A Map of Human Type 1 Diabetes Progression by Imaging Mass Cytometry

SUMMARY

Type 1 diabetes (T1D) results from the autoimmune destruction of insulin-producing β cells. A comprehensive picture of the changes during T1D development is lacking due to limited sample availability, inability to sample longitudinally, and the paucity of technologies enabling comprehensive tissue profiling. Here, we analyzed 1,581 islets from 12 human donors, including eight with T1D, using imaging mass cytometry (IMC). IMC enabled simultaneous measurement of 35 biomarkers with single-cell and spatial resolution. We performed pseudotime analysis of islets through T1D progression from snapshot data to reconstruct the evolution of β cell loss and insulitis. Our analyses revealed that β cell destruction is preceded by a β cell marker loss and by recruitment of cytotoxic and helper T cells. The approaches described herein demonstrate the value of IMC for improving our understanding of T1D pathogenesis, and our data lay the foundation for hypothesis generation and follow-on experiments.

eTOC BLURB

Using imaging mass cytometry of pancreas sections from human donors, XXX et al generate a detailed marker-based timeline and find that that loss of β cell markers precedes β cell destruction and that cytotoxic and helper T cells are recruited simultaneously to β cell-rich islets in type 1 diabetes.

Graphical Abstract

INTRODUCTION

Type 1 diabetes (T1D) is a chronic condition thought to result from an autoimmune attack on insulin-producing β cells in the pancreatic islets of Langerhans (Atkinson et al., 2014). The disorder, characterized by overt hyperglycemia, develops from a poorly understood combination of genetic and environmental factors and is thought to involve complex interactions between islets and cells of the immune system (Boldison and Wong, 2016). The study of human T1D has been limited by sample availability. Further, imaging of the pancreatic islets cannot be performed in vivo, and pancreatic biopsies, which would enable longitudinal studies of the changes of islet cell composition and cell-cell interactions, are not undertaken due to the associated investigational risks. Consequently, much of our knowledge regarding T1D pancreas pathology is derived from mouse models and limited analysis of individual specimens obtained from autopsy collections of human pancreata (Gepts, 1965; Foulis and Stewart, 1984). To enable more systematic studies of human pancreatic T1D tissue, biobanks such as the Network for Pancreatic Organ donors with Diabetes (nPOD) or the Human Pancreas Analysis Program (HPAP) (Campbell-Thompson et al., 2012a; Marchetti et al., 2018) have been established to collect tissues from organ donors.

In recent years, studies of human pancreas samples from these biobanks have provided valuable information regarding the pathogenesis and natural history of this disease. For example these analyses revealed the presence of islet-specific autoreactive CD8⁺ T cells in the pancreata of donors with T1D (Coppieters et al., 2012). It has also become evident that insulitis is less aggressive amongst T1D patients than it is in mouse models of T1D (Meier et al., 2005; Coppieters et al., 2012). Heterogeneity is observed on the cellular level, as the loss of functional β cell mass is not only caused by β cell destruction but also by dysfunctional β cells (Chen et al., 2017). On the individual pancreas level, insulitis often occurs in a lobular fashion and even islets in proximity show the presence and absence of insulitis (Gianani et al., 2010; Rodriguez-Calvo et al., 2015). A large degree of inter-patient heterogeneity is observed in terms of insulitis frequency and composition, residual β cell mass, and pace of β cell destruction (Campbell-Thompson et al., 2016), with insulitic lesions sometimes persisting for decades after diagnosis (Meier et al., 2005; Coppieters et al., 2012).

Despite these findings, our understanding of T1D progression in human patients is limited. An understanding of how cell types, cell states, and cell-cell interactions evolve during T1D development is essential to designing strategies to cure or halt this disorder. The lack of a comprehensive spatially resolved cellular analysis of the T1D lesion is due, at least in part, to the fact that only recently have highly multiplexed tissue imaging technologies emerged that allow for comprehensive studies of tissue from rare donor samples. These approaches include imaging techniques that are based on cyclic immunofluorescence (Schubert et al., 2006; Gerdes et al., 2013; Zrazhevskiy and Gao, 2013; Lin et al., 2015; Goltsev et al., 2018; Gut et al., 2018), epitope-targeted mass spectrometry (Angelo et al., 2014; Giesen et al., 2014), and RNA detection (Ke et al., 2013; Lee et al., 2014; Chen et al., 2015; Schulz et al., 2018). These methods provide a comprehensive view of marker distribution and tissue composition and are, therefore, poised to unlock some of the mysteries surrounding the pathogenesis of T1D and other complex disorders (Bodenmiller, 2016).

Here we used imaging mass cytometry (IMC) to analyze pancreas tissue from four control organ donors as well as four at onset and four with long-term T1D. IMC combines laser ablation with mass cytometry to enable simultaneous evaluation of up to 50 biomarkers from tissue sections stained with metal-tagged antibodies (Giesen et al., 2014). Using IMC data, we generated a map of islet heterogeneity as well as interactions between islet and immune cells. Pseudotime analysis at the islet-level enabled us to reconstruct islet evolution through T1D progression from the snapshot data from the 12 donors. These analyses revealed that a large fraction of islets from patients with recent-onset T1D retain a near-normal fraction of β cells but that the phenotype of these β cells is altered prior to β cell destruction. Finally, we investigated the dynamics of islet-associated immune cells and the evolution of cell-cell interactions during T1D progression. We observed that, although association between islet cells and immune cells is rare, both cytotoxic and helper T cells are simultaneously recruited to β cell-rich islets during T1D onset, indicating the assembly of a destructive immune motif.

RESULTS

Preprocessing of Human T1D Samples and Image Acquisition

In IMC, tissue sections are labeled with metal-conjugated antibodies (Figure 1A). During image acquisition, a laser ablates the tissue spot-by-spot at a resolution of 1 µm, and the clouds of tissue particles are analyzed by mass cytometry. The metal composition of the ablated tissue and thus, by inference, the antibody and epitope content is used to create images pixel-by-pixel, each pixel corresponding to one laser shot (Giesen et al., 2014). To enable IMC measurement of pancreas specimens from human donors with T1D, we designed a novel panel of 35 metal-tagged antibodies targeting key islet cell antigens as well as immune cell markers (Table 1 and STAR methods).

Figure 1. Acquisition of Single-Cell and Islet Level Data by IMC.

A. IMC workflow. A laser ablates the tissue spot-by-spot, generating clouds of particles that are atomized and ionized in an inductively coupled plasma before analysis by mass cytometry. Based on ion cloud composition, an image of the ablated area is created for each measured marker.

B. IMC images, here represented by a combination of proinsulin (magenta), glucagon (cyan), and somatostatin (yellow) (top), were segmented into cells. The cell mask is shown as a white overlay (bottom).

C. Marker expression was measured by applying cell masks onto IMC images (left). Protein abundance in single cells can be visualized using tSNE (right).

D. Imaged cells were classified into cell types (left). Cells were first divided into four categories: islet (red), exocrine (blue), immune (green), and “other” (yellow) (left). Each category was then subdivided into cell types. Islet cells were classified as α (cyan), β (magenta), δ (yellow), or γ (red) cells (center). This allowed determination of islet cellular composition (right).

E. Cell masks enabled retrieval of spatial data, such as cell area (left). Neighborhood information (center) was used to systematically analyze cell-cell interactions (right).

F. The expression of endocrine cell markers, such as synaptophysin (red) (top), was used to identify islets and generate islet masks (white overlay, bottom).

G. By combining cell and islet masks, each cell was related to the islet to which it belongs (left). The distance of cells to the islet rim was also measured (center), allowing comparison of cell positions relative to islets (right).

H. Spatial information (left) and protein abundance (center) were measured at the islet level using islet masks. Islet level data were used to perform pseudotime analyses (right). Scale bars: 100 µm.

Table 1.

Imaging Mass Cytometry Panel

Compartment	Short name	Antigen	Main target cells	Metal tag
Pancreas	INS	Insulin	β	141Pr
	PIN	Proinsulin	β	154Sm
	GCG	Glucagon	α	156Gd
	SST	Somatostatin	δ	159Tb
	PPY	Pancreatic polypeptide	γ	153Eu
	NKX6–1	Nkx6.1	β	169Tm
	PDX1	Pdx1	β, δ, ductal	158Gd
	IAPP	Amylin	β	167Er
	PCSK2	Proprotein convertase 2	α	144Nd
	SYP	Synaptophysin	Endocrine	160Gd
	CD99	CD99	Endocrine	145Nd
	SLC2A1	Glucose transporter 1	Endocrine	148Nd
	PTPRN	Receptor-type tyrosine-protein phosphatase-like N	Endocrine	174Yb
	AMY2A	Pancreatic amylase	Acinar	150Nd
	KRT19	Cytokeratin 19	Ductal	161Dy
	CD44	CD44	Exocrine	143Nd
Immune	CD45	CD45	Immune	162Dy
	CD45RA	CD45RA	Immune	164Dy
	CD3e	CD3ε	T	152Sm
	CD4	CD4	Helper T	171Yb
	CD8a	CD8a	Cytotoxic T	165Ho
	CD20	CD20	B	149Sm
	CD68	CD68	Monocytes, macrophages	146Nd
	MPO	Myeloperoxidase	Neutrophils	147Sm
	FOXP3	Forkhead box P3	Regulatory T	163Dy
	CD38	CD38	Immune	142Nd
Misc	CDH	E-/P-cadherin	Epithelial	173Yb
	CD31	CD31	Endothelial	172Yb
	SMA	Smooth muscle actin	Stromal	115Yb
	KI-67	Ki-67	Proliferating	168Yb
	pH3	Phospho-histone H3	Proliferating	170Yb
	pRB1	Phospho-retinoblastoma	Cycling	175Yb
	cPARP1	Cleaved caspase 3 + poly (ADP-ribose) polymerase	Apoptotic	176Yb
	CA9	Carbonic anhydrase IX	Hypoxic	166Er
	H3	Histone H3	Nuclei	113In
	Ir191, Ir193	(DNA intercalators)	DNA	191Ir, 193Ir

We selected 12 cases from the nPOD biorepository to represent recent-onset T1D (< 0.5 years, N = 4), long-standing T1D duration (≥ 8 years, N = 4), and controls without diabetes (N = 4) and for each donor analyzed two sections originating from different anatomical regions of the pancreas (tail, body, or head) (Table S1). Because the speed of data acquisition prevents imaging of entire pancreas sections, we used immunofluorescence (IF) to perform a pre-selection of areas of interest (Figure S1A and STAR Methods). We then stained the same sections with metal-labeled primary antibodies (Table 1) and imaged the selected areas by IMC. Our measurements yielded 845 multiplexed images that included 1581 islets (each with ≥ 10 cells); data were obtained in 37 channels corresponding to the 35 antibodies and two DNA counterstains in our panel (Figure S2).

Extraction of Single-Cell and Islet Level Data

Cell segmentation is essential to recover quantitative single-cell information from highly multiplexed images (Carpenter et al., 2006). We used supervised machine learning and computer vision algorithms to generate cell and islet segmentation masks, which represent pixels belonging to the same cell or islet, respectively (Figure 1B, Figure S3 and STAR Methods) (Kamentsky et al., 2011; Sommer et al., 2011). Applying these masks over high-dimensional pictures allowed retrieval of phenotypic and functional marker expression, spatial information and neighborhood information. We also combined cell and islet masks to extract additional information, such as the distance from cells to the islet rim (Figure 1C-H). Although cell populations can be defined using clustering approaches, we sought a more accurate way to define cell types in our dataset and turned to supervised machine-learning approaches (Figure S3 and STAR Methods) (Sommer et al., 2011). We first trained a classifier to segregate cells into four main categories (i.e., islet, immune, exocrine, other) and then performed sub-classification within each category in order to identify individual cell types. Together, these approaches enabled extraction of a wide range of biological information that can be explored in downstream data analyses to gain deeper insights into cell phenotypes and tissue function.

Evolution of Islet Cellular Composition and Architecture during T1D Progression

Even within a single pancreas, islet size and cell type composition are highly heterogeneous (Brissova et al., 2005; Cabrera et al., 2006). Whether this heterogeneity influences T1D progression remains unknown. We, therefore, sought to determine how islet structure and cell type composition change when T1D progresses. The 1581 imaged islets displayed striking heterogeneity in terms of cell number and cell type composition (Figure 2A). We also observed large inter-donor variations (Figure 2B). As compared to non-diabetic controls, β cell fraction was reduced by 62% in donors with recent-onset T1D. Remarkably, two of the four samples from donors at T1D onset had a proportion of β cells approaching those observed in some of the non-diabetic individuals (nPOD cases 6414 and 6380 had 57% and 72% of the average β cell fraction in control donors, respectively). As expected, however, pancreata from donors with prolonged disease duration were almost entirely devoid of β cells. Next, we examined intra-pancreas heterogeneity to determine whether different regions of the same pancreas differed in cell type composition. Within each pancreas, islet cellular composition was remarkably homogenous. We observed a significant difference in β cell loss between the pancreas body and tail in only one of the donors with recent disease onset (nPOD case 6380) (Figure 2C). Previous reports indicate that the head region of the pancreas is enriched in γ cells (Malaisse-Lagae et al., 1979; Brereton et al., 2015). We made a similar observation in the only pancreas head section analyzed (nPOD case 6264).

Figure 2. IMC Exploration of Islet Heterogeneity.

A. Cell type composition of the 1581 islets (≥ 10 islet cells) captured in our dataset. Each stacked bar represents a single islet with colors indicating β (red, n = 56,594), α (light red, n = 101,592), δ (blue, n = 18,833), and γ (light blue, n = 3721) cells.

B. Average islet cell type fraction for each of the 12 donors in our dataset.

C. Average islet cell type fraction per pancreas section. The region of origin of each section is indicated on the x-axis. Significant P values for comparison of β cell fractions between sections of the same pancreas are indicated.

D. Average islet cell type fraction versus islet size. For each bin size, β cell fraction was compared to the average β cell fraction for the central bin (35–100 cells).

E. Islet cell type fractions as a function of distance to islet rim. Cell types β (red, n = 39,557), α (light red, n = 73,174), δ (blue, n = 14,396), γ (light blue, n = 3431) cells. Bin sizes, 5 µm. The leftmost bin represents cells located within 5 µm of the islet rim and increasing bin numbers indicate cells located closer and closer to the islet center.

F. Islet cell type fractions as a function of distance to blood vessels. Cell types: β (red, n = 42,948), α (light red, n = 89,782), δ (blue, n = 16,114), γ (light blue, n = 2734) cells. Bin sizes, 5 µm. The leftmost bin represents cells located within 5 µm of the blood vessel, and increasing bin numbers represent cells located farther and farther away.

*, P < 0.01; **, P < 0.001; ***, P < 0.0001 for β cell fractions; †, P < 0.01; ††, P < 0.001; †††, P < 0.0001 for cell fractions (Mann-Whitney U test with Holm adjustment). Error bars, SEM.

Next, we examined how cell type composition varied with islet size. We found that in non-diabetic donors, the percentage of β cells was highest in islets of intermediate size, and lowest in very small (< 10 cells) and very large (> 250 cells) islets (Figure 2D). The distribution of α cells exhibited the opposite trend, whereas δ cell fractions were fairly constant. We noticed a similar pattern in islets from recent-onset donors, whereas the few residual β cells found in donors with long-standing T1D were most often found isolated or in very small islets, in line with previous observations (Wasserfall et al., 2017).

Finally, we measured the distance of islet cell subsets to the islet rim and to blood vessels in order to evaluate human islet architecture. Because the risk of overestimating distances in two-dimensional images increases with the measured distance, we limited the range of our analyses to cells located within 25 µm of islet rims or blood vessels. We observed that in non-diabetic donors, α cells were preferentially located at the islet periphery than in the islet core, whereas β cells exhibited an opposite tropism (Figure 2E). Similarly, α cells were more frequently found along blood vessels than were β cells (Figure 2F). These findings support earlier observations (Bosco et al., 2010; Brereton et al., 2015). Further, we observed that this islet organization is globally conserved between non-diabetic and recent-onset T1D donors, whereas in islets from donors with long-standing T1D architecture is altered due to β cell loss (Figure 2E-F).

Evolution of Islet Profiles during T1D Progression

It would be ideal to observe changes of cell phenotypes and cell-cell interactions that take place inside islets during T1D progression. Unfortunately, the study of T1D pancreas pathology relies on samples collected post-mortem, only allowing for snapshot analysis. To generate a time resolved map of T1D pathogenesis, we sought to reconstruct islet evolution from cross-sectional data. To achieve this, we first determined islet expression profiles by quantifying abundances of 12 islet-expressed markers, averaged over the islet area. Hierarchical clustering of islet profiles distinguished non-diabetic from diseased donors but did not clearly separate recent-onset donors from those with long-standing disease (Figure S4A). To visualize individual islet profiles, we performed a dimensionality reduction analysis with tSNE (van der Maaten and Hinton, 2008). tSNE confirmed that islets from donors with or without T1D exhibited clearly distinct profiles (Figure 3A and Figure S4B), but the separation between islets from recent-onset and long-duration T1D donors was only partial. In particular, a large cluster composed of islets from donors with different disease durations was found (Figure 3A, circled in dotted blue line). These results indicated that islets from donors at very different T1D stages can have similar profiles.

Figure 3. Evolution of Islet Profiles with T1D Progression.

A. tSNE map representing the expression profiles of 1494 islets (≥ 15 cells) from 12 donors at different stages of T1D (one circle = one islet, 90 to 155 islets per donor, see also Figure S4B). The dotted blue circle indicates a cluster composed of islets from donors with different disease durations.

B. Pseudotime analysis of islet expression profiles using SCORPIUS. Islets, represented by circles, were first projected into a two-dimensional space (Components 1 and 2). A trajectory, represented by the black curve, was then inferred in the reduced space. Pseudotime was measured by projecting the islets onto the trajectory (see also Figure S4C-D).

C. Heatmap showing expression of islet markers with columns representing islets arranged according to pseudotime progression.

D. Violin plots showing islet distribution over pseudotime for the three disease stages (left). Pseudostages are indicated in blue and are defined based on the density distribution of islets for each disease stage (right). The percentage of islets in each stage and pseudostage is indicated in blue. Horizontal blue lines, defined by the intersections of islet density distributions, indicate pseudostage separations (see also Figure S4E).

E. Islet β cell fraction for each pseudostage. Pseudostage 1, n = 514 islets; pseudostage 2, n = 289 islets; pseudostage 3, n = 691 islets.

F. Islet extent and solidity measured for each pseudostage. For both metrics, a decrease indicates a more concave or irregular shape.

G. Evolution of β cell marker expression in β cells through pseudostages. Expression levels were measured in single β cells and averaged for each islet (see also Figure S5A). Pseudostage 1, n = 38,479 β cells from 514 islets; pseudostage 2, n = 16,103 β cells from 289 islets; pseudostage 3, n = 618 β cells from 691 islets.

Pairwise comparisons of the three pseudostages: *, P < 0.01; **, P < 0.001; ***, P < 0.0001, n.s., not significant (Mann-Whitney U test with Holm adjustment).

To better understand how islet expression profiles change with disease progression, we performed a pseudotime analysis using the trajectory inference algorithm SCORPIUS (Figure 3B and Figure S4C) (Cannoodt et al., 2016; Saelens et al., 2018). This approach revealed that the abundance of β cell markers decreased with pseudotime progression (Figure 3C and Figure S4D). Strikingly, visualizing islet distribution over pseudotime revealed a bimodal distribution of islets from donors with recent T1D onset (Figure 3D, left). To analyze this observation in more detail, we calculated the density distributions in the assigned pseudotime for each stage (Figure 3D, right). We defined three pseudostages based on the intersections between these distributions, thereby allowing stratification of islets according to their individual expression profiles rather than on donor disease duration. Using this approach, islets from non-diabetic donors were distinct from islets from donors with T1D, but also about half of the islets from donors categorized as recent onset exhibited the phenotype found in long-duration T1D (Figure 3D). We observed a similar bimodal distribution in samples from each of the four donors with recent-onset T1D (Figure S4E). Together these analyses revealed that islet profiles do not necessarily correlate with disease duration.

Because pseudotime was computed at the islet level rather than at the single-cell level, the observed decreased expression of β cell markers could be due to a loss of β cells or to downregulation of β cell markers in β cells or both. To separate these phenomena, we first compared the islet β cell fractions at the three pseudostages. The proportion of β cells decreased only slightly from pseudostage 1 to pseudostage 2, whereas pseudostage 3 was essentially composed of pseudo-atrophic islets, i.e. which are devoid of β cells (Figure 3E). In addition, we examined how islet shape changed with pseudostage progression. Islet extent (islet area divided by islet bounding box) and solidity (proportion of pixels in the islet convex hull that are also in the islet), two metrics that reflect shape regularity, were similar for pseudostages 1 and 2 and decreased for pseudostage 3 (Figure 3F). This reflects altered islet shape at pseudostage 3, likely because of β cell destruction.

We then examined how the expression of β cell markers evolved with pseudotime. We assessed the abundance of β cell markers in the β cells of each islet as a function of pseudotime. This approach revealed that expression of four β cell markers (INS, PIN, IAPP, and PTPRN) was strongly downregulated between pseudostages 1 and 2 but remained relatively stable between pseudostages 2 and 3 (Figure 3G and Figure S5A). By contrast, the expression of β cell transcription factors, PDX1 and NKX6–1, and pan-endocrine markers, SYP and CD99, did not vary substantially through the three pseudostages. Together, these results indicate that progression from pseudostage 1 to 2 is driven by a downregulation of β cell-specific markers, while progression to pseudostage 3 reflects β cell death. Similar results were obtained using Monocle2 (Qiu et al., 2017) instead of SCORPIUS for pseudotime analysis (Figure S5B-G), indicating that our results are independent of the pseudotime mapping algorithm used. We performed similar analyses at the β cell level. Despite heterogeneity, β cells from non-diabetic controls and donors with T1D were largely separated by both hierarchical clustering and tSNE and had limited pseudotime overlap (Figure S5H-K). As in the islet-level analysis, this separation was mainly driven by a relative decrease in the abundance of INS, PIN, IAPP and PTPRN (Figure S5L). In summary, these findings indicate that islet evolution during T1D progression can be reconstructed from cross-sectional data and suggest that, at the islet level, alteration of the β cell phenotype precedes β cell destruction.

Tc and Th Cells are Simultaneously Recruited to β Cell-Rich Islets at T1D Onset

Autoimmunity is thought to be the driving force behind β cell death in T1D (Atkinson et al., 2014; Boldison and Wong, 2016). We therefore explored the evolution of immune cell composition through disease duration. We used supervised machine learning to identify the main immune cell types in our dataset: cytotoxic T cells (Tc, CD3⁺CD8⁺CD45RA-), naïve (possibly EMRA) Tc cells (naïve Tc, CD3⁺CD8⁺CD45RA⁺), helper T cells (Th, CD3⁺CD4⁺), B cells (CD20⁺), monocytes/macrophages (MΦ, CD45⁺CD68⁺), and neutrophils (Neutr., MPO⁺). In addition, remaining CD45⁺ or CD45RA⁺ cells, to which no specific cell type could be attributed, were classified as “other”. Overall, T cells were more abundant in donors with recent-onset T1D and the number of B cells was very low in most pancreata, except for one donor with recent-onset T1D (nPOD case 6228) (Figure 4A). We then measured the density of each immune cell type (i.e., number of cells per mm²) across T1D duration. All measured immune cell types were more abundant in recent-onset than in control donors (Figure 4B). In long-duration T1D donors, MΦ and neutrophil densities were elevated relative to donors without T1D, whereas Tc and Th cell abundances were decreased compared to recent-onset donors but remained higher than in donors without T1D (Figure 4B and S6A). We then tested whether immune cell densities in islets differed from those in the surrounding exocrine tissue. Strikingly, the density (cells per mm²) of immune cells was lower in islets than in the exocrine compartment for all immune cell types except B cells at all disease stages (Figure 4C). Yet, while immune cell densities never exceeded 20 cells per mm² in the exocrine tissue, a few islets exhibited much higher values. Next, we investigated whether positions of immune cells relative to islets varied but did not observe an accumulation of immune cells at islet peripheries (Figure 4D). The exception to this was observed in a donor with recent-onset T1D (nPOD case 6414) with an accumulation of Tc and Th cells around the islet periphery (Figure S6B). Together, these results indicate that more immune cells are found in pancreata from donors with T1D than in non-diabetic controls. Most of these immune cells are not directly associated with islets, however.

Figure 4. Dynamics of Immune Cell Recruitment during T1D Progression.

A.Distribution of immune cell types per donor and T1D stage. Number of cells analyzed: Tc, 15,367; Naïve Tc, 890; Th, 5417; B, 869; MΦ, 16,021; Neutr., 20,800; Other, 10,299.

B. Density (cells per mm²) of immune cell types in pancreata from controls and donors with T1D (see also Figure S6A). P values, pairwise comparisons of T1D stages.

C. Immune cell densities inside islets (green lining) and in the exocrine tissue (blue lining) in controls and donors with T1D. P values, comparison of densities in islets versus exocrine tissue.

D. Average number of immune cells over binned distance to islet rim (see also Figure S6B). Bin size, 10 µm. Negative and positive values represent bins located outside and inside islets, respectively. The dotted vertical lines indicate the islet rim. B cells were not numerous enough to be meaningfully plotted. Error bars, SEM.

E. Fraction of islets containing given numbers of associated immune cells (inside the islet or within 20 µm of the islet rim), in function of pseudostage (see also Figure S6C).

F. Number of immune cells associated with islets at different pseudostages. Number of cells analyzed: Tc, 3533; Naïve Tc, 264; Th, 1309; B, 165; MΦ, 1915; Neutr., 2234; Other, 1868. P values, comparison of immune cell numbers for consecutive pseudostages.

G. Number of Tc (red) and Th (blue) cells associated with individual islets plotted versus islet pseudotime. Curves: smoothed conditional means (local polynomial regression fitting), grey areas: 95% confidence intervals.

H. Number of Tc and Th cells per islet at different pseudostages. Pearson correlations are indicated, P values were < 1e⁻¹² for the four pseudostages.

*, P < 0.01; **, P < 0.001; ***, P < 0.0001, n.s., not significant (Mann-Whitney U test with Holm adjustment).

To further explore interactions between islets and immune cells in T1D, we sought to correlate islet pseudostage information, which reflects the islet marker expression profile, with our immune cell data. We counted the number of immune cells associated with each islet (i.e., the number of immune cells located either inside the islet or within 20 µm of the islet rim) and calculated the fraction of islets with immune cell infiltration at different disease stages (Figure 4E and Figure S6C). Most islets from non-diabetic controls and donors with long-duration T1D had no or few associated immune cells. By contrast, around half of the islets from donors with recent-onset T1D had six associated immune cells or more. Strikingly, insulitis (> 15 associated immune cells) was present in over 30% of β cell-rich (pseudostage 2) islets from recent-onset donors, but in less than 10% of β cell-depleted (pseudostage 3) islets from the same donors. We then examined how immune cell numbers varied with pseudostage progression. Islet-associated Tc and Th cells were more abundant in pseudostage 2 islets but these numbers decreased in pseudostage 3 islets from recent-onset T1D donors and further decreased in pseudostage 3 islets from donors with long-duration T1D, suggesting that T cells progressively leave islets after β cell destruction (Figure 4F). Together, these results indicated that the number of islet-associated T cells depends on both disease duration (recent-onset versus long-duration) and the islet profile (presence of surviving β cells).

We observed that Tc and Th cell numbers followed similar trajectories through pseudotime (Figure 4G), suggesting that these cells are recruited simultaneously to individual islets in T1D. This was confirmed by the positive correlation between the numbers of Tc and Th cells associated with individual islets, in particular at pseudostage 2 (Figure 4H). These findings indicated that the immune-mediated attack on β cells involves both Tc and Th cells, is maximal around the time of T1D onset, and focuses on islets that retain a high proportion of β cells. Moreover, our results show that pseudotime analysis based on islet expression profiles can be meaningfully combined with immune cell data to recreate the dynamics of interactions between islets and immune cells through T1D duration.

Neighborhood Analysis Reveals Dynamic Changes in Cell-Cell Interactions in T1D

In health and disease, tissue function depends on interactions between different cell types. This is also the case in the pancreas: Normal islet function depends on paracrine interactions between islet cells, and T1D progression is determined by interactions between immune and β cells. Understanding how cell-cell interactions change from one disease stage to the other is, therefore, key to a better understanding of T1D. To systematically probe interactions between cell types, we performed a large-scale neighborhood analysis (Schapiro et al., 2017). We measured, for each image, significant association or avoidance according to spatial proximity of all cell type pairs in our dataset (Figure 5A). Hierarchical clustering of individual images based on cell-cell interactions revealed a clear separation between images containing pseudostage 3 islets and those containing islets at pseudostages 1 or 2. This separation was essentially caused by β cell destruction, as illustrated by the rarity of significant association between β cells and other cell types in images containing pseudostage 3 islets (Figure 5A, green rectangle). This result confirmed that islet structure is maintained between pseudostage 1 and pseudostage 2 islets but is altered in pseudostage 3 islets. We then focused specifically on islets and compared the frequency of interactions between cell types at each pseudostage (Figure 5B). As expected, we observed frequent association between islet cells (Figure 5B, green rectangles), but interactions involving β cells were strongly reduced or absent in pseudostage 3 islets, as a result of β cell loss. Association between immune cells was also frequent, in particular at pseudostages 2 and 3 (Figure 5B, magenta rectangles). By contrast, islet and immune cells essentially remained isolated from each other at all disease stages, confirming our previous observation that immune cells are rarely found within islets.

Figure 5. Neighborhood Analysis Reveals Dynamic Changes in Cell-Cell Interactions in T1D.

A. Heatmap showing interactions between all cell type pairs (columns) for each individual image in our dataset (rows). Colors indicate significant (P < 0.01) cell type association (red) or avoidance (blue). Rows are ordered by hierarchical clustering. Non-present and non-significant interactions are indicated in white. The green rectangle highlights interactions between β cells and other cell types in images containing pseudostage 3 islets.

B. Frequency of cell-cell interactions in islets and in the 20 µm surrounding area for each pseudostage. Colors indicate the percentage of donors that display significant association (red) or avoidance (blue) between two cell types. Green rectangles indicate interactions between islet cells and magenta rectangles indicate interactions between immune cells.

C. Average number of interactions per β cell, with either Tc or Th cells, averaged by islet. Only islets containing at least one β cell were considered.

D. Average number of interactions per Tc or Th cell with β cells, averaged by islet. Only islets with at least one associated T cell were considered.

E. Fraction of islets showing association between β and T cells, in function of the percentage of β cells associated with T cells.

F. Fraction of islets showing association between T and β cells, in function of the percentage of T cells associated with β cells. *, P < 0.01; **, P < 0.001; ***, P < 0.0001, n.s., not significant (Mann-Whitney U test with Holm adjustment).

Cytotoxic and helper T cells are thought to be directly involved in the destruction of β cells in T1D (Boldison and Wong, 2016). Yet, our global neighborhood analysis did not detect association between T cells and β cells, even in donors with recent-onset T1D, where the rate of β cell death and islet infiltration are maximal. We, therefore, sought to study specifically the evolution of interactions between β and T cells through disease progression. First, we counted how often β cells were associated with Tc, respectively Th cells. This number remained extremely low in all islets but was highest for β cells from pseudostage 2 islets, i.e. β cell-rich islets from donors with recent-onset T1D (Figure 5C). Conversely, T cells were more often associated with β cells located in pseudostage 2 islets (Figure 5D). Next, we calculated for each islet the fraction of β cells associated with either Tc or Th cells. Again, more islets displayed such contacts at pseudostage 2 (Figure 5E). Even in those islets, though, the percentage of β cells associated with T cells was rarely above 10%. Conversely, the fraction of T cells associated with β cells was also higher in pseudostage 2 islets (Figure 5F). The small fraction of β cells in contact with T cells likely explains why our global neighborhood analysis did not allow detecting those associations. Together with the results of our immune cell analysis, these results suggested that not only pseudostage 2 islets have more associated T cells, but also that, in those islets, T cells contact β cells more frequently. Nevertheless, the frequency of interactions between β and T cells remained very low at all disease stages.

DISCUSSION

Here we used highly multiplexed imaging of human pancreas sections by IMC to generate a comprehensive map of T1D progression. Analysis of islet cell composition showed that two donors with recent-onset T1D had a proportion of β cells in islets similar to that observed in non-diabetic controls (Figure 2B). The discrepancy with β cell fractions usually reported can be, at least in part, explained by the presence of insulin-negative β cells, which we detected by co-expression of Pdx1 and Nkx6.1. These β cells with altered phenotypes are not detected when β cell identification relies exclusively on insulin (Rodriguez-Calvo et al., 2017; Wasserfall et al., 2017). The question remains, however, how patients with such high numbers of remaining β cells develop clinical symptoms of T1D. We found that roughly half of the islets from donors with recent-onset T1D retained a “normal” fraction of β cells with an unaltered islet structure (Figure 3E). The β cells in these islets, however, exhibited reduced expression of β cell markers as compared to β cells from healthy control islets (Figure 3G). This altered phenotype may be a consequence of stress and/or increased metabolic demand or, potentially, may allow escape from detection by the immune system (Rui et al., 2017). This finding is crucial from a therapeutic point of view because the presence of a high number of surviving β cells at T1D onset implies that if factors driving disease progression could be halted or slowed these cells might have the opportunity to regain functionality.

Much of our knowledge regarding T1D pancreas pathology has resulted from analyses of sections collected from deceased pancreas donors, and the lack of sampling over the course of the disease has clearly hampered study of this disorder. We show here that performing pseudotime ordering at the islet level represents a powerful method to reconstruct time trajectories from snapshot data, and studies such as ours may overcome the sample limitation issues. In particular, we investigated the dynamics of immune cell recruitment around T1D-affected islets and found that the number of islet-associated T cells depends on both disease stage and islet profile. In addition, we observed simultaneous recruitment of Tc and Th cells to individual islets before β cell destruction in subjects already diagnosed with T1D. This result confirms that these cell types form a destructive immune motif that plays a pivotal role in β cell demise. Finally, our analyses indicate that immune cells are rarely in direct contact with islet cells. This could be explained by the fact that infiltrating immune cells are more often located at the islet periphery (peri-insulitis), than inside islets (intra-insulitis) (Pugliese, 2016). The position of β cells, which are rarely located in the outer islet layer (Figure 2E), may represent an additional barrier. Finally, the chronic nature of the disease, with β cell destruction taking place over long-time periods, is consistent with a model in which contacts between T and β cells are relatively short and infrequent, making their detection in fixed tissue sections difficult.

For this first effort toward analysis of human pancreata using highly multiplexed imaging, we designed a broad antibody panel that enabled discrimination of the main pancreatic and immune cell types and selected cases that represented short and long duration T1D. We envision that future studies will attempt to address specific questions with tailored antibody panels and higher numbers of donors in order to increase the statistical power, to better capture heterogeneity among donors, and to explore other patient-related parameters known to influence T1D progression, such as age. A number of questions regarding T1D development remain open. We could not determine, for instance, whether recruitment of immune cells to pancreatic islets precedes or follows downregulation of insulin and other β cell markers. Because β cell phenotypes are already altered in recently diagnosed donors, determining the cause of this alteration and the underlying mechanisms will require studying samples from pre-diabetic subjects.

The techniques and concepts described in the current study, together with the co-submitted manuscript (Wang et al., 2018), will open new avenues for the exploration of T1D pancreas pathology and will help to address the crucial questions that still surround the disorder. Moreover, the approaches described herein should be applicable to the study of other diseases and will be especially valuable for deciphering complex cellular interactions.

Limitations of study

Some limitations remain when IMC is used for analysis of pancreas tissues. First, all results are a function of the selected antibodies. Antibodies used in this study were thoroughly validated (STAR Methods). In IMC, antigens are most often directly detected with primary antibodies so options for signal amplification are limited. Moreover, due to the multiplexed nature of IMC, a single staining protocol must be used, whereas in IF or IHC the conditions can be optimized for each antibody. Conjugation can also affect the performance of some antibodies. New antibody conjugation methods will become available in the near future (Allo et al., 2018). Second, the area that can be imaged by IMC is constrained by acquisition speed. To overcome this caveat, we used IF to select areas of interest for IMC analysis. Yet, the number of slides that can be imaged remains limited. In this study, we have analyzed two sections per donor, but we have acquired a large number of islets per section, such that variations related to the sectioning plane should average out. In this regard, we note that islet cellular compositions were globally similar between sections from different anatomical regions of the same pancreas. Third, among the most challenging processes in multiplexed image analysis is cell segmentation. Because cell borders are not always clearly defined and because individual pixels may contain information from more than one cell (Schapiro et al., 2017), tissue imaging data are noisier than what is typically obtained with suspension analysis methods. We have, therefore, implemented supervised machine learning for cell type attribution to overcome this challenge. These computational pipelines and new tools to be developed in the future will lead to further improvements in the quality of IMC-generated data.

STAR METHODS

KEY RESOURCES TABLE.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Mouse monoclonal anti-Alpha smooth muscle actin	Abcam	Cat#ab2376549, RRID:AB_2223019
Goat polyclonal anti-Carbonic anhydrase IX	R and D Systems	Cat#AF2188, RRID:AB_416562
Rabbit monoclonal anti-Caspase 3	BD Biosciences	Cat#559565, RRID:AB_397274
Rabbit monoclonal anti-CD3ε	Cell Signaling Technology	Cat#85061, RRID:AB_2721019
Goat polyclonal anti-CD4	R and D Systems	Cat#AF-379-NA, RRID:AB_354469
Mouse monoclonal anti-CD8a	Thermo Fisher Scientific	Cat#14–0085-82, RRID:AB_11150240
Mouse monoclonal anti-CD20	Thermo Fisher Scientific	Cat#14–0202-82, RRID:AB_10734340
Mouse monoclonal anti-CD31	Novus	Cat#NB600–562, RRID:AB_10002476
Rabbit monoclonal anti-CD38	Abcam	Cat#ab108403; RRID: AB_10890803
Rat monoclonal anti-CD44	BD Biosciences	Cat#550538, RRID:AB_393732
Mouse monoclonal anti-CD45	Thermo Fisher Scientific	Cat#14–9457-82, RRID:AB_11063696
Mouse monoclonal anti-CD45RA	BioLegend	Cat#304102, RRID:AB_314406
Mouse monoclonal anti-CD68	Thermo Fisher Scientific	Cat#14–0688-82, RRID:AB_11151139
Mouse monoclonal anti-CD99	BioLegend	Cat#318002, RRID:AB_604112
Mouse monoclonal anti-E-Cadherin	BD Biosciences	Cat#610182, RRID:AB_397581
Rabbit monoclonal anti-Keratin 19	Cell Signaling Technology	Cat#13092, RRID:AB_2722626
Mouse monoclonal anti-Foxp3	Thermo Fisher Scientific	Cat#14–4777-80, RRID:AB_467555
Rabbit monoclonal anti-Glucagon	Cell Signaling Technology	Cat#8233, RRID:AB_10859908
Rabbit monoclonal anti-Glucose Transporter 1	Abcam	Cat#ab115730, RRID:AB_10903230
Donkey polyclonal anti-Goat IgG	Thermo Fisher Scientific	Cat#PA1–28659, RRID:AB_10981355
Rabbit monoclonal anti-Histone H3	Cell Signaling Technology	Cat#4499, RRID:AB_10544537
Rat monoclonal anti-Histone H3 Phospho (Ser28)	BioLegend	Cat#641002, RRID:AB_1227659
Rabbit monoclonal anti-Insulin	Cell Signaling Technology	Cat#3014, RRID:AB_2126503)
Rabbit polyclonal anti-IAPP	Atlas Antibodies	Cat#HPA053194, RRID:AB_2682076
Mouse monoclonal anti-Ki-67	BD Biosciences	Cat#556003, RRID:AB_396287
Rabbit polyclonal anti-Myeloperoxidase	Dako	Cat#A0398, RRID:AB_2335676
Rabbit monoclonal anti-Nkx6.1	Cell Signaling Technology	Cat#54551, RRID:AB_2722625
Rabbit polyclonal anti-Pancreatic Amylase	Abcam	Cat#ab21156, RRID:AB_446061
Mouse-monoclonal anti-Pancreatic Polypeptide	R and D Systems	Cat#MAB62971, RRID:AB_11127208
Mouse monoclonal anti-cleaved PARP (Poly [ADP-Ribose] Polymerase)	BD Biosciences	Cat#552597, RRID:AB_394438
Goat polyclonal anti Pdx1	R and D Systems	Cat#AF2419, RRID:AB_355257
Mouse monoclonal anti-Proinsulin	Abcam	Cat#ab8301, RRID:AB_306453
Rabbit polyclonal anti-Prohormone Convertase 2	Merck	Cat#AB15601-I, RRID:AB_2722624
Rabbit polyclonal anti-PTPRN	Sigma-Aldrich	Cat#HPA007179, RRID:AB_1079714
Rabbit polyclonal anti-phospho Rb (Ser807/811)	Cell Signaling Technology	Cat#9308, RRID:AB_331472
Rabbit polyclonal anti-Somatostatin	Dako	Cat#A0566, RRID:AB_10013726
Rabbit monoclonal Synaptophysin	Abcam	Cat#ab32127, RRID:AB_2286949
Donkey polyclonal anti-rabbit IgG, Alexa Fluor 488	Thermo Fisher Scientific	Cat#A-21206, RRID:AB_2535792
Biological Samples
Human pancreas sections	nPOD, University of Florida, Gainesville, FL 32610–0275	https://www.jdrfnpod.org
Chemicals, Peptides, and Recombinant Proteins
Hoechst	Sigma Aldrich	Cat #94403–1ML
Iridium intercalator	Fluidigm	Cat #201192B
Critical Commercial Assays
MaxPar X8 Multimetal Labeling Kit	Fluidigm	Cat# 201300
Deposited Data
IMC images	This paper	http://dx.doi.org/10.17632/cydmwsfztj.1
Software and Algorithms
CellProfiler 2.2.0	(Kamentsky et al., 2011)	http://cellprofiler.org
Fiji (ImageJ 1.51u)	(Schindelin et al., 2012)	https://fiji.sc
histoCAT 1.73	(Schapiro et al., 2017)	https://github.com/BodenmillerGroup/histoCAT/releases
Ilastik 1.2.2	(Sommer et al., 2011)	http://ilastik.org
Python 3.4.3	Python Software Foundation	https://www.python.org
R 3.4.1	R Development Core Team	https://www.r-project.org
Tools for IMC data processing	Bodenmiller lab	https://github.com/BodenmillerGroup/imctools
IMC data segmentation pipeline	Bodenmiller lab	https://github.com/BodenmillerGroup/ImcSegmentationPipeline
Custom CellProfiler plugins	Bodenmiller lab	https://github.com/BodenmillerGroup/ImcPluginsCP
Code and guidelines for spillover compensation	Bodenmiller lab	https://github.com/BodenmillerGroup/cyTOFcompensation
IF-to-IMC coordinate transformation	This paper	http://dx.doi.org/10.17632/cydmwsfztj.1
Code related to neighborhood analysis	This paper	http://dx.doi.org/10.17632/cydmwsfztj.1
SCORPIUS 1.0	(Cannoodt et al., 2016)	https://github.com/rcannood/SCORPIUS
Monocle2	(Qiu et al., 2017)	http://cole-trapnell-lab.github.io/monocle-release/docs
R implementation of tSNE	Justin Donaldson	https://cran.r-project.org/web/packages/tsne

CONTACT FOR REAGENT AND RESOURCE SHARING

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Bernd Bodenmiller (bernd.bodenmiller@imls.uzh.ch).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human Organ Donors

Pancreas samples were obtained from nPOD (www.jdrfnpod.org). All procedures were approved by the University of Florida Institutional Review Board (IRB201600029) and the Zurich Cantonal Ethics Committee (2017–02302). Procedures were also approved by the United Network for Organ Sharing (UNOS) according to federal guidelines, and informed consent was obtained from each donor’s legal representative. For each donor, a medical chart review was performed and C-peptide measured (Campbell-Thompson et al., 2012a) with T1D diagnosis determined or confirmed according to the ADA guidelines. Demographics, length of hospitalization, and organ transport duration information was obtained from hospital records or UNOS. Donor pancreata were recovered, shipped in transport media on ice via organ courier to the University of Florida, and processed by a licensed pathology assistant (Campbell-Thompson et al., 2012b). The cases were selected to represent recent onset T1D (< 0.5 years), long-standing T1D (≥ 8 years), and non-diabetes control. Four groups of three donors with different disease duration but matched by age and gender were analyzed. Additional donor information is listed in Table S1.

METHOD DETAILS

Antibody Testing and Validation

All antibodies in the panel (Table 1) were initially tested by IF on lymphoid tissues (spleen, tonsil, and lymph node) for antibodies targeting immune markers and on pancreas for the other markers. Only antibodies that produced expression patterns consistent with the literature and a high signal intensity were conjugated to metals. After metal labeling, all antibodies were tested again by IMC to ensure that antibody specificity was not affected by conjugation. Finally, the entire antibody panel was used to stain two tissue arrays: an array containing large cores from six different tissues and a microarray containing 78 cores representing 20 different tissues from healthy subjects and cancer patients. In this way, each antibody could be positively validated on tissues known to express the target antigen and negatively validated on tissues known to be negative for that antigen. The main observations used to validate each antibody are listed in Table S2.

Antibody Conjugation and Titration

Antibodies were conjugated to metals using the MaxPar X8 Multimetal Labeling Kit (Fluidigm) according to the manufacturer’s protocol. To determine optimal antibody concentrations, the antibody panel was divided in “odd” and “even” channels, depending on the atomic mass of the metal tag, and two-fold serial antibody dilutions were performed. Sections of human pancreas and lymphoid tissue (spleen, tonsil, pancreatic lymph node) were divided in four areas, and each area was stained with a different antibody dilution. Separate slides were used for antibodies conjugated to odd and even metal isotopes. After IMC measurement, each channel was examined visually and the dilution that produced the highest signal-to-noise ratio was selected. In the few cases where notable spillover was detected in the adjacent (+1 and −1) channels, a higher antibody dilution was selected.

Tissue Staining

Two sections per pancreas, from head, body, or tail, were stained with our full antibody panel (Table 1). Slides were incubated for 1 h at 60 °C i n a dry oven, deparaffinized in fresh Xylol, and rehydrated through a graded alcohol series. Antigen retrieval was performed in a decloaking chamber (Medite) for 30 min at 96 °Cin Tris-EDTA, pH 9.2. After blocking in buffer containing 10% normal horse serum, slides were incubated overnight at 4 °C with three primary antibodies: rabbit anti-glucagon-¹⁵⁶Gd, rabbit anti-synaptophysin-¹⁶⁰Gd, and goat anti-Pdx1 (unconjugated). The next day, slides were incubated with anti-rabbit IgG-Alexa Fluor 488 and anti-Goat IgG-158Gd secondary antibodies for 1 h at room temperature, followed by Hoechst dye counterstaining (5 min at room temperature). Slide scanner imaging was performed at this stage (see Selection of Areas of Interest below). Two successive overnight incubations at 4 °C were then performed for labeling with the remainder of primary antibodies (13 primary antibodies on the first night and 19 on the second night). Finally, slides were counterstained with iridium intercalator for 5 min at room temperature, quickly dipped in doubly distilled H₂O, and immediately dried with pressurized air.

Selection of Areas of Interest

Pancreas sections stained for glucagon and synaptophysin and counterstained with Hoechst were imaged with a slide scanner (Zeiss Axio Scan.Z1) at 20X magnification. Areas of interest were chosen manually based on islet size (islets representing the complete spectrum of size observed but with an enrichment for large islets) and islet position (to ensure distribution of sampling over the whole section) (Figure S1A). Area size was defined as the axis-aligned minimum bounding box containing all islets in the area, expanded by 100 µm in all directions. The coordinates of areas of interest were recorded using ImageJ and transformed to IMC coordinates using a custom Python script (see Key Resources Table).

IMC Image Acquisition

Data acquisition was performed on a Helios time-of-flight mass cytometer (CyTOF) coupled to a Hyperion Imaging System (Fluidigm). Prior to laser ablation, optical images of slides were acquired using the Hyperion software, and the areas to ablate were selected as described above. Laser ablation was performed at a resolution of approximately 1 µm and a frequency of 200 Hz. Slides were acquired in a randomized order over 25 consecutive days. To ensure performance stability, the machine was calibrated daily with a tuning slide spiked with five metal elements (Fluidigm). In total, 851 image stacks from 24 pancreas sections were acquired. Three were removed due to data corruption, one due to poor image quality, and two because the areas did not contain any islets. All analyses were performed on the 845 remaining image stacks.

QUANTIFICATION AND STATISTICAL ANALYSIS

IMC Data Preprocessing

Text format (.txt) files generated during data acquisition were converted to tiff images using a custom Python script (https://github.com/BodenmillerGroup/imctools). For visualization, false color images were prepared from tiff images using Fiji (https://fiji.sc).

Spillover Compensation

Although spectral overlap between pure metal isotopes is very limited, channel crosstalk caused by isotopic impurity, oxide formation, and imprecision in ion detection still exist (Chevrier et al., 2018). To measure crosstalk from one channel to the other, we generated a slide spotted with the same metals used for antibody conjugation (one metal isotope per spot). With this strategy, each metal can be measured individually while having all channels open for detection, thus allowing for spillover quantification across all measured channels. The slide was measured three times during the 25 days of IMC data acquisition, and an average spillover matrix was generated from the resulting data (Figure S1B). This matrix was used to correct for spillover in our dataset (Figure S1C), as described (Chevrier et al., 2018).

Image Segmentation

We generated cell segmentation masks in two steps: First, we used the interactive machine learning software ilastik to classify pixels as nucleus, cytoplasm/membrane, and background compartments. Second, we imported the probability maps generated by ilastik into CellProfiler to identify individual nuclei, define cell borders, and generate cell masks. We adopted a similar approach to segment islets (Figure S3).

Pixel Classification

For cell segmentation, the pixels in each image were defined as belonging to the nucleus, cytoplasm, or background compartment using the pixel classification module of the open source software ilastik (Sommer et al., 2011) as described (Schulz et al., 2018) (https://github.com/BodenmillerGroup/ImcSegmentationPipeline). Training of the Random Forest classifier was performed on 125 × 125 pixels substacks generated from the original images and containing only informative markers (i.e., H3, CDH, CD99, CD44, MPO, SYP, KRT19, CD45, CD31, SMA, and Iridium). This allowed for the generation of maps that integrate for each pixel the probability of belonging to each of three compartments, as shown in Figure 1B (center). Based on the trained classifier, probability maps were generated for the whole dataset and exported as tiff files in batch mode.

Cell Segmentation

CellProfiler (Kamentsky et al., 2011) was subsequently used to define cell masks and quantify marker expression. To define cell borders, nuclei were first identified as primary objects based on ilastik probability maps and expanded through the cytoplasm compartment until either a neighboring cell or the background compartment was reached. Cell masks were generated for identification of single cells and used to extract single-cell information (marker abundance, spatial and neighborhood data) from the original images. The results were exported as csv tables for further analysis.

Islet and Blood Vessel Segmentation

Islet and blood vessel segmentation was performed in a similar manner as cell segmentation with the following modifications: i) for ilastik pixel classification, images were divided into islet, blood vessels, and “other” (that is non-islet, non-vessel) compartments based on substacks containing informative markers (i.e., SYP, CD99, CD31, CD45, AMY2A, KRT19, and Iridium); ii) in CellProfiler, islets and blood vessels were directly defined as primary objects. Distance to islet rim was measured by combining islet and cell masks, using a custom CellProfiler plugin (https://github.com/BodenmillerGroup/ImcPluginsCP).

Cell Type Definition

The object classification module of ilastik was used to define cell types (Figure S3). First, binary cell masks and substacks containing relevant markers (i.e., CD99, SYP, AMY2A, KRT19, CD44, CDH, CD45, CD3, CD8, CD20, MPO, CD68, and SMA) were loaded into ilastik and a classifier was trained in order to divide the cells in four main categories (islet, immune, exocrine, and “other”). The resulting object predictions were exported as tiff images and imported into CellProfiler to generate cell masks specific to each cell category. The process was iterated to subdivide each category into cell types. For islet cell type definition, masks containing only cells defined in the first step as belonging to the islet category were imported into ilastik, and a new training was performed to attribute α, β, δ, and γ cell types based on the expression of islet-expressed markers (i.e., INS, PIN, GCG, SST, PPY, PDX1, NKX6–1, IAPP, PCSK2, CD99, and SYP). The same approach was adopted to define B cells, cytotoxic and helper T cells, monocytes/macrophages, and neutrophils among immune cells (markers used: CD45, CD3e, CD4, CD8a, CD20, CD68, MPO, and CD45RA). In addition, exocrine cells were divided into acinar and ductal cells (markers used: KRT19, AMY2A, CD44, and CDH), and “other” cells were divided into three cell types: endothelial (also includes peri-endothelial cells), stromal, and unknown (markers used: SMA, CD31, CD68, and CDH).

Pseudotime Analysis

Pseudotime analysis was performed using the trajectory inference algorithm SCORPIUS (Cannoodt et al., 2016), and results were confirmed using the Monocle2 DDRTree algorithm (Qiu et al., 2017). Only islets containing 15 cells or more were considered. Islet expression profiles and pseudotime ordering were based on the expression of islet-expressed markers averaged over the islet sectional area. All measured islet markers were considered (INS, PIN, 814 GCG, SST, PDX1, NKX6–1, IAPP, PCSK2, SLC2A1, PTPRN, SYP, and CD99), except for PPY. The presence of a γ cell-rich pancreas head section in one donor with long-duration T1D tended to confound the algorithms, resulting in classification of the very few γ cell-rich islets from non-diabetic donors into the long-duration (pseudostage 3) category. Raw IMC counts averaged over islet areas were used as an input.

Neighborhood Analysis

Neighborhood analysis was performed using R and histoCAT (Schapiro et al., 2017). Custom scripts were used to import cell type and pseudotime information into histoCAT (see Key Resources Table). An expansion of 3 pixels was used to detect cell neighbors. The presence of significant cell type association or avoidance was determined using 999 random permutations of cell type labels and a significance threshold of P < 0.01. All other parameters and methods were chosen as suggested in the original publication (Schapiro et al., 2017).

Data Analysis

Data were analyzed and graphed using R and histoCAT (Schapiro et al., 2017). In particular, an R implementation of tSNE (https://cran.r-project.org/web/packages/tsne) was used to generate tSNE maps (van der Maaten and Hinton, 2008). Unless otherwise stated, raw IMC counts were 99^th-percentile-normalized and scaled from 0 to 1 (scaled counts).

Statistics

Statistical tests were performed using R. Statistical parameters are reported in the figure legends. For pairwise group comparisons, we systematically used a non-parametric statistical test (unpaired, two-tailed Mann-Whitney U test) and applied the Holm post-hoc adjustment method to correct for multiple comparisons. Overall, the P value significance threshold was defined as 0.01. In text and figures, N represents the number of donors and n represents either the number of cells or the number of islets (as indicated in figure legends). In histograms, bars represent the means and error bars represent standard errors of the mean. Correlation analyses in Figure 4H were performed using Pearson’s r correlation coefficient. As mentioned above (see the IMC Image Acquisition paragraph), six out of 851 IMC images were excluded because of data corruption, poor image quality, or absence of islets. No other data or values were excluded from our analyses.

HIGHLIGHTS.

• Highly multiplexed imaging of pancreas tissues from human donors with T1D.

• Islet evolution through T1D progression inferred from snapshot data.

• The phenotype of β cells is altered prior to β cell destruction.

• T cell recruitment depends on both disease stage and islet profile.