Introduction

Diabetes plagues approximately 537 million people each year throughout the world, with that number expected to increase by 50% over the next 20 years1. Prolonged exposure to hyperglycemia leads to a series of associated long-term complications including neuropathy, nephropathy, cardiovascular disease, stroke, and retinopathy which reduce life expectancy2. Current diabetes treatments (i.e. incretin-based therapeutics) improve glycemic control through a combination of improving beta cell function and insulin sensitivity3. Despite the combination of safety and efficacy profiles superior to insulin and insulin secretagogues, GLP-1 receptor agonism is associated with undesirable side effects that risk non-compliance4. Therefore, iterative therapeutic options are needed to fully address this serious public health issue throughout the world.

While diabetes is characterized by both beta cell failure and insulin resistance in peripheral tissues, the central nervous system is an attractive target for long term treatment of diabetes since brain systems should be intact in patients unlike the pancreas, liver, and muscle5. Indeed, the brain can sense changes in blood glucose and generate responses to both raise and lower blood glucose, when necessary6. Evidence suggests that the current generation of diabetes treatments, in part, act to restore normal function of brain systems associated with metabolic function7. Central nervous system action is crucial to the metabolic benefits of incretin-based therapeutics8. However, glucoregulatory benefits due to central actions are secondary to energy balance impacts.

Dysfunctions in brain systems have been previously observed in obesity and diabetes models, and it has been established that treatment in the brain can have therapeutic benefits in pre-clinical models of obesity and diabetes. There is evidence for inflammation in response to obesity in both rodent and humans that could explain dysfunction of brain function that underlies diabetes9. ER stress has also been noted in hypothalamic brain areas critical for energy homeostasis, and targeting ER stress in the brain does improve glucose dyshomeostasis and obesity in pre-clinical models10. While there is evidence for central nervous system insulin resistance in diabetes, the relative contribution is not clear for peripheral glycemic control11. However, the therapeutic potential of intact glucose control mechanisms in diabetes driven by the central nervous system is clear. ICV leptin administration can restore glycemic function in streptozotocin-induced diabetes, indicating insulin-independent glucose lowering mechanisms can function in the absence of beta cells12. As further evidence of the therapeutic potential for targeting the brain for diabetes, a single dose of Fibroblast growth factor 1 (FGF1) administration in both mice and rats leads to remission in both leptin-deficient and diet-induced obese models of diabetes13. However, we still do not fully understand how to leverage these mechanisms for therapeutic benefits in diabetic and obese patients.

Due to the sheer complexity of the neural systems that have only been partially identified, we still do not understand exactly how the brain can be harnessed for diabetes treatment. This complexity is underscored by the fact that brain regions contain a heterogeneous mix of cell types that perform distinct functions and respond to specialized sets of stimuli. Furthermore, isolated cells are only one node along neurocircuits that can involve complex regulation throughout the brain for the purpose of tight regulation. One way to better understand the changes that can be made to treat diabetes is through the identification of gene co-expression modules. Using this bioinformatic process, clusters of genes can be identified within key subsets of cells that are associated with remission from a disease. These represent genes or pathways that could be directly targeted therapeutically in the future or used as biomarkers for impaired functioning in patients. Co-expressed gene modules are often shared within cells of a certain type, and particularly those responsive to a given treatment. In both Alzheimer’s disease14 and cancer15, this has been used to identify targetable cell types not previously realized for treatment, and similar types of strategies have been used in autoimmune diseases such as crohn’s disease, inflammatory bowel, and lupus16,17,18. In addition, investigators have used these strategies in diabetes, but not in the brain19,20. However, this could be particularly valuable for neuroscience, as the brain processes function along a neural network that involves numerous cell types that could be targetable for future interventions, but our understanding of the neural network, outside of several critical nodes (e.g. arcuate nucleus), is very limited. While there may be some regions that can be targeted as a whole for a function in metabolic dysfunction, these cells communicate through cells that are intermingled in brain areas that participate in different functions. Hence, Indeed, this type of transfer learning strategies have been used in development across populations in the brain in development of neurological disiorders21 and presumably will also be valuable for identifying neurons along a circuit. Advancing tools to establish cell types and markers within new regions throughout the brain that are part of the neural network activated by known players may expedite progress towards correcting these neural systems in patients.

We first performed gene co-expression network (GCN) mining on a previously published single cell sequencing dataset in the mediobasal hypothalamus of mice in diabetes remission following FGF1 administration. Our goal is to use these data, not to identify mechanisms driven by FGF1 administration (e.g. median eminence and Arcuate nucleus), but to identify overall modules in the brain that are associated with normalization of glucose homeostasis. Here we use two datasets to study the gene regulatory relationships related to T2D remission22 and the preoptic area (POA)23. We selected the POA because this is a region with limited characterization of the underlying mechanisms related to diabetes, despite substantial evidence suggesting a role. Warm-responsive cells in the POA can impair glucose tolerance, presumably by suppressing glucose uptake24. Local norepinephrine administration can induce hyperglycemia25. Furthermore, structural and functional changes in the preoptic area have been linked to streptozotocin treatment, which is a model for diabetes26,27. In addition to glycemic control, intermingled POA cells control thermoregulation, feeding, sleep-wake cycles, social behaviors, and activity presumably from separate cells, but all of these functions have been linked to the human disease28,29,30. Moreover, cells in this region receive direct information from hypothalamic centers31,32, endocrine action33, and nutritive signals34 that presumably lead to these cells being particularly responsive to diabetes remission. In support of this notion of a key set of POA cells in diabetes, the POA receives efferent communication from both the feeding center of the brain (i.e. Arcuate Nucleus) and the glucoregulatory center of the brain (i.e. ventromedial hypothalamus). Thus, dysfunctional afferent communication will logically lead to changes in function of cells who received that information (i.e. hypothalamus-> POA connections). The markers and cell types that we identify could be leveraged in the future for novel targets for metabolic control and may even be able to bypass damage and inflammation that has been observed in the mediobasal hypothalamus of human patients, which predict impaired glucose homeostasis35,36. We used an integrative workflow to analyze these data to determine gene regulatory relationships at the single cell level in the POA. These efforts will hopefully identify markers that could be relevant for diabetes treatment.

Methods

Data collection

Counts data for the preoptic hypothalamus region associated with Moffit et al. was acquired from NCBI GSE113576 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE11357623 and denoted Mof< CELLTYPE > for each subset of cells of a given cell type. Single nucleus counts data from the hypothalamus produced by Bentsen et al. was obtained from GSE148946 https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE14894637 and denoted Ben< CELLTYPE > for each subset of cells of a given cell type. The full datasets including all cell types are denoted MofALL and BenALL.

Neuron filtering

Quality control was performed on the original dataset to remove doublets using the package Scrublet and samples with less than 400 or more than 4000 expressed genes were filtered and removed, as well as any cell with more than 5% mitochondrial and/or ribosomal transcripts. Additionally, genes that were detected in less than 5 cells were removed. Analysis was performed in RStudio (version 4.3.1) using Seurat (version 5.0.2). The counts for both datasets were converted to a Seurat object and preprocessed using the standard Seurat workflow38. Cell type annotations for BenALL were taken from the available metadata, with neurons comprising a total of 11,986 cells. Neuron subtypes were assigned by assessing the expression levels of the excitatory neuron (EXC) marker gene Slc17a6 and the inhibitory neuron (INH) marker genes Gad1, Gad2, and Calcrl23 in distinct Seurat clusters. Overall, 10,100 inhibitory neurons and 1,886 excitatory neurons were assigned using these marker genes. Clusters expressing both inhibitory and excitatory markers were classified as hybrids and removed from the primary analyses.

MASTDEG analysis

Seurat facilitated MASTDEG analysis (MAST version 1.28.0) was used to identify differentially expressed genes (DEGs) in the BenEXC and BenINH data sets according to the standard MASTDEG workflow39. Specifically, we identified DEGs that were associated with FGF1 induced diabetes remission (day 5 vs. day 1) and the same temporal changes in the Vehicle (Veh) group (day 5 vs. day 1). The DEGs that were identified in the day 5 vs. day 1 comparison in the FGF1 treated samples but were insignificant in the same comparison of Veh treated samples were then retained for downstream analysis. These DEG sets were described as BenEXC DEGs and BenINH DEGs respectively (i.e., significant in FGF1 but not in Veh between timepoints). MAST was used due to its specialization in scRNA-seq and was the best fit for our dataset due to the large variance in cell type39. p-values were corrected using the stringent Bonferroni correction with an adjusted p-value cutoff of 0.01 and a log 2 fold change (log2FC) cutoff of 0.32.

hdWGCNA analysis

Gene co-expression modules were identified in excitatory and inhibitory neurons were identified using hdWGCNA14. In summary, consensus network analysis was performed and then the module eigengenes were calculated for each sample.

RNA scope analyses

All experiment protocols were approved by the Committee on the Use and Care of Animals at the Indiana University School of Medicine (Approval number 22072). All methods were carried out in accordance with IACUC and IBC guidelines and regulations. All methods are reported in accordance with ARRIVE guidelines. All mice were provided with standard chow diet and water ad libitum and kept in a temperature-controlled (23 °C) room on a 12-hour light dark cycle. Male C57 wild type mice were acquired ~ 10 weeks of age from Jackson Laboratories (Bar Harbor, ME). Mice were group-housed upon arrival and throughout the study. After mice were a minimum of 10 weeks of age, mice were sacrificed and brains were harvested via transcardial perfusion under deep anesthesia with isoflurane with phosphate buffered saline (PBS) followed by 10% neutral buffered formalin (NBF) (Fisher brand, catalog #245685). Brains were removed and placed into 10% NBF overnight, followed by 30% sucrose for at least 36 h. Brains were cut into 14 mm sections on a freezing microtome in eight series where 6 series were in PBS for mounting and 2 series were stored in antifreeze solution (25% ethylene glycol, 25% glycerol). A predetermined range (~ Bregma + 0.5 through − 1) of 9 individual brain sections were immediately mounted while remaining sections outside of the range were stored in antifreeze solution. Slides were then stored in − 20 °C freezer for a minimum of 2 h and then transferred to − 80 °C freezer for storage up to 3 months.

Slides were removed from − 80 °C freezer and immediately submerged in PBS for 5 min and baked in the HybEZ II Hybridization System (ACD, catalog #321711) at 60 °C. After baking slides were immersed in prechilled NBF for 15 min. Then a series of EtOH washes is conducted using 50%, 70% and 100% EtOH before a RNAscope Hydrogen Peroxide (ACD, catalog # 322381) incubation. Slides were then washed with distilled water and set in boiling 1X target retrieval solution diluted from RNAscope 10X Target Retrieval (ACD, catalog #322000) for 5 min. During this time HybEZ Humidifying Paper (ACD, catalog #310015) was dampened with distilled water and placed in the HybEZ Humidity Control Tray (ACD, catalog #310012). The Hybridization System was then set to 40 °C for all following incubations unless otherwise stated. Slides were again washed with distilled water followed by 100% EtOH before being set to air dry for approximately 10 min. An ImmEdge Hydrophobic Barrier Pen (Vector Laboratory, catalog #H-4000) was then used to draw a thick border around the tissue sections. After the barrier was set RNAProtease III (ACD, catalog #322381) solution was dropped directly on the slide and incubated. After incubation the slides are rinsed in distilled water and the target probes are hybridized. Probes are mixed at 1 volume C2 probe, 1 volume C3 probe, and 50 volumes C1 probe. Channels are separated in a way that allows up to 3 genes of interest to be target at once with impressive specificity. The probes demonstrated in this paper are Mm-Car10-C2 (ACD catalog #492371-C2), Mm-Dgkg-C1 (ACD catalog #535381-C1), Mm-Slit2-C2 (ACD catalog #449691-C2), and Mm-Rorb-C3 (ACD catalog #444271-C3). Probes with the same channel designation, i.e. Mm-Car10-C2 and Mm-Slit2-C2, cannot be ran simultaneously. Of our four probes, Mm-Slit2-C2 and Mm-Rorb-C3 were hybridized concurrently, while Mm-Car10-C2 and Mm-Dgkg-C1 were hybridized with probes not relevant to the scope of this experiment. The probe mix is incubated on the slides for 2 h. After incubation the slides are washed in 1X wash buffer; diluted from RNAscope 50X Wash Buffer (ACD, catalog #310091), before a series of 3 amplification steps incubated and separated by washes in 1X wash buffer. RNAscope Multiplex Fl v2 AMPs 1–3 as well as all HRP developers, blocker, and DAPI are found in the RNAscope Multiplex Fluorescent Detection Reagents v2 Kit (ACD, Catalog #323110). Following amplification each channel signal is developed individually beginning with incubation of its designated HRP solution, washed with 1X wash buffer, incubation of the assigned TSA Vivid fluorophore, and finished with another wash and incubation of the HRP blocker solution to prevent continued channel binding. The fluorophores used in the experiment were TSA Vivid Fluorophore 570 (ACD, Catalog #323272), and TSA Vivid Fluorophore 650 (ACD, Catalog #323273). This process is repeated for each desired channel up to 3 channels. Once signal development is completed for all channels, DAPI is dropped on each slide, incubated for 30 s at room temperature, then immediately removed and replaced with mounting medium and coverslip. Slides are stored in the dark at 2–8 °C and within two weeks.

Images from 14 mm sections were collected in a 4-image grid pattern with slight overlap on each side to capture most of the section including the BNST and POA at Bregma sites 0.26, 0.02, − 0.34, and − 0.70 using an Echo Fluorescent microscope (San Diego, CA) with a 4X objective in the CY5 and TRITC channels. The target regions were identified by overlapping the corresponding bregma level from an anatomical reference using the Mouse Brain in Stereotaxic Coordinates Fourth Edition (Paxinos and Franklin) on microscope images. Images were collected with the same exposure setting on the microscope. Each image of the 4-image grid pattern was opened in Affinity Photo 2 and stitched together using the Panorama function creating one larger image of the section. This stitching process was repeated for each bregma site per each targeted gene. All large images were then opened in ImageJ software to normalize brightness, contrast, and color throughout each image and each targeted gene.

Functional enrichment

Functional enrichment was performed using the clusterProfiler (version 4.10.1) package40. Results were annotated with the AnnotationDBi (version 1.64.1)41 org.Mm.eg.db (version 3.18.0) package and visualized as bar plots.

Dataset integration

The datasets were integrated using Seurat (version 5.0.2) dataset integration features with a CCA reduction and comparison heatmaps were generated using the heatmap3 (version 1.9.9)42 package in R. Other integration methods through Seurat were tested, including RPCA and Harmony, but the CCA reduction was determined to be most effective through a comparison of average LISI scores calculated for cell type cohesion and dataset mixing.

Validation methods

We validated genes that were identified in the DGE and GCN analysis using RNA scope. We picked genes based both on their ranking in the DGE and GCN results as well as the availability of viable reagents for those genes, i.e., effective probes for RNA scope analysis. Based on these results, we evaluated the localization of these genes to specific regions in the preoptic area.

Results

FGF1 induced diabetic remission alters metabolic genes

In this study of scRNA-seq from mouse brain tissues (Fig. 1), we performed independent analyses on two datasets (i.e., Bentsen and Moffit) including DEG and GCN analyses to better understand diabetes remission in the preoptic area. Then to assess the robustness of our results across brain regions we integrated the two distinct datasets (i.e., Bentsen and Moffit). Of the tested integration methods, CCA was the best performing integration technique that had the highest LISI for dataset and lowest LISI for cell type (Fig. S1, Table S1) denoting well integrated data with well stratified cell types.

Fig. 1
Fig. 1
Full size image

Study design and workflow for bioinformatics analysis detailing data accession, tools used for analysis, and outcome of the study design.

The scRNA-seq counts from BenALL were segregated by provided cell type designations. UMAP clustering using the Seurat package was performed (Fig. 2A). The cell type specific marker genes identified by HypoMap were examined in the clustered data and were found to be consistent with expected expression levels in the merged datasets for the cell types with the highest representation in the dataset (Fig. S2)43. The gene expression for the excitatory marker gene Slc17a6 was used to determine the BenEXC, i.e., excitatory neurons from the Bentsen dataset (Fig. 2B)23. The same process was performed using the inhibitory marker genes Calcrl, Gad1, and Gad2 and the clusters containing BenINH, i.e., inhibitory neurons from the Bentsen dataset (Fig. 2C–E). Neurons comprising a total of 11,986 cells with 10,100 BenINH and 1,886 BenEXC. These groups were separated and MASTDEG analysis was performed between remission (day 5) and diabetic (day 1) mice (i.e., FC=remission/diabetic) treated with FGF1. Then the same analysis was performed for Veh with the final tables only including genes that were significant in the FGF1 analysis but not the Veh analysis, i.e., BenEXC (Table 1) and BenINH (Table 2).

Fig. 2
Fig. 2
Full size image

Identification of neuronal subtypes and analysis of DEGs. Counts for each cell type by dataset are reflected in Table S4, counts for each cell type by treatment for BenALL are reflected in Table S5. (A) Seurat UMAP clustering of all neuron cell types identified in GSE148946. (B) Violin plot generation identifying neuron clusters that highly express excitatory marker GAD2. (C–E) Violin plot generation identifying neuron clusters highly expressing inhibitory markers (Calcrl, Gad1, Gad2) as identified by Moffit et al. (F, G) Volcano plots generated from MASTDEG analysis of excitatory and inhibitory identified neuron clusters (adjusted p-value < 0.05, log2FC < −0.58 and log2FC > 0.58) with highly significant positive and negative regulated genes labeled. (H, I) Functional enrichment results for negative regulated BenEXC and BenINH DEGs (higher expression in T2D).

Table 1 Top 10 most upregulated & downregulated genes based on log2FC from BenEXC between day 5 and day 1 in the FGF1 treated mice that were not significant in the Veh treated mice. Note that P denotes p-value, log2FC denotes the average log2FC for the FC=remission(day 5)/diabetic(day 1), Prop1 denotes the proportion of cells expressing the gene in the remission group (day 5), Prop2 denotes the proportion of cells expressing the gene in the diabetic group (day 1), and Padj denotes the bonferroni adjusted p-value.
Full size table
Table 2 Top 10 most upregulated & downregulated genes based on log2FC from BenINH between day 5 and day 1 in the FGF1 treated mice that were not significant in the Veh treated mice. Note that P denotes p-value, log2FC denotes the average log2FC for the FC=remission(day 5)/diabetic(day 1), Prop1 denotes the proportion of cells expressing the gene in the remission group (day 5), Prop2 denotes the proportion of cells expressing the gene in the diabetic group (day 1), and Padj denotes the bonferroni adjusted p-value.
Full size table

BenEXC cells revealed a general downregulation of metabolic genes in the remission state compared to the diabetic state (Fig. 2F), including the metabolic regulator genes Junb, Pfn1, Avp, Cck, and Cox5a (Table 1, File S1), that may indicate metabolic gene regulation by diabetes remission. We also evaluated the higher expression genes (i.e., > 25% cells with detectable expression in either diabetes (day 1) or remission (day 5) groups) and still identified Cox5a in the top DEGs (Tables S2 & S3). Additionally in the BenINH cells, several genes involved in hormonal and neuroendocrine signaling (Oxt and Syt2) were downregulated while another gene in this category (Meis2) was upregulated (Table 2, File S2). DEGs identified in BenINH were downregulated and aside from Meis2; most are involved in metabolic regulation and hormonal signaling. DEGs identified in BenINH showed downregulation of inflammation and immune system regulation associated genes (Cdkn1a, Socs3) that were not seen in BenEXC cells (Fig. 2F–G; Table 2, File S1–S2). In both BenEXC and BenINH DEGs we observed that Timp1, Sprr1a, Junb, and Gm1673 had decreased expression (Tables 1 and 2, File S1–S2).

Gene ontology functional enrichment of the downregulated DEGs from BenEXC cells showed downregulation of cytoplasmic translation, oxidative phosphorylation, cellular respiration, and various other translational and energy producing functions (Fig. 2H). Whereas BenINH cells showed downregulation of many functions involved in energy production including aerobic electron transport chain, oxidative phosphorylation, mitochondrial electron transport cytochrome c to oxygen, and mitochondrial ATP synthesis coupled electron transport (Fig. 2I).

FGF1 induced diabetes remission decreases expression of energy production associated genes

GCN analysis via hdWGCNA grouped genes from BenEXC and BenINH into 10 and 3 distinct modules, respectively (Fig. 3A–B), with each eigenvalue matching the general directionality of the genes within the corresponding GCN module (Fig. 3C–D). Of note, the BenEXC Turquoise module was significantly downregulated in the FGF1 treated samples at day 5 (Fig. 3C; Table 3) and the BenINH Turquoise module was also (Fig. 3D; Table 4). Functional enrichment of the genes classifying this module indicates significant expression of genes regulating aerobic respiration, oxidative phosphorylation, and cellular respiration (Fig. 3E). Cross-referencing genes identified in the BenEXC Turquoise with MASTDEG results revealed 507 down regulated genes (Table 5). Functional enrichment of these genes revealed excitatory processes that may influence energy expenditure such as regulation of synapse organization and regulation of glutamatergic synaptic transmission (Fig. 3F).

Fig. 3
Fig. 3
Full size image

The hdWGCNA analysis reveals unique GCN modules for BenEXC and BenINH cell types. (A, B) PCC matrices of the BenEXC derived GCN modules in the BenEXC cells (A) and the BenINH derived GCN modules in the BenINH cells (B). (C, D) GCN module eigengene values for BenEXC and BenINH cells compared between remission (day 5) and diabetes (day1). Each bar direction is matched with the directionality of expression for identified DEGs. (E) Functional enrichment results for BenEXC Turquoise module. (F) Functional enrichment results for genes identified in both BenEXC Turquoise module and MASTDEG analysis.

Table 3 Eigengene significance of each BenEXC module between FGF1 treatment and vehicle at day 1 and day 5. Note that the log2FC column denotes the log2FC between the eigengene values in FGF1 and Veh, i.e., log2FC(FGF1/Veh).
Full size table
Table 4 Eigengene significance of each BenINH module between FGF1 treatment and vehicle at day 1 and day 5. Note that the log2FC column denotes the log2FC between the eigengene values in FGF1 and Veh, i.e., log2FC(FGF1/Veh).
Full size table
Table 5 Number and directionality of mastdegs identified in each HdWGCNA module for the Bentsen dataset.
Full size table

Distinct cell types in the POA correspond to diabetes remission gene modules

After identifying novel DEGs and gene modules associated with diabetes remission, we next wanted to see if they corresponded with specific sets of POA cell types from Moffit et al.23 Based on the integrated analysis, we were successfully able to merge these two datasets (Fig. 4A–B), had representation in all of the major cell types from each of the datasets (Table S4), and representation from both diabetes/remission groups from the Bentsen datasets (Table S5). Genes identified in the GCN modules in BenEXC and BenINH were also cross-referenced for their expression levels in MofEXC and MofINH, respectively, to evaluate the strength of their relationship (Fig. 4C–D), which showed some conserved relationship especially in the turquoise modules. We additionally compared the DEGs identified in BenEXC and BenINH in the FGF1-treated group with their expression in MofEXC or MofINH, respectively. This showed subsets of cells in MofEXC and MofINH that expressed remission-associated DEGs (Fig. 4E–F).

Fig. 4
Fig. 4
Full size image

Comparative results for MofEXC and MofINH in relation to BenINH and BenEXC derived GCN modules and remission DEGs. (A–B) UMAP plots demonstrating the integration between the Moffitt and Bentsen datasets. (C–D) PCC matrices of the BenEXC derived GCN modules in the MofEXC cells (C) and the BenINH derived GCN modules in the MofINH cells (D). The column colors designate different cell types in the Moffit cells. (E–F) Heatmap of the expression of MofEXC subtype markers in BenEXC cells (E) and MofINH subtype markers in BenINH cells (F). Genes (i.e., rows) annotated with the blue bar were downregulated in the Bentsen remission cells, whereas genes designated with a red bar were upregulated in the Bentsen remission cells. Subtype marker expression across all cell types can be found in Fig. S3C–D.

Validation of uncovered markers for cell types in the mouse preoptic area

A total of 155 genes were identified by MERFISH in the pre-optic area by Moffitt et al. and characterized as either preselected genes as known markers for major cell classes or hypothalamus-specific neuronal functions or neuronal cluster markers. This list included the genes Trpc4, Dgkg, and Ryr3, which were specifically chosen due to being identified as co-expressed in both the Moffitt and Bentsen datasets through weighted gene co-expression network analysis or MASTDEG analysis (Tables 6 and 7, see also Fig. S3A-B). To validate the location of these targets for populations in the preoptic area, we turned to RNA scope (Fig. 5). Using this procedure, we were able to locate the expression level of these genes within the mouse preoptic area. The distribution of Trpc4, Dgkg, Ryr3 and all similarly were expressed throughout the mouse preoptic area, but the distribution varied, particularly in the dorsal-ventral and lateral-medial axes. All three were expressed primarily in the dorsal regions. Ryr3 was typically medial and rostral, Dgkg was also more rostral but both medial and lateral, and Trpc4 was more laterally expressed. These genes along with other markers of interest were expressed as either gene or protein in neurons from human brains from the Human Protein Atlas (Fig. S4). These findings support the expression of these markers in the mouse preoptic region, but do not validate their function in diabetes-related mechanisms.

Table 6 Overlap of marker genes identified by Moffitt et al. with Bentsen gene co-expression modules.
Full size table
Table 7 Selected RNA scope genes along with their inclusion in GCN modules and DEGs from MAST. Note that the first number in the DEG columns is the log2FC and the second number is the bonferroni adjusted p-value.
Full size table
Fig. 5
Fig. 5
Full size image

Validation of targets for subpopulations in the mouse preoptic area. To identify markers of subpopulations in the POA, we turned to RNA scope. We observed expression of ryr3, Dgkg, and trpc4 throughout the preoptic region from Bregma 0.26 to − 0.7. Included are representative images from at least 3 male and 3 female mice. Anatomical markers for each Bregma level are noted on the top panels. Scale bar = 180 μm. BNST, Bed Nucleus of the Stria Terminalis, LS, Lateral Septum, AC, Anterior Commissure, 3 V, Third Ventricle, POA, Preoptic area, AHA, Anterior Hypothalamic Area, SCN, Suprachiasmatic Nucleus.

Discussion

We sought to identify new markers and cell types in the mouse preoptic area responsive to a correction in metabolic function. To do this, we first identified gene co-expression modules that are corrected in the brain with FGF1 treatment, which is a treatment that can powerfully induce diabetes remission from mouse models of diabetes (ob/ob mice in this instance) with a single injection. We integrated datasets with multiple cell types from different regions of the brain, though integration methods showed that expression levels across both datasets for each cell type were consistent enough that they clustered together. We tested several other methods of integration through Seurat’s functions including RPCA and Harmony (Fig. S1) before ultimately settling on CCA integration due to the more effective dataset integration and cell type clustering (Table S1). In Bentsen et al., the authors identified a total of 15 neuron subtypes that were used throughout their analysis. It is important to note that these datasets are from different brain regions, and the integration would have ideally been stronger. However, CAA integration performed the best of the ones tested despite the suboptimal integration. This limitation should be noted when interpreting the DEG and GCN results. However, due to the differences in the neuron subtypes identified in Moffitt et al., we reclassified these neuron clusters as either inhibitory, excitatory, or hybrid depending on cell-type specific gene markers. We acknowledge that this removes some specificity in neuron subtypes and their connection to FGF1-induced T2D remission and future direction for this study would include expanding on neuron subtypes from samples derived from MofALL. Additionally, we acknowledge that the presence of ambient RNA contamination may have led to some anomalies with marker gene expression outside of their specific cell type clusters. While the results we identified were consistent with expected gene expression for this region of the brain (Fig. S4), we plan to introduce methods of ambient RNA removal for future single-cell pipelines, though it is currently outside of the scope of this paper. With this in mind, our findings show a number of differentially expressed genes and related pathways in remission cells as opposed to cells derived from T2D specimens that can be compared back to non-T2D-associated cells from MofALL. We emphasized in our analysis genes that were differentially expressed in excitatory and inhibitory neurons, which were identified in both datasets.

We think that remission modules can be an important tool for investigators focused on neural mechanisms in diabetes and obesity. Metabolically relevant functions are controlled by neural networks that contain cells throughout the brain. As such, there are critical nodes that haven’t been well-established for this role, including the POA. This goal is difficult to reach because many of these structures, including the POA, contain cell types that perform unrelated functions. As an example, the POA contains cells that are involved in ocular, reproductive, temperature, and behavioral functioning. An important goal of our lab is to identify the hallmarks for metabolically relevant cells in the POA. But one weakness of this approach is that we assume these effects will be translatable from efforts with FGF1 that have focused on the Median Eminence and Arcuate nucleus to the preoptic area. However, this has been used successfully for translating effects between different types of cells in the brain in the development of neurological disorders21. In further support of this notion, sustained inhibition of NPY/AGRP in the hypothalamic Arcuate nucleus is a key site for FGF1 action, but these signals act through the POA to initiate their effects; these studies can help understand efferent communication in both FGF1 pharmacology and neural mechanisms involved in diabetes remission44,45. Importantly, we are not claiming that this strategy will alone answer these questions and instead think that this is one tool among many that will identify new mechanisms in the brain for metabolic control. This approach will not replace single cell sequencing alone but will be a complement to single cell sequencing and could identify mechanisms for a project without the resources to complete a full sing cell sequencing studies across both sexes and with multiple treatments. The realistic goal is not to identify mechanisms alone but will help reduce false positives that are inherent problems with any omics type approach - particularly in a tissue as heterogeneous as the brain.

Besides DEGs, gene regulatory networks (GRNs) may also play an important role in determining disease mechanisms. For this reason, the commonly used approach of GRN analysis is used to identify sets of genes, i.e., gene modules, that are correlated with one another and may represent disease-associated GRNs. These modules can be jointly measured as an individual unit, i.e., an eigengene, and thus can be measured between disease states46. Gene coexpression module analysis has already yielded insights in numerous neurological47 and metabolic diseases48. Here we apply it to study the effects of T2D remission in the preoptic area of the brain.

Our data shines new light on the functioning of the POA in metabolic disease. Indeed, the POA controls both thermoregulation and energy expenditure, with distinct neuronal populations that each play specialized roles. Some markers have been established, such as leptin receptor for energy expenditure, BRS3 for thermoregulation, and BDNF/PACAP for torpor49,50,51. In addition to these functions that play an important role in overall metabolic tone, POA cells have been connected with inducing changes in glucose homeostasis. Taken together, it would be surprising if this wasn’t an area that would be responsive to diabetes and obesity. More work is needed to tie the regulation of the region to cases of diabetes and obesity to tie our data to remission from diabetes and obesity, in a manner that would be consistent with what FGF1 does.

While there are likely similarities in how neurons respond in remission from disease, there is a great deal of heterogeneity in the brain. Brain regions don’t respond to the same kinds of signals and don’t respond in the same way. In particular, this is the case with high fat diet exposure. Microglia and Astroglia recruitment only occur in the Arcuate nucleus but no other hypothalamic regions. Therefore, cells within this particular hypothalamic nucleus are sensitive to neurodegeneration. That being said, the Arcuate nucleus sends projections to the POA32,52, which would in turn lead to hypoactivation of a subset of cells in the POA. Thus, there could be secondary neurodegeneration that would impact the POA and still reveal similar normalization in the remission state. This could help explain the surprising lack of a signal within traditional glucose sensing mechanisms in our data. There is no evidence currently for direct glucose sensing in the POA, but there is evidence or POA neurons regulating glucose homeostasis24,53. This could be explained by the fact that glucose information is sent to the POA by efferent projections coming from places like the ventromedial hypothalamus and arcuate nucleus. If this is the case, it would not be a surprise that we observed mechanisms unique to how the POA processes information related to diabetes.

Since neurons can vary greatly in their gene expression and function depending on whether they contain either excitatory or inhibitory transmitters, we separated out our analyses by the expression level of slc17a6 (i.e. Vglut2) or slc32a1 (i.e. Vgat). Presumably, we will have less variability, particularly cell functions and gene expression, by separating our analyses. Cell clusters that did not have high confidence in their classification as excitatory or inhibitory neurons in BenALL were classified as hybrids and excluded (n = 2163) to reduce variability in downstream analysis. By doing this, our data points to a key role of excitatory cells represented by the turquoise module in diabetes remission. We found that this set of cells had a pronounced decrease in the average change in eigenvalue from control, dramatically more than any other group in the analysis. In addition, we performed functional enrichment and found an overall downregulation in genes associated with cellular processes, such as aerobic respiration, oxidative phosphorylation, cellular respiration, and mitochondrial gene expression. These findings are consistent with an overall reduction in metabolic activity of these cells which could be due to less activation and/or neurodegeneration. Future studies will examine the function of this population and if we can reverse these changes for therapeutic benefits in diabetes.

For genes downregulated, we observed that both excitatory and inhibitory datasets had a decrease in Timp1, Sprr1a, Junb, and Gm1673. As with the downregulated genes, none of these genes have been associated with functioning in the POA. However, Timp1, Sprr1a, and Junb have been associated with diabetes. Timp1 is required for high fat diet-induced glucose intolerance and hepatic steatosis54, and knockouts also induce anxiety like behavior55. This is consistent with a function in the POA, in which a downregulation may be protective for diabetes in remission. However, future studies will need to test that hypothesis. Junb is a transcription factor that is part of the Ap-1 subunit connected with ER stress and apoptosis in pancreatic beta cells56. Since beta cells and neurons share genetic similarities, downregulation of Junb could be part of the protection in remission from diabetes in the POA, but that would also have to be tested. Sprr1a has been associated with autophagy and the progression of beta cell failure in diabetes57, but has also been connected with heart failure and colon cancer58,59,60. Nothing is known about Gm1673. Thus, it is not at all clear what this observation could mean to diabetes remission.

There are potential limitations to the interpretations of this data. For one, the remission modules were generated from mice on an ob/ob background, whereas preoptic area analyses were performed in wild type C57 mice fed a high fat diet to induce diabetes and obesity. It is possible that the different backgrounds could negatively impact our results. However, we assume that the most meaningful findings will not be dependent of the diabetes model chosen for these analyses. It is also possible that these strategies used in other diseases and with different tissues may be less generalizable within the brain. Particularly, these studies were originally performed with conclusions in oligodendrocytes and we are applying them to neurons in a different brain region. However, we do think that adaptations and responses are shared along a circuit, as these dysfunctional signal operate from the brain to the periphery, and shared upregulation or downregulation throughout the brain as it adapts to disease. That being said, we do not assume that remission modules alone will be able to find new mechanisms in the brain in metabolic control, but they can be one tool used towards that end goal.

Conclusion

In conclusion, our studies use data from diabetes remission to form remission modules that helped us define new regulated functions in the murine POA. By defining these subsets, we identified new genes connected with diabetes remission, many of which already have connections in the literature with diabetes. Future studies will tease apart the functions and how they could be targeted for future treatments.