Identification of type 2 diabetes- and obesity-associated human β-cells using deep transfer learning
Curation statements for this article:-
Curated by eLife
eLife assessment
This is a useful study that used DEGAS, a deep transfer learning tool, to identify distinct pancreatic beta cell subpopulations that could be associated with type 2 diabetes (T2D) and/or obesity status. The data supporting the authors' findings is solid and demonstrates that DEGAS will be a helpful tool for analyzing cell-specific transcriptomic phenotypes. This study will be of interest to researchers studying the genetics of T2D.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
Abstract
Diabetes affects >10% of adults worldwide and is caused by impaired production or response to insulin, resulting in chronic hyperglycemia. Pancreatic islet β-cells are the sole source of endogenous insulin and our understanding of β-cell dysfunction and death in type 2 diabetes (T2D) is incomplete. Single-cell RNA-seq data supports heterogeneity as an important factor in β-cell function and survival. However, it is difficult to identify which β-cell phenotypes are critical for T2D etiology and progression. Our goal was to prioritize specific disease-related β-cell subpopulations to better understand T2D pathogenesis and identify relevant genes for targeted therapeutics. To address this, we applied a deep transfer learning tool, DEGAS, which maps disease associations onto single-cell RNA-seq data from bulk expression data. Independent runs of DEGAS using T2D or obesity status identified distinct β-cell subpopulations. A singular cluster of T2D-associated β-cells was identified; however, β-cells with high obese-DEGAS scores contained two subpopulations derived largely from either non-diabetic or T2D donors. The obesity-associated non-diabetic cells were enriched for translation and unfolded protein response genes compared to T2D cells. We selected DLK1 for validation by immunostaining in human pancreas sections from healthy and T2D donors. DLK1 was heterogeneously expressed among β-cells and appeared depleted from T2D islets. In conclusion, DEGAS has the potential to advance our holistic understanding of the β-cell transcriptomic phenotypes, including features that distinguish β-cells in obese non-diabetic or lean T2D states. Future work will expand this approach to additional human islet omics datasets to reveal the complex multicellular interactions driving T2D.
Article activity feed
-
-
-
Author response:
Public Reviews:
Reviewer #1 (Public Review):
Summary:
In this manuscript, Roy et al. used the previously published deep transfer learning tool, DEGAS, to map disease associations onto single-cell RNA-seq data from bulk expression data. The authors performed independent runs of DEGAS using T2D or obesity status and identified distinct β-cell subpopulations. β-cells with high obese-DEGAS scores contained two subpopulations derived largely from either non-diabetic or T2D donors. Finally, immunostaining using human pancreas sections from healthy and T2D donors validated the heterogeneous expression and depletion of DLK1 in T2D islets.
Strengths:
(1) This meta-analysis of previously published scRNA-seq data using a deep transfer learning tool.
(2) Identification of novel beta cell subclusters.
(3) Identified a relatively …
Author response:
Public Reviews:
Reviewer #1 (Public Review):
Summary:
In this manuscript, Roy et al. used the previously published deep transfer learning tool, DEGAS, to map disease associations onto single-cell RNA-seq data from bulk expression data. The authors performed independent runs of DEGAS using T2D or obesity status and identified distinct β-cell subpopulations. β-cells with high obese-DEGAS scores contained two subpopulations derived largely from either non-diabetic or T2D donors. Finally, immunostaining using human pancreas sections from healthy and T2D donors validated the heterogeneous expression and depletion of DLK1 in T2D islets.
Strengths:
(1) This meta-analysis of previously published scRNA-seq data using a deep transfer learning tool.
(2) Identification of novel beta cell subclusters.
(3) Identified a relatively innovative role of DLK1 in T2D disease progression.
We thank the reviewer for their constructive critiques and positive feedback. We hope to further apply deep transfer learning tools in future scRNA-seq meta-analyses.
Weaknesses:
(1) There is little overlap of the DE list of bulk RNA-seq analysis in Figure 1D and 1E overlap with the DE list of pseudo-bulk RNA-seq analysis of all cells in Figure S2C.
We thank the reviewer for this insightful thought and plan to perform additional analyses and comparisons to address this comment.
(2) The biological meaning of "beta cells had the lowest scores compared to other cell types" is not clear.
We agree with the reviewer and will amend this statement to clarify in the revised manuscript. In summary, the relatively lower T2D-DEGAS scores for beta cells overall compared to all other cell types (alpha cells, acinar cells, etc) reflects the fact that in T2D, beta cell-specific genes can be downregulated. This is also possibly due to beta cell loss in T2D and would be reflected in bulk islet RNAseq data. This affects the DEGAS model which is reflected in the scores of all cells in the scRNA-seq data (Fig 3A). For this reason, subsetting the beta cells and replotting them on their own (Fig 4B) is an important step to identify relative differences in DEGAS scores between different subsets of beta cells.
(3) The figures and supplemental figures were not cited following the sequence, which makes the manuscript very difficult to read. Some supplemental figures, such as Figures S1C-S1D, S2B-S2E, S3A-S3B, were not cited or mentioned in the text.
We apologize and thank the reviewer for pointing out these errors. All of the annotated errors will be amended in the revised manuscript.
(4) In Figure 7, the current resolution is too low to determine the localization of DLK1.
We will include the original highest-resolution confocal images in our resubmission. We will also improve the color combination to improve visibility of colocalization of DLK1 with Insulin.
Reviewer #2 (Public Review):
Summary:
The manuscript by Gitanjali Roy et al. applies deep transfer learning (DEGAS) to assign patient-level disease attributes (metadata) to single cells of T2D and non-diabetic patients, including obese patients. This led to the identification of a singular cluster of T2D-associated β-cells; and two subpopulations of obese- β-cells derived from either non-diabetic or T2D donors. The objective was to identify novel and established genes implicated in T2D and obesity. Their final goal is to validate their findings at the protein level using immunohistochemistry of pancreas tissue from non-diabetic and T2D organ donors.
Strengths:
This paper is well-written, and the findings are relevant for β-cell heterogeneity in T2D and obesity.
We thank the reviewer for their constructive critiques and positive feedback. We believe this study can improve our understanding β-cell heterogeneity in the context of T2D and obesity.
Weaknesses:
The validation they provide is not sufficiently strong: no DLK1 immunohistochemistry is shown of obese patient-derived sections. Additional presumptive relevant candidates from this transcriptomic analysis should be screened for, at the protein level.
Thank the reviewer for this suggestion. We are planning to perform new immunostaining of DLK1 in human pancreas tissue sections from non-diabetic lean, non-diabetic obese, T2D lean, and T2D obese donors. We also note that Table S6 contains the patient metadata for the pancreas samples we show in the current manuscript. Two of the T2D donors have BMI > 30 (obese). However, the non-diabetic donors have BMI between 26-29. Our new planned studies should address the question of differential DLK1 expression / beta cell heterogeneity in the context of both diabetes and obesity.
-
eLife assessment
This is a useful study that used DEGAS, a deep transfer learning tool, to identify distinct pancreatic beta cell subpopulations that could be associated with type 2 diabetes (T2D) and/or obesity status. The data supporting the authors' findings is solid and demonstrates that DEGAS will be a helpful tool for analyzing cell-specific transcriptomic phenotypes. This study will be of interest to researchers studying the genetics of T2D.
-
Reviewer #1 (Public Review):
In this manuscript, Roy et al. used the previously published deep transfer learning tool, DEGAS, to map disease associations onto single-cell RNA-seq data from bulk expression data. The authors performed independent runs of DEGAS using T2D or obesity status and identified distinct β-cell subpopulations. β-cells with high obese-DEGAS scores contained two subpopulations derived largely from either non-diabetic or T2D donors. Finally, immunostaining using human pancreas sections from healthy and T2D donors validated the heterogeneous expression and depletion of DLK1 in T2D islets.
Strengths:
(1) This meta-analysis of previously published scRNA-seq data using a deep transfer learning tool.
(2) Identification of novel beta cell subclusters.
(3) Identified a relatively innovative role of DLK1 in T2D disease …
Reviewer #1 (Public Review):
In this manuscript, Roy et al. used the previously published deep transfer learning tool, DEGAS, to map disease associations onto single-cell RNA-seq data from bulk expression data. The authors performed independent runs of DEGAS using T2D or obesity status and identified distinct β-cell subpopulations. β-cells with high obese-DEGAS scores contained two subpopulations derived largely from either non-diabetic or T2D donors. Finally, immunostaining using human pancreas sections from healthy and T2D donors validated the heterogeneous expression and depletion of DLK1 in T2D islets.
Strengths:
(1) This meta-analysis of previously published scRNA-seq data using a deep transfer learning tool.
(2) Identification of novel beta cell subclusters.
(3) Identified a relatively innovative role of DLK1 in T2D disease progression.
Weaknesses:
(1) There is little overlap of the DE list of bulk RNA-seq analysis in Figure 1D and 1E overlap with the DE list of pseudo-bulk RNA-seq analysis of all cells in Figure S2C.
(2) The biological meaning of "beta cells had the lowest scores compared to other cell types" is not clear.
(3) The figures and supplemental figures were not cited following the sequence, which makes the manuscript very difficult to read. Some supplemental figures, such as Figures S1C-S1D, S2B-S2E, S3A-S3B, were not cited or mentioned in the text.
(4) In Figure 7, the current resolution is too low to determine the localization of DLK1.
-
Reviewer #2 (Public Review):
Summary:
The manuscript by Gitanjali Roy et al. applies deep transfer learning (DEGAS) to assign patient-level disease attributes (metadata) to single cells of T2D and non-diabetic patients, including obese patients. This led to the identification of a singular cluster of T2D-associated β-cells; and two subpopulations of obese- β-cells derived from either non-diabetic or T2D donors. The objective was to identify novel and established genes implicated in T2D and obesity. Their final goal is to validate their findings at the protein level using immunohistochemistry of pancreas tissue from non-diabetic and T2D organ donors.
Strengths:
This paper is well-written, and the findings are relevant for β-cell heterogeneity in T2D and obesity.
Weaknesses:
The validation they provide is not sufficiently strong: no DLK1 …
Reviewer #2 (Public Review):
Summary:
The manuscript by Gitanjali Roy et al. applies deep transfer learning (DEGAS) to assign patient-level disease attributes (metadata) to single cells of T2D and non-diabetic patients, including obese patients. This led to the identification of a singular cluster of T2D-associated β-cells; and two subpopulations of obese- β-cells derived from either non-diabetic or T2D donors. The objective was to identify novel and established genes implicated in T2D and obesity. Their final goal is to validate their findings at the protein level using immunohistochemistry of pancreas tissue from non-diabetic and T2D organ donors.
Strengths:
This paper is well-written, and the findings are relevant for β-cell heterogeneity in T2D and obesity.
Weaknesses:
The validation they provide is not sufficiently strong: no DLK1 immunohistochemistry is shown of obese patient-derived sections. Additional presumptive relevant candidates from this transcriptomic analysis should be screened for, at the protein level.
-
