85). However, this classification was not designed for individual diagnosis, nor has it been tested for clinical management. Development and application of similar classification algorithms in different prospectively followed cohorts with new-onset T2D should be a priority for translational research. If validated, especially with predictive capacity (as with the Aß classification for KPD), they would fulfill an important unmet need in clinical practice (86).

Conceptual Frameworks of Diabetes Heterogeneity

Heterogeneity as Reflecting an Overlap Between Diabetes Types

An acknowledged model to accommodate variant forms of T2D is to consider them as “overlaps” with other diabetes types. The most established example of this is latent autoimmune diabetes in adults (LADA), which defines patients who display key characteristics of both T1D and T2D, such as presentation in middle-late adulthood, positivity for a single islet autoantibody, risk alleles for both T1D and T2D, and a faster rate of progression to insulin requirement than autoantibody-negative T2D patients (87, 88). The prevalence of LADA ranges from 6% to 12% depending on the T2D cohort studied (89-96). However, the scope and variety of the overlapping phenotypes may be much greater if T-cell reactivity to islet autoantigens (ie, cellular autoimmunity) rather than autoantibodies (ie, humoral autoimmunity) alone were used to define the autoimmune basis of beta cell dysfunction. In a representative subset of T2D participants in the Glycemic Reduction Approaches in Diabetes–A Comparative Effectiveness (GRADE) study, 41% of participants had evidence for cellular islet autoimmunity, while 13% had humoral islet autoimmunity, and only 5% had evidence for both (97). These differences in autoimmune processes, apart from revealing a much higher prevalence of islet autoimmunity than previously considered in T2D, had physiologic and clinical relevance as participants with cellular islet autoimmunity manifested decreased beta cell function and worse glycemic control (97). These data support the case that the overlap model is not a simple construct but encompasses heterogeneous pathophysiologic pathways and clinical syndromes of T2D.

Clinically accessible tools now exist to identify patients with phenotypes that lie within the T1D-T2D overlap. A genetic risk score (GRS) for T1D (98) that incorporates T1D-associated genetic risk variants can differentiate autoimmune T1D from monogenic diabetes among adults and young adults (58, 99). A shortcoming of this GRS is its use of gene variants predominant in persons of European ancestry, so recently others, such as an African ancestry-specific GRS, have been developed (100). The predictive value of these GRSs makes them useful for diagnosis, prognosis and screening of patients’ family members (101). Studies are ongoing to test the role of GRS in classifying pediatric diabetes.

Phenotypic heterogeneity in T1D and KPD has been dissected using markers of insulin secretion, and this approach may also help to place the phenotypic heterogeneity in T2D on an etiological footing. In T1D, clinical trials of immune modulation to preserve residual beta cell function may have been confounded by pathogenic heterogeneity not well reflected by the current diagnostic tests. A measure of insulin secretion may be built into patient characterization at the time of diagnosis. For example, Index60 combines glucose and C-peptide values measured during an oral glucose tolerance test (102). In autoantibody-positive individuals with dysglycemia, those in the upper quartile of the Index60 distribution were significantly younger, with lower BMI and higher frequency of autoantibody positivity compared with those in the lower quartile (103), suggesting that the latter group may have a different pathogenesis, not adequately addressed by immunomodulatory therapies. Tools like Index60 and GRS may be effective at resolving heterogeneity within the T1D/T2D overlap.

The Palette and Threshold Models of Heterogeneity

Diabetes may result from multiple possible mechanisms that combine and interact to disrupt metabolic homeostasis. These mechanisms may vary by individual and, within the same individual, as time goes by. This has led to the concept that genetic mechanisms that are typically associated with either T1D or T2D may operate in the manner of a palette, ie, with varying combinations in different individuals to generate a range of phenotypes (104-106). A parallel concept lies in the “threshold hypothesis” (107) wherein interaction of both genetic and environmental factors may force an individual’s metabolic pathways to exceed a normal threshold and progress to diabetes. The palette model and the threshold hypothesis may be integrated to explain a wide, combinatorial variety of diabetes phenotypes (108). For instance, excess BMI combines with autoantibody positivity in both children (109) and adults (110), with different thresholds by sex and age, to elevate the risk of T1D and modify the rate of progression of islet autoimmunity (111). As another example, T2D-associated genetic variants in the TCF7L2 locus are more frequent in both children and adults with new-onset autoimmune T1D who carry a single islet autoantibody (a marker of only mild islet autoimmunity) or lack T1D-associated HLA genotypes (112-114). In these individuals, diabetes results from the combination of relatively slow progression of autoimmune islet destruction and additional diabetogenic mechanisms that are typically associated with T2D, resulting in atypical T1D characteristics (115, 116). This conceptual model shows that diabetes heterogeneity at the individual level requires recognition of multiple dynamic pathophysiological processes in the same person. While it does not readily evolve into an etiological classification, it provides useful guidance in clinical and research practice, such as including BMI as an important predictor for progression from preclinical to clinical T1D (117), using of metformin in pediatric T1D to improve insulin sensitivity (118), and recognition that islet autoantibody positivity accelerates progression to insulin dependence in T2D (50).

The Spectrum Model of Heterogeneity

As described in an earlier review by one of us (A.B.), variant forms of diabetes could fall into a spectrum of diseases or syndromes between the 2 poles of “typical” autoimmune T1D and “typical” T2D (119). The spectrum includes the currently known, etiological forms of atypical diabetes, including monogenic diabetes and the atypical subgroups of KPD, as well as LADA and T2D associated with T-cell mediated islet autoimmunity. The spectrum would also accommodate a range of currently unknown atypical forms of diabetes. Approaches to identify these unknown forms would require a concerted and dedicated effort beginning with selection of patients (from specialty clinics, registries of monogenic, lipodystrophic, syndromic, and KPD, cohorts of relatively undifferentiated “T2D” patients enrolled in clinical trials, as well as from patient self-referral or clinician referral), moving through a process to identify variant T2D phenotypes by clinical and biochemical criteria (eg, early or pre-pubertal onset, leanness, Mendelian inheritance of the trait, absence of autoantibodies in the face of severely deficient insulin production) and employing the full technological force of comprehensive “omics” analysis and physiological testing on the patients and their biosamples to delineate the pathophysiologic mechanisms. The outcome would be a clustered compendium of T2D variants classified by etiologic mechanisms. Such is the approach and goal of the Rare and Atypical Diabetes Network (RADIANT) (see below).

Future Directions Toward an Etiological Classification of T2D

To prepare the ground for a new, etiological classification of T2D, research is needed to sharpen our understanding of the pathogenic basis of the phenotypically variant forms of the condition. If the study of atypical forms of T2D could open windows into distinct pathophysiologic pathways leading to T2D, then comprehensive strategies to identify patients and families with these syndromes should be scientific priorities for clinical investigation. One of the challenges in this effort is the potential for wide variation in genetic contributions to pathogenesis within each form. If the patients were identified based on early onset of diabetes without evidence of islet autoimmunity, a Mendelian pattern of inheritance, or syndromic features that have failed to produce a clear diagnosis despite standard clinical evaluations (including MODY genetic testing), the cause of the diabetes might be pinpointed to novel, rare variants in 1 or 2 genes in the nuclear or mitochondrial genome that have a strong phenotypic effect. At specialized centers to which such patients may be referred, genetic testing is often recommended. The patients and their families are enrolled in genetics registries, and their data made available to researchers. If the phenotype is more complex, including late onset of diabetes and an uninformative pedigree in adult patients, then rare gene variants are less likely to contribute significantly to the etiology. In these cases, complex genetics, epigenetics, environmental factors, or combinations therein may underlie the pathogenesis. Identifying atypical forms within this group requires different methods that account for this complexity. Electronic searches and clustering analyses of extensive datasets from very large numbers of diabetes patients may discern discrete forms of atypical diabetes based on both genetic and phenotypic data.

Multiple approaches could help identify and define forms of atypical diabetes, investigations of which are likely to uncover pathogenic mechanisms relevant to more common forms of T2D. These include:

Population Diabetes Cohort/Diabetes Study Clinic Approach

Existing diabetes cohorts, registries, and studies have large longitudinal databases, stored biosamples of fasting serum, DNA, or peripheral blood mononuclear cells, broad geographic spread, and a wide range of age and ethnic groups likely enriched in atypical forms of diabetes. Unique study designs could further enrich some cohorts in patients with atypical forms of T2D—for example, the SEARCH cohort is exceptionally large and diverse and includes every form of pediatric and adolescent diabetes from across the United States. The TODAY cohort focused on new-onset “T2D” in children, which is likely enriched in atypical forms of higher complexity. Cohorts of recent-onset diabetic patients from North and South India include different regional ethnic populations who likely harbor unique forms of “lean” T2D. “Lean T2D” has proven, even with clinical and biochemical measures, to aggregate into atypical subgroups, such as ketosis-resistant diabetes of youth (27, 28) and the more recently described insulin-resistant obese diabetes (IROD) and combined insulin-resistant and deficient diabetes (CIRDD) (120). A filtering process, such as has been developed by the RADIANT study (see below), could serve to cull out from these databases persons with atypical features. These processes could be adapted to perform similar searches among in electronic medical records (especially if natural language processing software are available) of clinics and health systems and very large national or international databases such as the Botnia project, UK Biobank, the US Department of Veterans Affairs’ Million Veterans Program, or the DPV cohort in Germany. More circumscribed existing registries for specific atypical forms of diabetes (monogenic diabetes registry at the University of Chicago, lipodystrophy registry at the University of Michigan, Ketosis-Prone Diabetes registry at Baylor College of Medicine) are smaller but have potential for a higher yield of such patients.

Genetics Registries Approach

Patients referred to clinical geneticists and genetic testing laboratories are a potentially rich source to mine for patients with atypical T2D. A review of patient phenotypes referred to the Baylor College of Medicine clinical and research genetics clinics (comprising the Clinical Genetics laboratory, the mitochondrial genetics laboratory, the nuclear whole-exome sequencing laboratory, and the Center for Mendelian Genetics) illustrates this possibility. In the >100 000 cases referred to these laboratories collectively for genetic testing, approximately 800 were referred for phenotypic characteristics that included diabetes. Diabetes patients in the nuclear exome sequencing laboratory and mitochondrial sequencing laboratory data sets were reviewed for the following atypical criteria: age at testing (early-onset or later-onset with features atypical of T2D), strong family history (multigenerational), syndromic features, existing variant interpretations and genotype-phenotype associations, and expert review by clinical diabetologists (121). Suspicion for atypical diabetes was raised in 31 of the 287 nuclear exome sequencing referrals and 87 of the 115 mitochondrial DNA sequencing referrals. Since genetic testing is often ordered for atypical presentation of a common disease, clinical genetic diagnostic lab data are a rich source for discovery of patients with atypical forms of diabetes.

Machine Learning Approaches for Subgrouping T2D

Machine learning models and statistical techniques offer exciting opportunities to uncover heterogeneous forms of diabetes within T2D by clustering individuals with similar biochemical, genetic, and clinical characteristics. The limitations are largely related to the nature and coding of the stored data and the selection of variables to circumscribe the phenotypes.

Alqvist et al performed unsupervised hierarchical cluster analysis on clinical data from ~9000 Swedish adult patients with newly diagnosed diabetes (122), using 6 variables collected systematically over many years in longitudinal Scandinavian registries. The variables were GAD65-Ab, age at diagnosis, BMI, and measures of insulin sensitivity and insulin secretion. The analysis uncovered 5 clusters of T2D: severe autoimmune diabetes (SAID), severe insulin-dependent diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD) and mild age-related diabetes (MARD). The clinical relevance of the clusters (each representing a distinctive syndrome of T2D) was evident in their natural histories and risks of complications. This clustering analysis demonstrates that the heterogeneity of T2D is susceptible to useful, predictive subclassification based on key differences in phenotypic markers such as insulin sensitivity, metabolic syndrome markers, BMI, and age of onset. SAID and SIDD patients had the poorest glycemic control and became insulin dependent over the shortest period. SIDD patients had higher frequency of diabetic retinopathy than other clusters. SIRD patients had the highest prevalence of nonalcoholic fatty liver disease, and, despite a relatively low HbA1c, developed diabetic kidney disease at the fastest rate. Limited analysis of T2D risk alleles indicated that SIRD might be genetically distinct from other forms of T2D; the TCF7L2 rs7903146 variant that is strongly associated with T2D overall was associated with SIDD, MOD, and MARD but not with SIRD, and the T2D genetic risk score was associated with all the clusters but not with SIRD.

Udler et al used nonnegative matrix factorization (123) in over 17 000 patients with T2D genetic variants and 47 traits from genome-wide association studies of T2D cohorts (64). They also identified 5 clusters, each differentially enriched for tissue-specific regulators of gene expression (enhancers and promoters). Recently a large T2D cohort database in South India was utilized to identify 4 clusters prone to young-onset T2D (119). Two clusters (SIDD and MARD) were similar to those identified in the Scandinavian study. Two new clusters unique to this ethnic group were identified, denoted IROD and CIRDD. Bancks et al (124) analyzed T2D subgroups in participants in the Look AHEAD weight loss study and observed that intensive lifestyle intervention was associated with higher risk of cardiovascular disease in the subgroup with poorest glucose control. Wagner et al recently proposed that clusters can be identified even in the prediabetic period, drawing upon variables that would not be confounded by secondary effects of gluco- and lipotoxicity (125). In adults at risk for T2D, these authors identified 6 clusters with different risk of progression to T2D despite, in some cases, similar glycemic levels. Thus, although there is not a straight path from current cluster analysis to an etiological classification scheme for T2D, the clusters permit clinicians to refine screening approaches and provide better informed therapy to different groups of individuals with T2D.

Another limitation of extant clustering / bioinformatics analyses arises from the inclusion of more or less undifferentiated forms of T2D in the cohorts studied. Rare or atypical forms of diabetes may not find a taxonomic home using this approach. There is an ongoing attempt to first filter a T2D population for atypical features, and then perform cluster analysis in those persons, in a Mexican American cohort in South Texas. The approach is to manually identify patients who do not fit the conventional criteria of T1D or T2D and then use a machine learning procedure with a filtering process utilizing variables such as lean habitus, islet autoantibody, age of onset, insulin dependence, and C-peptide, followed by a robust spatial clustering algorithm (Parikh H et al, unpublished). Validation of this and similar approaches in larger and more geographically and ethnically varied cohorts are needed. Electronic medical records can be leveraged for the purpose of identifying atypical diabetes but this requires the development of algorithms that are adapted to the specific target population (126). Overall, a variety of approaches will be needed to determine the true prevalence of atypical diabetes and their forms. This is a goal of RADIANT, a multicenter collaborative project funded by the National Institutes of Health, launched in October 2020. This innovative project should provide detailed insight into the pathophysiology of a wide range of atypical forms of diabetes, with the promise through clustering analyses of these mechanistically defined phenotypes to better understand and develop an etiological classification of T2D.

Glossary

Abbreviations

BMI

body mass index

CIRDD

combined insulin-resistant and deficient diabetes

DKA

diabetic ketoacidosis

GRS

genetic risk score

HbA1c

glycated hemoglobin A1c

IROD

insulin resistant obese diabetes

KPD

ketosis-prone diabetes

LADA

latent autoimmune diabetes in adults

MARD

mild age-related diabetes

MOD

mild obesity-related diabetes

MODY

maturity onset diabetes of the young

SAID

severe autoimmune diabetes

SIDD

severe insulin-dependent diabetes

SIRD

severe insulin-resistant diabetes

T1D

type 1 diabetes

T2D

type 2 diabetes

Additional Information

Disclosures: The authors affirm no conflicts of interest related to the contents of this paper.

Data Availability

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.