Spatial Transcriptomics Revealed Alzheimer’s Disease Associated Molecular Markers in Parvalbumin Interneurons

Created on October 07, 2024
Updated on May 20, 2026

Acknowledgement

I would like to sincerely thank Dr. Justin Chun, a pioneer in spatial transcriptomics, who generated both the GeoMx DSP and CosMx SMI datasets for his diabetic kidney disease (DKD) project. Working with these datasets gave me the opportunity to develop the Spatial Omics Toolkit (SOTK), although the DKD project itself did not ultimately take advantage of this methodology. Using SOTK, I subsequently analyzed a separate GeoMx DSP dataset together with Xenium In Situ data, which was to become publicly available as the data owner was preparing their manuscript for publication; this is why InSituType (NanoString), originally developed for CosMx data, was used in this study. I thank the Applied Spatial Omics Centre and the contributing authors for their contributions to this paper. I am profoundly grateful to the Snyder Institute, and I gratefully acknowledge its support, which allowed me to focus on the revision process.

Introduction

• Aim of the study
• Utilized GeoMx DSP and Xenium In Situ for spatial transcriptomic analysis of PV+ and PV− neurons in the retrosplenial cortex (RSC) of a mouse model of Alzheimer’s Disease (AD).
• Discovered distinct disease-associated molecular markers specifically in PV+ neuron populations.
• Integrated datasets from GeoMx and Xenium for comprehensive spatial transcriptomic insights.
• Established the methodological coherence between GeoMx DSP and Xenium In Situ platforms, setting a framework for future integrated studies.

Materials

1. 5xFAD transgenic (TG) and wild-type (WT) mice
2. Spatial transcriptomics datasets

[GeoMx] Discovery cohort Sample, N=7: Four TG-female mice Three WT-female mice Platform: Nanostring GeoMx Digital Spatial Profiler (DSP) Panel: Whole Transcriptome Atlas (WTA) Segment, N=4: PV (purple, interneurons) NeuN (green, neurons) Amyloid (red) TN (triple negative) ROI (Region of interest): 14 AOI (Area of illumination): 46 [Xenium] Validation cohort Sample, (consecutive sections): N=6: Three TG-female mice Three WT-female mice Platform: 10X Genomics Xenium In Situ Panel: 247 genes FOV (Field of view): 349 Number of cells: 303,158 cells

Sample labels and corresponding tissues — concatenated GeoMx DSP and Xenium In Situ slides showing the manually labelled mouse half-brain sections used in this report



3. Number of samples:

Platform	TG	WT	Total
GeoMx	4	3	7
Xenium	3	3	6
Total	7	6	13



4. Sample naming convention for GeoMx DSP data
Group|No|Sex - Segment - ROI
1. Group: [TG | WT]
2. No (mouse model ID): as is
3. Sex: [M | F]
4. Segment: [PV | NeuN | Amyloid | TN]
5. ROI: three-digit number
6. Example: TG1M-PV-010

Methods

1. Non-negative Matrix Factorization (NMF) ^{Lee and Seung (1999)}
• Identification of metagenes ^{Brunet et al. (2004)}
2. InSituType ^{Danaher et al. (2022)}
• Cell-typing with the reference datasets (supervised)
3. SOTK (Spatial Omics Toolkit): Select the best rank from NMF based on correlation network in a data-driven way
• sotk2 SOTK deprecated

Results

1. Preprocessing, QC, and XeniumRanger
• Nanostring GeoMx DSP
• 10X Genomics Xenium In Situ
1. TG2F (Region 2)
2. TG3F (Region 4)
3. TG4F (Region 6)
4. WT1F (Region 1)
5. WT2F (Region 3)
6. WT3F (Region 5)


2. [GeoMx] Normalization
• Variance Stabilizing Transformation (VST) was applied using DESeq2 ^{Love et al. (2014)}

RNA-seq tends to follow the negative-binomial distribution, their SD tends to increase with mean A shifting and scaling of the data to stabilize the variance (generalised logarithm) for each cohort/profile In this study, Count matrix was fed in to VST and upper quartile [UQ] normalization was not applied

Effect of variance-stabilizing transformation (VST)



• Segment-level boxplots:

Count	VST



• Individual-level boxplots:

Count	VST



3. [GeoMx] Identify Interneuron- and Neuron-specific genes
• Differential Expression (DE) analysis between PV and NeuN segment in WT
1. Welch's t-test, FDR <0.05 and | Log2 fold-change | > 0.2
• Results:
1. Stats results in XLSX
   
   Differential expression between PV and NeuN segments in wild-type (WT)


4. [GeoMx] Feature selection

Selects a subset of relevant features while keeping the original feature space intact Profiles often too big for downstream analysis Highly variable genes, N=2000: Coefficient of variation < 0.01 Order by standard deviation

Highly variable gene selection



5. [GeoMx] Transcriptome deconvolution using NMF
• NMF parameters
1. Method: brunet
2. Rank: {2, 3, 4, 5, ..., 10}
3. Iteration: 1000
4. Seed: 123456
• Input
1. PV: 2,000 features X 12 samples
2. NeuN: 2,000 features X 14 samples
• Output

Segment	Rank survey	Consensus matrix	Coefficient H matrix (Metagene X Sample)	Basis W matrix (Gene X Metagene)
PV
NeuN



6. [GeoMx] Identify overrepresented metagenes
• Workflow:
  
  Schematic overview of how the Spatial Omics Toolkit (SOTK) works: a stepwise workflow for detecting communities from spatial data using NMF
1. Include a subset of the NMF results: max rank at half size of the population
2. Concatenate basis W matrices column-wide from NMF results
3. Calculate all pairwise coefficient correlations using Spearman
4. Generate a correlation network with positive correlations (Spearman rho >0.5)
5. Identify communities using the fast greedy algorithm
• Overrepresented communities, or biological modules, identified:
1. PV: four communities identified from 20 metagenes
2. NeuN: five communities identified from 27 metagenes


7. [GeoMx] Select an optimal Rank from NMF using SOTK
• Mark and label metagenes at each rank:

Rank = 3

Rank = 2 Rank = 4 Rank = 5 Rank = 6

• Select the best rank: the minimum rank that covers the most communities
1. PV, Rank = 3
2. NeuN, Rank = 4
• Sample-metagene relationships: based on the maximum coefficient values
1. PV, Rank = 3
2. NeuN, Rank = 4
• Metagene-associated genes ^{Tsukahara et al. (2021)}
1. The top 100 genes based on basis values for a given metagene
2. Correlation between the expression level and coefficient values for a given gene

PV	NeuN
Metagene 1 (97) Metagene 2 (94) Metagene 3 (79)	Metagene 1 (83) Metagene 2 (49) Metagene 3 (88) Metagene 4 (72)



8. [GeoMx] Metagene functional annotations
• Hypergeometric distribution - g:Profiler results:
1. PV
2. NeuN
• Top 10 enriched terms:
1. PV, Metagene 1
2. PV, Metagene 2
3. PV, Metagene 3
4. NeuN, Metagene 1
5. NeuN, Metagene 2
6. NeuN, Metagene 3
7. NeuN, Metagene 4




9. [GeoMx] Variability among samples
• Metagene-sample relationships on the correlation network per sample:
1. PV
2. NeuN
• Each sample shows different contribution of the metagenes
  
  Variability in metagene contribution among samples



10. [Xenium] Regions of Interest (ROIs)

The number of cells in each ROI Section WT1F WT2F WT3F TG2F TG3F TG4F Total RSC 3,415 2,335 1,498 2,085 391 187 9,911 SUB 293 142 161 297 N/A N/A 893 VIS 1,483 1,223 637 1,247 N/A 258 4,848 ENT 682 430 738 426 566 264 3,106 CA1 663 807 1,140 907 1,172 753 5,442 Total 6,536 4,937 4,174 4,962 2,129 1,462 24,200 RSC, retrosplenial cortex SUB, subiculum VIS, visual cortex ENT, entorhinal cortex CA1 region of the hippocampus

Xenium regions of interest (ROIs)



11. [Xenium and GeoMx] Correlation between and across consecutive sections in RSC
1. between sections:

   Expression correlation between consecutive sections

   PV WT1F WT2F WT3F TG2F TG3F TG4F

   NeuN WT1F WT2F WT3F TG2F TG3F TG4F

2. across sections:

   Sample correlation across GeoMx and Xenium sections

   GeoMx-PV and Xenium GeoMx-NeuN and Xenium



12. [Xenium] Unsupervised clustering using k-means clustering
• k = 7 (Assumption: at least seven cell types based on the NMF results [three PV and four NeuN metagenes])
• Results - import the following CSV files to the Xenium Explorer:

WT1F	WT2F	WT3F	Legend
TG2F	TG3F	TG4F
			WT1F WT2F WT3F TG2F TG3F TG4F



13. [Xenium] Deconvolution using NMF results
• Cell typing of Xenium data with the GeoMx data (NMF results) as a reference using InSituType ^{Danaher et al. (2022)}
  
  Xenium cell-type deconvolution using the GeoMx NMF reference
• Results - import the following CSV files to the Xenium Explorer:
1. WT1F
2. WT2F
3. WT3F
4. TG2F
5. TG3F
6. TG4F


14. [Xenium] t-distributed Stochastic Neighbor Embedding (tSNE)
• RSC section
• All cells from five sections


15. [Xenium] Differential expression analysis
• Wilcoxon rank sum test results in EXCEL
  
  Differential expression (log2 fold-change) across the five sections

Discussion/Limitations

• Developed a framework to deconvolute spatial transcriptomics data and identified TG- and WT-associated metagenes and features/genes.
• Dner, Gad1, and Pvalb are significantly down-regulated in TG compared to the WT in female mice.
• There is a limited intersection between the mouse whole transcriptomics atlas and the xenium 250 mouse brain panel.
• An arbitrary threshold value, i.e., Spearman rho >0.5, was used in the network construction. However, higher coefficient threshold filters in a limited number of nodes and lower values include doubtful biological modules.
• Xenium internally checks and filters out out-of-focus FOV (field-of-view). However, it is necessary to filter data at the FOV level since each FOV captures different morphology or regions that should be normalized across FOVs.

Data Availability

• GeoMx, raw data: GSE277793 from GEO (Gene Expression Omnibus)
• GeoMx, NMF inputs:
1. GeoMx.PV.RDS
2. GeoMx.NeuN.RDS
• GeoMx, SOTK objects:
1. sotkObj.PV.RDS
2. sotkObj.NeuN.RDS
• Xenium, raw data: GSE277463 from GEO
• Xenium, ROIs: Xenium.ROI.RDS

Code Availability

• [Zenodo] Workflows
• [Shiny app] SOTK demo
• [GitHub] sotk2 SOTK deprecated
• [Docker Hub] sotk2 image, run locally

References

Brunet et al. Metagenes and molecular pattern discovery using matrix factorization. Proc Natl Acad Sci 2004;101(12):4164-9.
Danaher et al. Insitutype: likelihood-based cell typing for single cell spatial transcriptomics. bioRxiv 2022.10.19.512902.
Lee and Seung Learning the parts of objects by non-negative matrix factorization. Nature 1999;401(6755):788-91.
Love et al. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15(12):550.
Tsukahara et al. A transcriptional rheostat couples past activity to future sensory responses. Cell 2021;184(26):6326-6343.e32.

Abbreviations