Subcellular Feature-Based Classification of α and β Cells Using Soft X-ray Tomography

Abstract

The dysfunction of α and β cells in pancreatic islets can lead to diabetes. Many questions remain on the subcellular organization of islet cells during the progression of disease. Existing three-dimensional cellular mapping approaches face challenges such as time-intensive sample sectioning and subjective cellular identification. To address these challenges, we have developed a subcellular feature-based classification approach, which allows us to identify α and β cells and quantify their subcellular structural characteristics using soft X-ray tomography (SXT). We observed significant differences in whole-cell morphological and organelle statistics between the two cell types. Additionally, we characterize subtle biophysical differences between individual insulin and glucagon vesicles by analyzing vesicle size and molecular density distributions, which were not previously possible using other methods. These sub-vesicular parameters enable us to predict cell types systematically using supervised machine learning. We also visualize distinct vesicle and cell subtypes using Uniform Manifold Approximation and Projection (UMAP) embeddings, which provides us with an innovative approach to explore structural heterogeneity in islet cells. This methodology presents an innovative approach for tracking biologically meaningful heterogeneity in cells that can be applied to any cellular system.

Keywords: 3D cell mapping; Uniform Manifold Approximation and Projection (UMAP); cryogenic fluorescence microscopy; machine learning; pancreatic islets; soft X-ray tomography; α cells; β cells.

Figure 1

3D reconstruction and quantitative analysis of α and β cell morphology. (A) Orthoslice showing the XY plane through the soft X-ray tomogram of representative α and β cells (α_3 and β_6, respectively). Cell constituents and organelles are distinguished from one another based on their LAC values and are identified as follows: nucleus–blue arrowhead, mitochondria–pink arrowhead, glucagon vesicles–red arrowhead, and insulin vesicles–green arrowhead. The overall LAC value range of the orthoslice is between 0.15 and 0.4 μm⁻¹ to optimize contrast. Scale bar: 2 μm. (B) 3D reconstruction of the representative α and β cells (α_3 and β_6, respectively). In detail, the reconstruction shows the nucleus (blue), mitochondria (pink), glucagon vesicles ((left), in red), insulin vesicles ((right), in green), and plasma membrane (gray). (C) Cellular volume of both cell types, showing a significantly higher volume (*** p < 0.001) for β cells (1191 ± 277 μm³) compared with α cells (579 ± 247 μm³). (D) Nuclear volume of both cell types showed no significant difference (p = 0.76). (E) Comparison between mean nuclear occupancy for each cell type, with a significant increase (*** p < 0.001) in percentage occupancy of the nucleus for α cells (21 ± 5%) compared with β cells (10 ± 3%). (F) Number of insulin vesicles normalized by cytosolic volume indicating a significantly higher number of vesicles (* p = 0.03) per cytosolic μm³ for α cells (3.3 ± 1.4 vesicles/μm³) compared with β cells (2 ± 0.6 vesicles/μm³). (G) Plot of mean vesicle diameters of α and β cell vesicles demonstrating a higher vesicle diameter (*** p < 0.001) for α cell vesicles (212 ± 21 nm) compared with β cell vesicles (163 ± 13 nm). (H) Mean Vesicle LAC for secretory vesicles of α and β cells showing a significantly higher mean LAC (** p = 0.003) for α cell vesicles (0.37 ± 0.03 μm⁻¹) compared with β cell vesicles (0.33 ± 0.02 μm⁻¹). Error bars in each plot are representative of the standard deviation. Welch’s t-test was used as a statistical test. n = 8 for α cells (red) and n = 7 for β cells (green).

Figure 2

3D reconstruction and quantitative analysis of pooled insulin and glucagon vesicles. (A) (left) XY Orthoslice through the SXT of representative α and β cells. Glucagon vesicles (red arrowheads) and insulin vesicles (green arrowheads) can be identified based on their high LAC values. The overall LAC value in the orthoslice is thresholded between 0.15 and 0.40 μm⁻¹ (scale bar: 0.5 μm). (right) 3D reconstruction of a section of representative α and β cells (α_3 and β_6, respectively). In detail, the reconstruction shows glucagon vesicles ((top), in red) and insulin vesicles ((bottom), in green), and plasma membrane (gray). (B) Histogram showing the size distribution of glucagon and insulin vesicles. The vesicles for each cell type are pooled together and show a significantly higher diameter (**** p < 0.0001) for insulin vesicles (194.2 ± 49 nm, green dotted line), compared with glucagon vesicles (157 ± 35 nm, red dotted line). (C) Histogram showing LAC distribution of glucagon and insulin vesicles demonstrating a significantly higher mean vesicle LAC values (**** p < 0.0001) for insulin vesicles (0.37 ± 0.04 μm⁻¹, red dotted line), compared with glucagon vesicles (0.33 ± 0.03 μm⁻¹, green dotted line). (B,C) n = 10,694 for glucagon vesicles (red) and n = 14,690 for insulin vesicles (green). Welch’s t-test was used as a statistical test.

Figure 3

Description and comparison of LAC-based parameters between insulin and glucagon vesicles (A) (top) XY Orthoslice through SXT of representative α cell (α_3). Scale bar: 0.5 μm. (middle) 3D reconstruction of a representative glucagon vesicle (red). (bottom) Histogram displaying the LAC distribution of the glucagon vesicle picture in the top and middle panels showing a mean LAC value of 0.34 μm⁻¹ for the vesicle. (B) (top) XY Orthoslice through SXT of representative β-cell (β_6). Scale bar: 500 nm. (middle) 3D reconstruction of a representative insulin vesicle (green). (bottom) Histogram displaying the LAC distribution of the insulin vesicle picture in the (top) and (middle) panels, showing a mean LAC value of 0.318 μm⁻¹ for the vesicle. (C) (top) A comparison of vesicle LAC parameters (minimum LAC, 25th quantile LAC, mean LAC, 75th quantile LAC, maximum LAC) between glucagon vesicles (red) and insulin vesicles (green) showing significantly higher values (**** p < 0.0001; one-way ANOVA with Bonferroni’s correction) for glucagon vesicles for all displayed parameters compared with insulin vesicles. (bottom) LAC histogram curve for a sample vesicle, with arrows indicating the value being compared in the (top) panel. (D) Plot showing a significantly higher (**** p < 0.0001; Welch’s t-test) standard deviation for glucagon vesicles (red) compared with insulin vesicles (green). Error bars in all plots are representative of the standard deviation. n = 10,694 for glucagon vesicles (red) and n = 14,690 for insulin vesicles (green).

Figure 4

Overview of machine learning strategy. (A) Data pre-processing for insulin and glucagon vesicles from β and α cells is conducted. A final vesicle feature matrix, including group labels (denoting which cell a vesicle is from) for vesicles, is used as input for machine learning. (B) Train/test split for grouping vesicles from α and β cells. The process of model building is described, with Leave One Group Out cross-validation used to estimate the performance of predicting vesicle identity from unseen cells. Model building and testing are repeated over 56 combinations to understand variability in performance.

Figure 5

Representing vesicle feature importances in UMAP embeddings. (A) Representative feature importances from each ML model listed in order of accuracy. The radius of circles is scaled to the magnitude of the permutation feature importances. Since the LAC standard deviation in the XGBoost model had the highest overall importance magnitude, the other parameters are scaled to it. In general, LAC mean, LAC standard deviation, and diameter seem to be the most important representative parameters. (B) UMAP embedding of vesicles colored by vesicle identity. Semi-distinct clusters of insulin and glucagon vesicles can be observed. (C) Vesicles colored by cellular origin. The overall trends in the pooled vesicle UMAP space do not seem to be driven by cell-dependent effects. (D) Embeddings colored by vesicle feature values. Gradients of LAC mean, standard deviation, and diameter correspond to regions of insulin and glucagon vesicles. The grouping of heterogeneous vesicle subpopulations can also be visualized.