Single-Cell Analysis of CD4 T Cells in Type 1 Diabetes: From Mouse to Man, How to Perform Mechanistic Studies

Type 1 diabetes is a CD4 T cell–mediated autoimmune disease that results in the destruction of the pancreatic β-cells that produce insulin. Type 1 diabetes is also the poster child of autoimmune diseases linked to genetic susceptibility. While more than 40 genes have been described in this inherited landscape (1), the association that stands out is with the HLA class II locus on chromosome 6 with a P value of 10⁻¹²³, suggesting not only influence but causality. This notion is reinforced by the fact that the fine mapping of this linkage with the HLA class II region identifies a single polymorphism at position 57 of the HLA-DQβ chain as being responsible for most of the association (2,3), while non-HLA loci and genes likely play a role in influencing the progression to disease onset (4). Only two common haplotypes of HLA-DQ carry a nonaspartic acid at position β57, unlike any other HLA-DR, -DQ, or -DP molecules, HLA-DQ2 (HLA-DQβ0201), and HLA-DQ8 (HLA-DQβ0302) (Fig. 1). This genetic terrain allows multiple environmental factors to emerge as disease triggers (5).

At first glance, an association of a CD4 T cell–mediated disease with HLA class II gene products, whose function is to present peptides to CD4 T cells, appears easily explainable. Over the years, the most obvious, nonexclusive theories have been tested: instability and poor peptide binding of diabetogenic HLA class II molecules (6), unique peptide repertoire of the same molecules (7), T cells focused on the recognition of HLA-DQβ57 (8), failed thymic selection of autoreactive T cells (9), and abnormal T-cell binding to autoimmune peptide–MHC complexes (10). While all of those might bear truth and give some level of understanding of what the β57 residue might do, none could formally associate the mutation to a molecular mechanism leading to diabetes. The closest one to explaining the association of the same mutation with a disease was in the context of celiac disease, where the same HLA-DQ molecules are strongly predisposing to onset and also promote a frequent association with type 1 diabetes (11). In this instance, it was shown in transgenic HLA-DQ8 mice, and for some human CD4 T-cell clones, that gliadin peptides were recognized by T-cell receptors (TCRs) bearing a negatively charged residue in the first segment of the CDR3β loop (12). Interestingly, most native gliadin peptides are glutamine rich and neutral unless deamidated by tissue transglutaminase (13) but can still be presented by HLA-DQ2 and -DQ8 molecules, which usually prefer peptides with negatively charged amino acids in their C-termini (7). The acidic residues occupy the P9 pocket of MHC and compensate for the loss of the aspartic acid at position β57, which is an integral part of the outside wall of this MHC pocket (14) (Fig. 1). The need for side chains anchoring into pockets for peptide binding as we see for MHC class I and HLA-DR molecules has been lost for HLA-DQ molecules (14,15), allowing more diverse and promiscuous binding. As a consequence of this mode of binding, peptide repertoire is much broader for HLA-DQ than for HLA-DR molecules, but the affinities of peptide binding are much lower, often in the mid to high micromolar range (12). This biophysical detail is often overlooked, although it informs us of two important convergent features of autoimmunity: first, because there is a threshold to activate T cells, low affinity peptides must be abundant to compensate for short binding half-lives; second, homozygosity of the susceptibility HLA genes, as often observed in autoimmunity, is essential to increase cell surface expression of the diabetogenic peptide–MHC complexes. In type 1 diabetes, this latter issue is further compounded by the fact that HLA-DQ2/DQ8 heterozygotes are also at a higher risk of disease due to the expression of transdimers in which the β57 position always lacks the normal aspartic acid (16). In an effort to understand how neutral peptides bound to HLA-DQ8 could select TCRs with a negatively charged residue in their CDR3β loop, we extended the observation to the mouse model and a neutral peptide from hen egg lysozyme in order to gain structural information (8). Indeed, the very low affinity of gliadin peptides for HLA-DQ8 precluded structural studies and also the making of T cell–detecting reagents such as MHC tetramers. The results from our structural studies were surprising; while we expected the CDR3β negative charge to sit in proximity to the positive patch of the β57 residue and establish a salt bridge, shifting the TCR over the COOH-terminal part of the p-MHC complex, the TCR was found in a normal diagonal position putting the CDR3β far away from β57 (8). However, biophysical studies demonstrated that the complementation of charges between TCR and p-MHC were operating through Coulombic interactions, a phenomenon that allows surfaces of opposite charges to enhance dehydration and increase on-rates of binding interaction.

In any case, this deep knowledge of I-A^g7 and HLA-DQ molecules could not establish a direct link between MHC class II, position β57, and type 1 diabetes. In the absence of a rodent model and MHC tetramers, the celiac disease observation could not be tested further, while the absence of antigen-specific reagents was impeding our studies in mice. In addition, studying type 1 diabetes offers additional challenges that were insurmountable for decades both in mouse and man. The two most challenging were the size of the organ that produces insulin (<1.5 g of tissue in a human) and the asynchrony of the lesions across the ∼1 million islets. These numbers and the low efficiency of the autoimmune process, which takes on average 15% of a life span to reach completion (5 years in humans, 15 weeks in mice), likely translate to a very small number of anti–β-cell–specific CD4 T cells locally and in circulation. This situation is very consequential, making the diagnosis of the preclinical phase of disease extremely difficult and mechanistic studies very challenging.

As often in science, advancements in technology open access to the next level of understanding. About a decade ago, microfluidic systems allowed the isolation of single cells in microchambers that can be used as reaction vessels. Concomitant with the rise of next-generation sequencing, this new engineering gave birth to the rapidly expanding world of single-cell technologies. We can now probe with single-cell resolution genetic differences, differential gene expression, genome-wide epigenetic modifications, and large sets of unique proteins. Most importantly, these approaches have allowed us to interrogate very small numbers of cells as we expect in biopsies or circulating blood. As it stands today, single-cell technologies have allowed us to refine our translational studies from mouse to man and approach mechanistic understanding of disease, and, most importantly, might open the possibility of an early preclinical diagnosis and the monitoring of treatment in man.

Bulk analysis techniques have been the primary method utilized thus far to understand autoimmunity in type 1 diabetes. While these have been invaluable in developing our basic understanding of the disease, they are built on the principle that populations are homogenous and static, and therefore readings of a particular marker can be averaged over the population (17). Unlike bulk techniques, single-cell techniques are powered to account for variability. There are excellent reviews that discuss in detail the application of single-cell approaches to address cellular and biological identity (17–20). All of these techniques rely on large data sets with computational and statistical analysis. While instrument manufacturers provide basic analysis packages, thorough analysis and the production of informative figures requires the involvement of computational biologists. This bottleneck seriously limits the access to these technologies and also creates a landscape where no one way is yet accepted as the standard.

Single-cell genomics are the most used single-cell techniques; they can interrogate both genomic DNA and the transcriptome (exome). These approaches have primarily been used to describe cellular diversity and representation in tissues for two main applications: embryogenesis/development (21) and cancer. While sequencing techniques are highly accurate, “depth of interrogation” and “coverage” of the genome or transcriptome are the two limiting factors that should be kept in mind. These two parameters vary dramatically based on the method used for isolating single cells (microfluidic vs. droplet isolation) and for sequencing (full length vs. 3′ counting methods) (22). Very recent technologies using the principle of fluorescent in situ RNA hybridization in a multiplex and repetitive format on tissues have added spatial information to exome data (23,24). Similarly, epigenetic information would nicely complement transcriptomics analysis, but single-cell epigenomics techniques remain difficult and limited (25).

In all cases, regardless of the depth of the single-cell genomics interrogation, profiles of RNA expression must be compared with the levels of proteins made by the cells of interest.

For blood cells and cells dissociated from tissues, this validation step at single-cell resolution is dominated by multiparametric flow cytometry analysis and variants of it such as mass cytometry. All are currently limited to 20–50 parameters, a dimensionality that is largely sufficient to separate immune cell populations for which differential markers have been very well described. For tissue sections, automated immunofluorescence techniques with successive rounds of hybridization-photobleaching or imaging mass cytometry achieve similar levels of resolution. The immune infiltrate in pancreatic islets from healthy donors, patients with recent-onset type 1 diabetes, and patients with established diabetes was recently published using these approaches (26).

While these cutting-edge technologies allow the description of phenomenology at its highest granularity and might generate interesting hypotheses, mechanistic immunological studies require one more level of technological advance to address the central theme of immune recognition: antigen specificity.

Our ability to interrogate specificity of T cells with MHC tetramers dates from 1996 (27). While relatively successful for MHC class I and the detection of CD8 T cells in patients (28,29), the use of tetramers for isolating CD4 T cells has been difficult and has not moved yet to the clinic. The main reasons for this situation are inherent to the MHC class II molecules associated with type 1 diabetes, HLA-DQ in humans and I-A in mice. This class of molecules, unlike MHC class I or the ancestral MHC class II molecules, HLA-DR and I-E, do not necessarily bind and select peptides using anchor side chains in binding pockets, but they can accommodate many more peptides by using an anchorless mode of binding (15). The consequences of this divergence are a promiscuity of binding, the ability to display multiple registers of a single peptide, and a much larger peptide repertoire. These characteristics are advantageous to fight infection, but they inherently impede an efficient deletion of autoreactive cells.

When it comes to expressing recombinant HLA-DQ/I-A molecules for functional and/or structural studies, low-affinity peptide binding results in poor protein stability and often a mix of different registers, while the inefficient MHC α and β chain dimerization further limits expression. If the latter issue can be addressed by leucine zippers forced pairing (30), low peptide affinity, and the display of single registers remain without universal answers. As a result, very few good reagents have been produced for mouse and human studies of CD4 T cells in autoimmunity; none have been brought to the clinic. Finally, an appreciable number of patients do not carry diabetogenic HLA genes and cannot be interrogated with MHC tetramers; therefore, detection of anti–β-cell CD4 T cells using non-HLA tetramer–based technologies must be developed.

The isolation of antigen-specific B cells using tetramerized antigens is a more recent approach that has proven useful in following anti-infectious responses (31,32) but remains confidential in autoimmunity (33).

In both instances, the tetramer isolation of single cells allows the determination of TCR and B-cell receptor chain sequences, their pairing, and most importantly their reexpression in surrogate cells and/or as recombinant soluble molecules for functional and mechanistic studies.

The debate about the value of the NOD mouse model for human type 1 diabetes is still current with no signs that either of the two sides will compromise anytime soon. In our opinion, arguments should always be ranked in biological order. In this respect, the confounding similarities of the genetics of the disease between mouse and man are at the top, especially the common mutation at position 57 of the β chain of MHC class II. The next arguments on the list are clinical and pathological: spontaneous onset, prolonged preclinical period (15% of a lifetime), allochronicity of the lesions, β-cell exclusivity, and similar comorbidity (thyroid, salivary glands). Therefore, we would argue that the great divide between our colleagues is more a methodological issue than an issue with the model itself. However, to bridge mouse and man, methodology is difficult and faces obvious challenges: first, in the two species the work is usually done at different phases of disease—preclinical in the mouse, established disease in man. We rarely keep mice under insulinotherapy to study the post-destruction phase of the disease, whereas in humans, access to preclinical samples is limited to the small TrialNet cohort and newly diagnosed cases remain difficult to obtain in large numbers. Beyond these fundamental discrepancies, the tissues that we sample are also dramatically different. Blood is the only readily available steady source of immune cells in humans but is challenging to survey in mice because of volume; conversely, internal tissues such as the spleen, lymph nodes, and infiltrating cells of the islets are easily accessible in mice but not in humans. To cap this difficult situation, informative samples from recently diagnosed patients, accessible through the Network for Pancreatic Organ Donors with Diabetes, are extremely rare and coveted by many laboratories. Being conscious of these important issues allowed us to redesign our studies to compare similar stages of disease, sample blood from single mice, use identical markers and reagents, and use similar protocols. Single-cell approaches are remarkably suited to this effort.

As mentioned, the top advantage of single-cell techniques is to allow the interrogation of very few cells. In the mouse, access to cell numbers was until now addressed by pooling together organs from a large number of animals. Indeed, the frequencies of autoreactive T cells in most organs is always low, from 0.05% to 0.2% for CD4 T cells, numbers that preclude any characterization at the very early preclinical stage of disease when only a few cells are present in only a few islets. Now the recovery of 100–200 cells from the 200–300 islets from a single mouse is exploitable and reaches statistical power to determine the diversity of the CD4⁺ T-cell population, the gene expression, and idiotypic receptor sequences of each cell in it. Similarly, we can isolate and study 25–50 CD4 anti-insulin T cells that we isolate from 0.5–1 mL of blood with MHC tetramers (frequencies in this tissue vary widely from 0.1% to 0.01% both in mice and humans). Because we know the biology of T cells in such detail, the examination of 96 genes by single-cell quantitative PCR allows us, based on gene module examination, to categorize subsets of CD4 T cells (Th1, Th2, Th17, Treg) while simultaneously evaluating their state of activation (Fig. 2) (34). Although this exercise of parsing cellular diversity remains very descriptive, it provides a very deep interrogation of the functional state of each cell and can easily indicate whether the cell is dormant, activated by cytokines, and/or activated through its TCR. Using this approach in NOD mice, we have been able to not only characterize CD4 T cell populations in thymus, spleen, pancreatic lymph nodes, and islets, but to determine the gene expression signature in each organ (35). The main conclusion of this examination was that, based on 96 “T-cell genes,” we could easily identify the residence of each T cell. While peripheral T cells (spleen as well as peripheral lymph nodes) expressed nearly no signaling genes, T cells in the pancreatic lymph nodes showed signs of cytokine activation but no stigma of TCR engagement, and intraislet cells displayed the bona fide signature of TCR triggering and activation. This first layer of examination indicated that the antigen-specific process was mostly limited to the islet, not the draining lymph node. The second layer of information that was gained was about the nature of the early antigen, insulin. MHC tetramers displaying insulin peptides showed that at 6 weeks of age, about two-thirds of all intraislet CD4 T cells recognized a single epitope of the B9-23 segment of insulin, B12-20, and that at 12 weeks of age this epitope dominance was gone. This series of experiments firmly established that insulin was the main initiator and driver antigen of anti–β-cell autoimmunity (35). Aside from this very important piece of information, our single-cell analysis was capable of supporting two more mechanistic studies. The first one was about the mode of T-cell recognition that allows breakage of T-cell tolerance and that we described earlier as the “P9 switch,” in which diabetogenic MHC molecules displaying peptides with neutral residues at P9 are recognized by TCRs bearing an acidic amino acid in the N-terminal segment of their CDR3β (8,12). The P9 switch was demonstrated by sequencing and reexpressing TCRs isolated with I-A^g7 B12-20 insulin tetramers and by showing their mode of activation in vitro. Most importantly, while the P9 switch dominated the anti-insulin response at 6 weeks of age, it was nearly gone by 12 weeks, indicating that its role is essential in the initiation of disease, not its progression. Also, this P9 switch–driven anti-insulin response was eliminated from mice in which the β57 mutation was corrected to an aspartic acid, and these mice were resistant to disease. The second mechanistic study that the single-cell approach gave access to was the examination of whether anti-islet CD4 T cells recirculated and were detectable in peripheral blood. To access this information, we first profiled intraislet CD4 T cells and then examined blood for cells with similar phenotypes and TCRs, tracing once again specificity with insulin I-A^g7 tetramers. Our early studies, both in mice and humans, are encouraging but still preliminary (Fig. 2). Anti-insulin cells were found in peripheral blood at a very early stage (3 weeks of age in mice), but it appeared that their state of activation varied greatly. If activation state correlates with the progression of autoimmunity, a single-cell profiling of gene expression of anti-insulin CD4 T cells from blood might constitute a very early diagnostic test of autoimmunity in which not numbers but activation status is evaluated (Fig. 2). The confrontation of this approach with the detection of anti-islet antibodies in at-risk patients will answer this very important question.

Other groups have made progress in the diagnostic isolation of circulating autoreactive anti–β-cell T cells and/or T-cell populations from the peripheral blood of at-risk or just diagnosed patients. In one study, multiparametric flow cytometry with HLA class I tetramers was used to isolate anti-ZnT8–specific CD8 T cells, whereas in the other flow cytometry was used to isolate CD4 T cells and examine those by single-cell RNA sequencing (29,36). It is likely that the technological progress for single-cell analysis and the examination of a larger cohort of patients will deliver a reliable diagnosis within the next 5 years. Of course, it is in at-risk populations that these approaches need to demonstrate their value; in this respect, the TrialNet cohort is of the greatest importance and a resource that each type 1 diabetes researcher should appreciate and advocate for. The added interest of assessing autoimmunity from peripheral blood is twofold: one is to establish precedent for other autoimmune diseases and generalize the approach, and the other one is to appreciate in real time or near real time the progression of disease and the impact of therapies on its course. Indeed, immunologically based interventions at diagnosis or in at-risk patients are currently monitored on residual β-cell mass, not on the driving autoimmune process itself as it should be (37,38).

It is also likely that a single-cell approach will add a deeper understanding of mechanisms of disease. If our animal studies translate to humans, TCR sequencing of single CD4 T cells could improve the accuracy and timing of diagnosis by showing T cells capable of recognition through a P9 switch; these cells should be the earliest to appear. We should also be able to confirm that the anti-insulin response comes first and drives disease progression. Deeper RNA sequencing and epigenome measurements should also shed light on the mechanisms leading to the breakage of tolerance. This issue is critical to address if we want to be able to understand environmental factors at the molecular level and, practically speaking, to parse the at-risk population into “at high risk” and “at low risk” of progression. In this process, it is also highly probable that we might uncover some explanation for the increasing heterogeneity in disease presentation and evolution.

This perspective would not be complete if we were not addressing some financial issues related to single-cell analysis. The lofty goal of diagnosing type 1 diabetes in its preclinical phase and monitoring immune interventions aimed at protecting β-cells is hindered by the cost of single-cell technologies and is a major issue. Between the cost of goods, the cost of processing (cell sorting on a cytometer, preparation of DNA libraries, sequencing, deconvolution of metadata), and the cost for highly qualified personnel, each mouse and each patient that are examined represent an investment of $10,000–$12,000. As the cost of DNA sequencing is now near bottom, the deployment of single-cell technologies in the clinic will require a complete change of the business model used by the private sector that provides equipment and consumables. It is understood that the current effort was intended for basic research, but the future is in the clinic with thousands of patients who will need regular evaluation over a lifetime.

Funding. This work is supported by National Institute of Diabetes and Digestive and Kidney Diseases grant 1R01DK117138 to J.P., M.G., and L.T. S.S. is a TL1 scholar supported by a National Institutes of Health Clinical and Translational Science Award issued to the Scripps Translational Science Institute (UL1TR002550, TL1TR002551). B.A. was a KL2 scholar (KL2TR001112) from the same Clinical and Translational Science Award.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.