Targeting of Pancreatic Glia in Type 1 Diabetes

NOD/LtJ, NOD.129P2(B6)-B2m^tm1Unc/J (NOD.β2m^−/−), and C57BL/6J (B6) mice were purchased from The Jackson Laboratory (Bar Harbor, ME). NOD AI4 T-cell receptor (TCR) transgenics were donated by Dr. David Serreze (The Jackson Laboratory). Mice were maintained in our conventional facility under approved protocols. Glucosuria was used to screen for diabetes (Diastix; Bayer, Toronto, ON, Canada) and confirmed by blood glucose measurements (>13.8 mmol/l) (OneTouch Ultrasmart; Lifescan Canada, Burnaby, BC, Canada).

A GFAP promoter expression cassette was donated by Dr. M. Brenner (12). NOD β2-microglobulin (β2m) cDNA was generated by RT-PCR from spleen using β2mBamF 5′-CGCGGATCCATGGCTCGCTCGGTG-3′ and β2mBamR 5′-CGCGGATCCTCACATGTCTCGATC-3′). Six founder mice were generated after ∼1 pL injection into hyperovulated, fertilized NOD ova, crossed to NOD.β2m^−/− and screened using the genotyping protocol B2m^tm1Unc (The Jackson Laboratories). Western blots used goat α-β2m (Santa Cruz Biotechnology, Santa Cruz, CA) and mouse α-goat horseradish peroxidase (The Jackson Laboratory).

Splenocytes (10⁷) from pooled spontaneously diabetic NOD females were injected intravenously in 100 μl PBS to NOD males, NOD.β2m^−/− females, or GFAP-β2m Tg female recipients irradiated 650 cGy, 24 h before transfer.

Mouse spleens were dissociated and red cells lysed by hypotonic shock. A total of 4.5 × 10⁵ splenocytes/well were cultured in adoptive immunotherapy media (AIM)-V media (Gibco) plus 5% fetal bovine serum (Gibco) containing antigen in 96-well flat-bottom plates (BD). After 72 h, cultures were pulsed (1 μCi [3H]thymidine/18 h) and counted by liquid scintillation. Supernatants were collected after 48 h for cytokine enzyme-linked immunosorbent assay following the manufacturer's instructions (BD). Patients with new-onset type 1 diabetes or first-degree relatives (FDRs) of type 1 diabetic patients were invited to participate under approved protocols and after signing informed consent at the Pittsburgh Children's Hospital. Heparinized peripheral blood was sent to the Toronto assay lab (H.-M.D) by overnight courier, and mononuclear cells were prepared and cultured as described (9). To normalize results from human and mouse experiments, some data are presented as stimulation index (SI) (SI = antigen cpm/media cpm). K^d inhibition used the SF1–1.1 monoclonal antibody (BD).

Splenocytes were stained with CD3-FITC, CD8-APC, and CD4-PE (BD Pharmingen, Mississauga, ON, Canada). Live events were collected based on forward-and side-scatter profiles on a FACScan flow cytometer (BD) and analyzed using FlowJo software (Stanford University).

Eight-week-old NOD females were immunized with 5 μg GFAP, emulsified in incomplete Freund's adjuvant (IFA) (Sigma) in the hind footpads (50 μl/foot), and boosted intraperitoneally 10 days later with 5 μg GFAP IFA. Spleens were harvested 10 days after the seecond injection for in vitro T-cell proliferation assay.

Overlapping 15-mer peptides with an offset of five amino acids were purchased from Mimotope (Clayton, Australia). The mus musculus sequence of GFAP (418 amino acids) from the NCBI protein database (Accession NP_034407) was used (13). Peptides were dissolved in PBS (Gibco) and for insoluble peptides, DMSO or acetic acid was added. GFAP 79–87, 96–110, 116–130, 216–230 and 253–261 were synthesized at >90% purity (Alpha Diagnostics International, San Antonio, TX). Other antigens included human GFAP (Calbiochem, San Diego, CA) and ovalbumin (Sigma). Human T-cell assay antigens have previously been described (5).

Frozen pancreas sections were fixed in 1% paraformaldehyde and blocked with 5% normal donkey serum (Jackson ImmunoResearch, Westgrove, PA). Primary antibodies: rabbit α-GFAP (Signet Pathology Systems, Dedham, MA), guinea pig α-insulin (DAKO, Carpinteria, CA), goat α-CD3 (Santa Cruz), goat α-glucagon (Santa Cruz), mouse α-β2m (Santa Cruz), biotinylated α-I-A^b (BD), and mouse anti-CD8-biotin (BD). Secondary reagents included donkey α-rabbit fluorescein isothiocyanate, donkey α-goat Cy5, donkey α–guinea pig Cy5, donkey α-goat Cy5 (Jackson ImmunoResearch), donkey α-goat AlexaFluor 546, Strepavidin AlexaFluor 633 (Molecular Probes, Eugene, OR), and propidium iodide or DAPI (Sigma). Immunofluorescence was visualized by confocal microscopy (Zeiss).

Islets were isolated using standard protocols (14). For pSC cultures, standard islet isolations were performed, but islets were cultured in laminin-coated dishes (Sigma) using Dulbecco's modified Eagle's medium (DMEM) with 2.5% fetal bovine serum, 0.2% bovine pituitary extract, and forskolin (Sigma) (15). Media were changed twice weekly. At ∼80% confluency, islets were handpicked from the culture followed by treatment with 10 mmol/l alloxan (Sigma) for 15 min consecutively over 2 days. At confluency, media was changed to 2.7 ml DMEM + 0.3 ml dissociation medium (10 ml DMEM + 250 mg Dispase [Gibco] and 50 mg collagenase type I [Sigma]) and incubated for 3 h at 37°C. A flame-narrowed Pasteur pipette was used to triturate the cells and, following two washes, seeded for cytotoxic T-lymphocyte (CTL) assay. Islets and pSC cultures were derived from either NOD.scid or 4-week-old B6 or NOD females.

Islets or pSC were seeded in 96-well flat-bottom plates (BD) at 4 × 10⁴ cells/well in AIM-V. pSCs were stimulated for 48 h with 100 units/ml recombinant mouse IFN-γ (Sigma) before 1 h labeling with ⁵¹Cr. Splenocytes or CD8⁺ T-cells (EasySep; Stem Cell Technologies, Vancouver, BC, Canada) in AIM-V were added at various effector:target cell ratios. After 8 h at 37°C, 70 μl supernatant was counted. RMA-S K^d targets were a gift of P. Santamaria (Calgary, AB, Canada) and used as previously described (16). RMA-S K^d cells were pulsed with 5 μg/ml peptide. Specific lysis was calculated as follows: (cpm sample release − cpm spontaneous release)/(cpm maximal release − cpm spontaneous release).

To date, pSC destruction has been inferred from immunohistochemistry of pancreas sections of NOD mice (9). To directly visualize pSC-directed autoimmune targeting, we generated pSC lines in vitro. Primary pSC cultures were derived from freshly isolated islets (Fig. 1A) by growth in selective media (15). Fresh islets contain the expected distribution of endocrine cells with insulin-stained β-cells (blue) in the center surrounded by glucagon-stained α-cells (red) (Fig. 1B). Only sparse numbers of islet-adherent GFAP⁺ pSC survive the stress of collagenase digestion and isolation procedures, which strip the pSC envelope (Fig. 1C). Variations in digestion time affected pSC yields dramatically. However, over the ensuing 3- to 5-week culture period, pSC proliferate extensively, generating lawns of GFAP⁺ glial cells (Fig. 1D–F). Growing pSCs generate a network and actively envelope islets, suggestive of a strong tropism between islets and pSC (Fig. 1D and E) (17). In the cultures used for subsequent experiments, contaminating islets were removed manually, followed by two consecutive treatments with alloxan. Remaining cells were >94% pSC as assessed by flow cytometry for p75 nerve growth factor receptor expression (data not shown) (15) and immunofluorescent staining for GFAP and insulin, with glucagon-negative, fibroblast-like cells as the minor contaminant.

GFAP-stained pSC cultures exhibited dendritic morphology reminiscent of astrocytes (Fig. 1G). To determine whether pSC expressed MHC class II, pSC cultures were treated with IFN-γ for 24 h and then observed by immunohistochemistry. MHC class II was undetectable in pSC cultures as expected (18), with added antigen-presenting cells serving as positive controls (Fig. 1H).

As NOD T-cells recognize pSC auto-antigens such as GFAP in proliferation assays (9), we asked whether pSC would be killed ex vivo by pSC-reactive T-cell pools. A spontaneous CTL assay using pSC targets was developed, initially with isolated NOD islets (Fig. 1I). Purified NOD pSC cultures were then generated and treated with IFN-γ (to upregulate MHC class I) before labeling and incubation with fresh splenocytes or negatively selected CD8⁺ T-cells from spontaneously diabetic NOD mice. As shown in Fig. 1J and L, ex vivo CTL activity toward IFN-γ treated pSC was detectable without prior CTL activation in culture or exogenous peptide pulsing. pSC lysis was significantly higher in the NOD (P < 0.03) than in the B6 control assays (Fig. 1J and L), although some background B6 killing was observed.

CTL-mediated lysis in this system appears to be dependent on MHC class I, as killing was negligible with NOD.β2m^−/−-derived pSC (Fig. 1K). That pSC destruction may be predominantly CD8⁺ T-cell dependent was corroborated by immunohistochemistry of spontaneously diabetic NOD BDC2.5 T-cell receptor transgenics. Pancreas sections in this CD4⁺ T-cell–mediated diabetes model revealed intact pSC mantles surrounding remaining α-cells (Fig. S1 [available in an online appendix at http://dx.doi.org/10.2337/db07-0226]). Thus, it appears that at least one mechanism of pSC destruction involves cognate CD8⁺ CTL.

To investigate the in vivo relevance of CTL reactivity toward pSC, we generated transgenic mice that limited CTL–β-cell contact to pSC. NOD.β2m^−/− mice do not develop spontaneous diabetes but do develop peri-islet T-cell aggregates >30 weeks of age without pSC damage (Fig. S2), further supporting the importance of CD8⁺ T-cells in pSC targeting. Importantly, NOD.β2m^−/− mice can become hyperglycemic several months after adoptive transfer with wild-type splenocytes from diabetic donors (19–21). To selectively restore β2m expression in Schwann cells, we used a well-characterized GFAP promoter (12) to drive NOD β2m expression (Fig. 2A). Injection of the transgene into fertilized NOD eggs yielded multiple founder lines (Fig. 2B), which were bred to NOD.β2m^−/− congenic stock (Fig. 2C). Homozygous β2m^−/− transgenics were then analyzed for β2m expression by Western blot and immunostaining in brain and pancreas (Fig. 2D–F).

Based on good breeding and nonleaky transgene expression, transgenic lines 9 and 14 were selected for further study, generating similar results. As expected, endogenous CD8⁺ T-cells were absent in spleen cells of GFAP-β2m transgenics (Fig. 3A–C), and they resisted development of spontaneous diabetes (Fig. 3D). In adoptive transfer experiments (22), splenocytes from diabetic females transferred hyperglycemia to irradiated wild-type NOD males within 5 weeks, while nontransgenic NOD.β2m^−/− mice were very slow to develop disease and in only a few animals (Fig. 3E). In contrast, GFAP-β2m mice developed diabetes at a significantly faster rate and higher incidence than nontransgenic littermates (P = 0.0016, log-rank test). Both founders gave similar results. Hyperglycemia in transgenic mice was associated with progressive autoimmune insulitis and destruction of pSC (Fig. 3F–I), with increased retention/recruitment of CD8⁺ T-cells to the islet locale (Fig. S3).

To ensure that disease acceleration was not due to MHC class I stabilization on β-cells from soluble β2m possibly derived from pSC, we employed a β-cell–specific TCR transgenic system. The AI4 TCR recognizes dystrophia myotonica kinase (23) and induces rapid spontaneous diabetes by 4 weeks of age (24). Compared with spontaneous NOD diabetes, AI4-mediated hyperglycemia occurs with only minor pSC disruption (Fig. 4A and B). Using a robust AI4 adoptive transfer protocol (25), rapid disease developed in wild-type recipients but not GFAP-β2m transgenics or NOD.β2m^−/− controls, excluding biologically relevant amounts of MHC class I stabilization on β-cells in our GFAP-β2m model (Fig. 4C).

To further characterize pSC autoimmunity, we sought to define NOD GFAP epitopes. The NOD MHC haplotype includes the common MHC class I alleles K^d and D^b and the rare MHC class II allele I-A^g7 and is I-E null (1). While search algorithms exist for the K^d and D^b alleles, none are available for the unique I-A^g7, although preferred peptide-binding motifs have been reported (26,27). To identify MHC class II GFAP epitopes, an overlapping 15-mer peptide map of the entire murine GFAP sequence was synthesized. Proliferative splenic responses to GFAP peptides were measured 10 days following immunization with whole GFAP protein in IFA. Pooled results from 10 individual assays are shown in Fig. S4, with top-responding peptides in Table S1. We selected IFA over CFA to avoid a strong TLR-mediated immune response and more closely mimic autoimmune targeting conditions inherent in the NOD mouse. The results presented here focus on epitope mapping with IFA. Complete Freund's adjuvant results yielded an alternate peptide (GFAP106–120), summarized in Fig. S5 and Table S2.

From the initial list of GFAP-IFA–generated putative CD4⁺ T-cell epitopes, peptides were selected that contained preferred I-A^g7 anchor residues, including large hydrophobic or positively charged residues at P6 and aromatic hydrophobic or positively charged residues at P9 (26,27). Using these criteria, GFAP epitopes with predicted I-A^g7 binding are listed in Table 1.

For MHC class I epitopes, the BIMAS (BioInformatics and Molecular Analysis Section) (http://thr.cit.nih.gov) and SYFPEITHI (www.syfpeithi.de) matrix-assisted algorithms were used to predict GFAP peptide sequences for K^d and D^b. BIMAS and SYFPEITHI scores for predicted GFAP K^d epitopes are presented in Table S3. These algorithms did not predict D^b binding peptides.

Spontaneous T-cell reactivity provides strong evidence for immunogenicity, natural processing, and disease relevance. Of the four MHC class II peptides synthesized, 116–130 and 216–230 consistently elicited positive responses in NOD mice, with peak responses occurring after weaning at 4 weeks of age. GFAP96–110 triggered responses to a lesser extent. GFAP321–335 did not respond (not shown). Figure 5 shows cpm data from four individual diabetic mice (Fig. 5A, C, and E) and kinetic results (Fig. 5B, D, and F) expressed as maximum SI). Maximum SI presents the highest proliferation observed irrespective of the peptide concentration (assayed from 5 to 250 μg/ml).

Of the predicted MHC class I epitopes, only GFAP79–87 and 253–261 gave positive T-cell responses (Fig. 5G–J). GFAP 112–120, 200–208, and 320–328 remained negative (not shown). The minimum dose at which a proliferative response was observed varied from mouse to mouse, the particular peptide, and their age, likely reflecting stochastic TCR usage and the nonsynchronous disease course. Nevertheless, responses to GFAP epitopes could be detected by 2.5 weeks of age, suggesting that priming and expansion of GFAP reactive T-cell pools occur early along with epitope specificity.

To further confirm the immunogenicity of the putative MHC class II GFAP peptides, 100 μg GFAP 96–110, 116–130, and 216–230 were each emulsified in CFA and recall responses assessed 10 days post–intraperitoneal immunization in 8-week-old NOD mice. All three peptides were immunogenic, as evidenced by boosted recall responses (Fig. S6a–c). Interestingly, CFA immunization with this panel of MHC class II GFAP peptides also elevated the low-dose whole protein GFAP response (Fig. S6d). This panel of GFAP peptides responds spontaneously in NOD mice but not in the nonautoimmune B6 (Fig. S7).

To ensure that the MHC class I peptides were indeed CD8⁺ T-cell specific, 8-week-old NOD females were immunized subcutaneously with GFAP 79–87 or 253–263 in CFA. Five days later, CD8⁺ T-cells were negatively selected from draining lymph nodes and incubated with peptide-pulsed RMA-S(K^d) targets. GFAP 79–87 and 253–263 immunization generated peptide-specific CTLs (Fig. 6A), confirming their MHC class I specificity.

This conclusion was tested by using the monoclonal antibody SF1–1.1 to block K^d in our in vitro proliferation assay. As shown in Fig. 6B, K^d inhibition reduced proliferation to GFAP 79–87 and GFAP 253–261, although the effect was more marked in GFAP 79–87. Differing affinities may account for the specific degree of inhibition; nevertheless, it is clear that the GFAP MHC class I peptides are susceptible to K^d inhibition. K^d blockade also partially reduced the response to whole-protein GFAP.

We next assessed the disease relevance of MHC class I GFAP epitopes by immunotherapy. 8-week-old NOD females were injected intraperitoneally with 100 μg GFAP 79–87 or 253–261 or vehicle control in IFA. At the dose and route used, GFAP 79–87, but not 253–261, provided significant protection from type 1 diabetes (P < 0.02) (Fig. 6C). Insulitis in normoglycemic GFAP 79–87–treated animals stalled predominantly at the peri-insulitis stage (data not shown). Intriguingly, diabetes protection was associated with a marked reduction in IFN-γ production to whole-protein GFAP (P < 0.03) (Fig. 6D, right side) despite equivalent proliferative responses in GFAP 79–87 and GFAP 253–261–immunized mice (Fig. 6D, left side). No differences in IL-10 production or recall responses to the peptides were observed at the end point (not shown); however, GFAP79–87–immunized mice exhibited divergent proliferative responses to whole GFAP and GFAP 216–230 at 10 days posttreatment (Fig. S8). No shifts in peptide responses were observed in GFAP253–261–immunized mice. Alterations in specific GFAP reactive T-cell pools may be responsible for the disease protection; however, the exact mechanism will require further investigation. Nevertheless, immunotherapy with a MHC class I GFAP peptide was capable of modifying diabetes, fulfilling our test of disease relevance.

Independently, the Nepom group has recently identified human MHC class I GFAP epitopes for HLA-A*0201 (28). Though we used the NOD mouse to identify class I and II restricted GFAP epitopes, the I-A^g7 MHC class II molecule shares the same nonasparagine polymorphism at position 57 of the β-chain as the diabetes-predisposing HLA-DQ8 (29). Therefore, we asked whether peripheral blood mononuclear cells from recent-onset diabetic children would recognize our panel of NOD-derived GFAP epitopes. As shown in Fig. 7, all three GFAP MHC class II peptides elicited proliferation at levels significantly different from healthy, autoantibody-negative FDRs. Notably, our I-A^g7 GFAP216–230 epitope overlaps GFAP214–222, an epitope reported by Nepom and colleagues to bind HLA-A*0201 and induce granzyme B production in type 1 diabetic CTLs (28). This is consistent with other type 1 diabetes autoantigens where MHC class I and II epitopes have been found in overlapping regions or in close proximity (30,31).

Although MHC class I genotyping was not performed in our patient set, MHC class I alleles are in linkage disequilibrium with DR3/DR4, and it is plausible that our K^d mapped peptides also bind risk-conferring HLA-A or -B alleles (32,33). Indeed, our K^d epitope GFAP 253–261, but not GFAP 79–87, gave proliferative responses statistically distinguishable between patients and healthy FDRs (P < 0.003, Mann-Whitney t test).

By analyzing T-cell responses to a broad array of self-antigens, an overall measure of T-cell reactivity can be derived (34). The T-cell score is defined as the number of positive (SI >1.5) proliferative responses to an array of type 1 diabetes–relevant test antigens/peptides (35). Using this system and an autoreactivity cutoff of T-cell score ≥3, all type 1 diabetic patients and 2 of 6 FDRs in this dataset can be considered autoreactive, with one borderline (Fig. 7B). Intriguingly, when considering global GFAP responses (whole protein and peptides), the three autoreactive FDRs generated 50% (8 of 16) GFAP responses above SI 1.5 (Fig. 7C). A prospective, large-scale, blinded study comparing diabetic subjects, FDRs, and normal control subjects is under way as part of the Trials to Reduce IDDM in the Genetically at Risk (TRIGR) diabetes prevention trial, with the goal of identifying autoimmune responses with disease-predictive value.

β-Cell destruction and subsequent insulin deficiency are responsible for the clinical symptoms associated with type 1 diabetes. Our finding that autoimmune islet destruction is not β-cell exclusive but includes pSC suggests that other islet constituents may also be involved in type 1 diabetes etiology (9). Here, we demonstrate that pSC autoimmunity involves cognate CD8⁺ T-cell targeting of GFAP.

Cultured pSC lacked MHC class II expression; therefore, we concentrated on cognate CD8⁺ T-cell interactions. pSC targets were lysed by ex vivo CD8⁺ T-cells from diabetic NOD mice, and transgenic rescue of MHC class I expression in pSC of NOD.β2m^−/− mice significantly accelerated adoptively transferred type 1 diabetes, providing evidence for disease relevance and an in vivo correlate for the cognate interaction observed between diabetogenic CD8⁺ T-cells and pSCs in vitro. Here, hyperglycemia required 12–16 weeks while β-cells remained intact in MHC class I deficient animals, reinforcing the importance of CD8⁺ T-cells in β-cell destruction (36).

This is the first report of primary pSC cultures facilitating molecular and functional studies of pSC, including their competence as antigen-presenting cells and association with and trophism for β-cells and relationship to peripheral glia. Physiologically, pSC loss may advance β-cell apoptosis by removal of neurotropic support, facilitating contact with pathogenic lymphocytes and disrupting islet architecture. Islet transplantation studies with or without pSC encapsulation and in combination with transgenic pSC modifications may be a new therapeutic strategy in type 1 diabetes.

GFAP was initially discovered as an autoantigen when analyzing brains from multiple sclerosis patients (10), which accumulate GFAP⁺ astrocytes in the multiple sclerosis plaque. In response to injury, astrocytes synthesize GFAP in a process called astrogliosis (37,38); similarly, pSCs undergo gliosis following islet insult (39). We now extend our study of GFAP autoimmunity in type 1 diabetes from the protein to the peptide level, mapping three I-A^g7 and two K^d epitopes. Although we cannot rule out the presence of other GFAP epitopes, the MHC class I peptides 79–87 and 253–261 and the MHC class II peptides 96–110, 116–130, and 216–230 stimulated T-cell responses in nonimmunized NOD mice, diabetic children, and some FDRs. Confirmation of GFAP 106–120, revealed by CFA immunization, is under further investigation.

We employed immunotherapy to demonstrate disease relevance and were successful with GFAP 79–87. Although GFAP 253–261 was ineffective, immunotherapy depends on route, timing, and optimal dose for efficacy (22). Differences in affinity may account for the apparent disparity between the two K^d peptides. Mechanistically, GFAP 79–87 immunization nearly abolished IFN-γ production to whole-protein GFAP. Although IL-10 production remained equivalent, immunoregulation involving transforming growth factor-β is a likely explanation (40). Alterations in GFAP epitope targeting (Fig. S8) is another possibility.

GFAP autoimmunity may contribute to other type 1 diabetes–associated diseases such as multiple sclerosis, chronic inflammatory demyelinating polyradiculoneuropathy, (41) and Crohn's disease. We and others have shown a connection between type 1 diabetes and multiple sclerosis (5,6) and linkage between chronic inflammatory demyelinating polyradiculoneuropathy and type 1 diabetes (42,43) and Crohn's and type 1 diabetes (44,45) have been reported. Whether glial cells are generally involved in organ-selective autoimmunity remains to be determined.

The identification of GFAP epitopes is a crucial step in the development of GFAP tetramers and TCR transgenics. Why widely expressed neuronal autoantigens such as GFAP contribute to local islet inflammation remains unknown. Differences in tissue access, antigen processing and presentation, costimulation, and regulatory T-cell populations might, in combination, explain the islet selectivity of type 1 diabetes infiltrates (7,46,47). Genetic ablation of B7-2 induces a CD4⁺ T-cell–mediated peripheral neuropathy, which presumably also targets GFAP, but with effector mechanisms distinct from those involved in type 1 diabetes (48). Whether the same GFAP peptides are targeted in B7-2–deficient NOD mice will be interesting to discern.

Recently, we have shown that sensory neurons are involved in a local regulatory circuit that controls islet inflammation and β-cell stress (4). Clearly, neuronal elements are involved in β-cell demise, and a complete understanding of type 1 diabetes pathogenesis will require a collaborative perspective between immunology and neuroscience.

H.T. received scholarships from RESTRACOMP (Hospital for Sick Children), the Banting and Best Diabetes Centre, and the OGS Trust Fund. This research was supported by the Canadian Institutes for Health Research (CIHR) (to J.R.E., D.J.B., P.S.O., and H.-M.D.); by a University-Industry grant from CIHR and SynX Pharma, Toronto (to H.-M.D.); and by the Juvenile Diabetes Research Foundation, the National Institutes of Health (GCRC MO1 RR00084 and RO1 DK24021 to D.J.B.), and the Renziehausen fund (to D.J.B.).

We thank Lily Morikawa for histology and Drs. David Serreze, Pere Santamaria, and Michael Brenner for reagents.