Modeling type 1 diabetes progression using machine learning and single-cell transcriptomic measurements in human islets

Abstract

Type 1 diabetes (T1D) is a chronic condition in which beta cells are destroyed by immune cells. Despite progress in immunotherapies that could delay T1D onset, early detection of autoimmunity remains challenging. Here, we evaluate the utility of machine learning for early prediction of T1D using single-cell analysis of islets. Using gradient-boosting algorithms, we model changes in gene expression of single cells from pancreatic tissues in T1D and non-diabetic organ donors. We assess if mathematical modeling could predict the likelihood of T1D development in non-diabetic autoantibody-positive donors. While most autoantibody-positive donors are predicted to be non-diabetic, select donors with unique gene signatures are classified as T1D. Our strategy also reveals a shared gene signature in distinct T1D-associated models across cell types, suggesting a common effect of the disease on transcriptional outputs of these cells. Our study establishes a precedent for using machine learning in early detection of T1D.

Keywords: autoantibody-positive; human islets; machine learning; single-cell RNA-seq; type 1 diabetes.

Graphical abstract

Figure 1

scRNA-seq reveals the cell populations of the human pancreatic islets in CTL, AAb+, and T1D donors (A) The complete workflow depicting the scRNA-seq and machine learning workflow using human pancreatic islet tissue samples. (B) Pie chart showing the number of cells and donor distribution across different biological conditions. (C) Uniform manifold approximation and projection (UMAP) plot showing the scSorter cell classification of islet cells. (D) Stacked bar chart showing the percentage-wise distribution of cell types across AAb+, control, and T1D donors. (E) Multiple feature plots UMAPs depicting the validation of cell-type-specific expression of marker genes. Acinar cells (PRSS1 high), alpha cells (GCG high), beta cells (INS high), delta cells (SST high), ductal cells (KRT19 high), endothelial cells (VWF high), epsilon cells (GHRL high), immune (NCF2 high), PP cells (PPY high), and stellate cells (COL1A1 high).

Figure 2

Classification performance of machine learning model on scRNA-seq islet data (A) A schematic workflow of XGBoost and performance. The machine-learning-based XGBoost model was built for gene selection and classification. The dotted lines show the training and testing procedures, where T denotes the gradient boosting tree models. The double lines show 100 repetitions of the entire workflow. (B) Boxplots depicting a pairwise comparison of the XGBoost method across all cells (unannotated) in the dataset. (C) Performance of XGBoost across major cell types for T1D vs. CTL comparison using boxplots. (D) Performance of XGBoost across major cell types for T1D vs. AAb+ comparison using boxplots. (E) Performance of XGBoost across major cell types for AAb+ vs. CTL comparison using boxplots.

Figure 3

Top-ranked genes selected from the machine learning model and pathway enrichment analysis (A) GO cluster analysis showing pathways based on genes obtained from T1D vs. CTL comparison across unannotated all cells. (B) Top 20 KEGG pathways based on ranked genes obtained from T1D vs. CTL comparison across all cells (unannotated) and different cell types (annotated) (Table S3). (C) UpSetR chart showing the common and unique count of genes across annotated cells in T1D vs. CTL comparison. (D) Pathways based on shared RRA_combined gene list in T1D vs. CTL annotated cells. (E) UpSetR chart showing the common and unique count of genes across annotated cells in T1D vs. AAb+ comparison. (F) UpSetR chart showing the common and unique count of genes across annotated cells in AAb+ vs. CTL comparison.

Figure 4

Expression of HLA-1 genes in beta cells across healthy and T1D donors (A) Comparison of significant KEGG pathways (FDR < 0.05) obtained from different pairwise classifiers. (B) Significant KEGG pathways for T1D vs. CTL (FDR < 0.05). (C–E) The top 3 modules were obtained from the PPI network using the MCODE algorithm. (F) Average expression of beta cells in non-diabetic controls and T1D donors.

Figure 5

Prediction of AAb+ cells using trained T1D-CTL classifier across major cell types (A) Distribution of cells misclassified as T1D in different cell types. (B) Comparing the average expression of HLA-I genes among beta cells from AAb+ donors classified as T1D with other conditions. (C) Comparing the average expression of HLA-I genes among alpha cells from AAb+ donors classified as T1D with other conditions. (D) Comparing the average expression of HLA-I genes among all cell types from two AAb+ donors classified as T1D with other conditions.

Figure 6

Prediction of AAb+ cells using trained T1D-CTL classifier across all cells together (A) Distribution of AAb+ cells predicted as T1D using trained T1D-CTL classifier for all cells. (B) Selection frequency of genes from HLA-I and HLA-II class. (C) Selection frequency of genes from non-HLA class relevant to T1D. (D) Comparing the average expression of HLA-I genes among all cells from HPAP092 AAb+ donor classified as T1D. (E) Comparing the average expression of HLA-II genes among all cells from HPAP092 AAb+ donor classified as T1D. (F) Comparing the average expression of non-HLA genes among all cells from HPAP092 AAb+ donor classified as T1D.