Human DLPFC Visium Spatial Transcriptomics Analysis

Author

Dayou Zou

Published

May 26, 2026

1 Introduction

1.1 Data Background

The human dorsolateral prefrontal cortex (DLPFC) exhibits well-defined cortical laminar architecture. 10x Visium spatial transcriptomics measures spot-level RNA expression while preserving tissue spatial coordinates, making it suitable for examining spatial expression patterns, cortical layer-associated markers, and the ability of different spatial clustering methods to identify laminar structure.

This project uses the human DLPFC Visium dataset published by Maynard et al., with sample 151673 as the analysis target. This sample includes manual annotations for Layer1–Layer6 and white matter (WM), and thus serves as a small-scale benchmark for comparing unsupervised clustering results against expert layer annotations.

1.2 Project Objectives

This project focuses on analytical workflow and method comparison, covering five parts:

  • QC, normalization, HVG selection, PCA, and expression graph construction on the spot-level expression matrix;
  • Identification of spatially variable genes using Moran’s I, and examination of known cortical/WM marker spatial distributions;
  • Comparison of Leiden, SpaGCN, and BayesSpace clustering results against manual layer annotations;
  • Estimation of major cell type spatial trends via Tangram deconvolution using human MTG snRNA-seq as reference;
  • Examination of spatial neighborhood relationships using neighborhood enrichment, co-occurrence, and Ripley’s L.

Manual layer annotations are not used in unsupervised clustering model fitting and serve primarily for downstream evaluation of method consistency with cortical laminar structure. The settings for Leiden resolution, SpaGCN n_clusters, and BayesSpace q reference the 7 expert-annotated regions; therefore, the clustering comparison in this report should be understood as a benchmark near a fixed target cluster count, rather than fully unconstrained model selection.

1.3 Data Overview

This analysis uses the publicly available Space Ranger output matrix and corresponding manual layer annotations for sample 151673. After initial loading, the data contains 3,639 spots and 33,538 genes; 3,611 spots carry manual layer annotations. After removing spots without layer annotations, Layer1–Layer6 and WM are all retained in the analysis.

Table 1: Spot composition by manual layer annotation
Region Spot count
Layer1 273
Layer2 253
Layer3 989
Layer4 218
Layer5 673
Layer6 692
WM 513
Total 3,611

2 Quality Control

Quality control is performed at the spot level, examining total_counts, n_genes_by_counts, and pct_counts_mt. Before filtering, the number of spots with manual layer annotations is 3,611. Filtering criteria are min_counts >= 500, min_genes >= 200, and pct_counts_mt < 30; genes detected in fewer than 10 spots are also removed. After filtering, 3,601 spots and 16,567 genes are retained.

Figure 1: Spatial distribution of main QC metrics

Figure 1 shows the spatial distribution of three main QC metrics. Gray matter regions generally display higher UMI counts and detected gene numbers; pct_counts_mt does not show obvious localized elevation. This project does not use multi-sample or external QC controls, so the QC results here serve only as basic filtering criteria for downstream analysis.

Table 2: Data dimensions before and after QC
Stage Spots Genes
Initial loading 3,639 33,538
With manual annotation 3,611 33,538
After QC 3,601 16,567
Figure 2: Manual layer annotation

Figure 2 shows the manual layer annotation for sample 151673. This annotation is used for downstream clustering evaluation and does not serve as training labels for Leiden, SpaGCN, or BayesSpace.

3 Spatial Expression Structure

3.1 Preprocessing and Expression Clustering Baseline

The post-QC matrix is library size normalized with target_sum=10,000 and log1p transformed. HVG selection uses the seurat_v3 method, retaining 3,000 highly variable genes. PCA is computed on HVGs with 50 components; the downstream kNN graph uses the first 17 PCs.

Leiden clustering serves as an expression clustering baseline that does not explicitly use spatial coordinates. This baseline is used for comparison with spatial methods such as SpaGCN and BayesSpace, rather than as a final layer annotation result.

Table 3: Preprocessing and expression clustering parameters
Step Setting
Normalization target_sum=10,000
Transformation log1p
HVG seurat_v3, n_top_genes=3,000
PCA 50 PCs
kNN graph n_neighbors=15, n_pcs=17
Leiden baseline resolution=0.62

3.2 Spatially Variable Genes

Moran’s I is computed on 3,000 HVGs and ranked in descending order. Moran’s I measures the spatial autocorrelation of expression values; higher values indicate that the gene tends to form local spatial clusters. Statistical significance is estimated via 100 permutations.

Figure 3: Top 10 spatially variable genes

Figure 3 shows the top 10 spatially variable genes ranked by Moran’s I. MBP, PLP1, MOBP, MAG, and other myelin/oligodendrocyte/WM-related genes rank high, suggesting that WM-related expression signal is one of the stronger spatial patterns in this sample. These results indicate spatial clustering trends of RNA expression only, and cannot be directly interpreted as functional enhancement, loss of function, or changes in regulatory mechanisms.

Table 4: Top 10 genes ranked by Moran’s I
Gene Moran’s I FDR
MBP 0.8075 <0.0001
PLP1 0.7019 <0.0001
SCGB2A2 0.7017 <0.0001
MOBP 0.6461 <0.0001
GFAP 0.6246 <0.0001
TF 0.5927 <0.0001
CNP 0.5717 <0.0001
CRYAB 0.5613 <0.0001
MAG 0.5459 <0.0001
PPP1R14A 0.5334 <0.0001

3.3 Known Marker Examination

To check whether spatial expression results are consistent with known anatomical directions, this project examines several commonly used markers. This examination serves as a sanity check and does not constitute independent validation.

Figure 4: Known marker expression

Figure 4 shows the spatial expression of 5 reference markers. MBP and MOBP high-expression regions are mainly located in WM; SNAP25 is primarily distributed in cortical gray matter. PCP4 shows some expression near Layer5, but its spatial autocorrelation strength is lower than the top WM markers. RELN ranks low in Moran’s I in this analysis.

Table 5: Known marker spatial expression examination
Marker Reference direction Spatial expression observation Moran’s I rank
MBP WM / myelin / oligodendrocyte-related Stronger signal in WM region 1
MOBP WM / myelin / oligodendrocyte-related Stronger signal in WM region 4
SNAP25 Neuron / gray matter-related Stronger signal in gray matter 20
PCP4 Layer5-related Some Layer5-associated trend 67
RELN Layer1-related Weak spatial autocorrelation 399

4 Spatial Clustering Method Comparison

4.1 Comparison Rationale

This project compares three unsupervised clustering approaches: Leiden uses only expression PCA/kNN graph; SpaGCN incorporates spatial coordinates and H&E image in addition to expression; BayesSpace uses expression PCA and Visium spot adjacency structure. Clustering results do not use manual layer annotations for training or fitting. Manual layer annotations serve only as reference labels for computing ARI and NMI.

Table 6: Clustering method settings
Method Input information Parameter settings Role in this analysis
Leiden Expression PCA/kNN graph resolution=0.62 Baseline without spatial information
SpaGCN Expression, spatial coordinates, H&E image beta=49, n_clusters=7 Spatial graph model
BayesSpace Expression PCA and Visium adjacency structure q=7, gamma=2, nrep=50,000 Spatial prior model

4.2 Spatial Distribution Comparison

Figure 5: Spatial clustering result comparison

Figure 5 compares the spatial distributions of manual layer annotations, Leiden, SpaGCN, and BayesSpace. BayesSpace preserves relatively clear locally continuous structure in this sample; Leiden primarily reflects expression similarity with weaker local spatial continuity; SpaGCN under the current parameters also does not surpass Leiden.

4.3 Quantitative Evaluation

Table 7: Clustering consistency with manual layer annotations
Method ARI NMI
Leiden 0.3882 0.5009
SpaGCN 0.3163 0.4593
BayesSpace 0.5355 0.6729

For sample 151673 under the current parameter settings, BayesSpace achieves the highest ARI/NMI. This result supports the utility of spatial priors in laminar structure identification for this sample. SpaGCN falls below Leiden under the current settings, possibly due to parameter selection, H&E information utilization, or smoothing effects; this report does not perform further validation and therefore does not generalize this result to a conclusion about method superiority.

6 Spatial Neighborhood Analysis

6.1 Neighborhood Enrichment

Neighborhood enrichment examines whether different regions or clusters tend to be spatially connected to themselves or adjacent regions. Here it is computed separately for manual layer annotations and BayesSpace clusters.

Figure 8: Ground truth neighborhood enrichment
Figure 9: BayesSpace neighborhood enrichment

Figure 8 and Figure 9 show that both manual layer annotations and BayesSpace clusters exhibit local continuity. This result can serve as a spatial structure check but cannot alone be used to determine whether BayesSpace accurately recovers inter-layer boundaries.

6.2 Co-occurrence and Ripley’s L

Co-occurrence analysis is used to observe the co-occurrence trends of specified layers at different distance scales. This project examines Layer1, Layer3, Layer5, and WM.

Layer1 co-occurrence

Layer1 co-occurrence

Layer3 co-occurrence

Layer3 co-occurrence

Layer5 co-occurrence

Layer5 co-occurrence

WM co-occurrence

WM co-occurrence

These plots serve as supplementary checks of the spatial organization of layer annotations. No additional statistical tests are performed on the co-occurrence curves or Ripley’s L here, so they are used as descriptive results only.

Figure 10: Ripley’s L curves

Figure 10 shows the Ripley’s L curves for different layer annotations at various distance scales. Curve differences are used primarily to observe spatial clustering and range differences across layers, and are not used alone as a criterion for clustering accuracy.

7 Discussion

7.1 Main Observations

This project starts from 3,611 spots with manual layer annotations and retains 3,601 spots and 16,567 genes after QC. Spatially variable gene analysis shows that MBP, PLP1, MOBP, MAG, and other WM/oligodendrocyte-related genes rank high in Moran’s I, suggesting that the expression difference between WM and gray matter is one of the stronger spatial signals in this sample.

7.2 Method-Level Results

Under sample 151673, the current parameters, and ARI/NMI metrics, BayesSpace achieves the highest consistency with manual layer annotations (ARI = 0.5355, NMI = 0.6729). This result supports the utility of spatial priors in laminar structure identification for this sample.

In Tangram reference mapping, the Oligodendrocyte high-score region is mainly located in WM, and the L2/3 IT high-score region is mainly in superficial cortical layers. Since Tangram output consists of relative scores derived from reference mapping, it cannot be interpreted as true cell proportions or functional mechanisms.

7.3 Limitations

This project is a single-sample, public-data, parameter-specific analysis. Clustering method comparison uses manual layer annotations as reference labels and sets the target cluster count near 7 expert-annotated regions; therefore, results should not be generalized to all DLPFC samples or all spatial clustering tasks.

RNA-level expression trends, Tangram mapping, and neighborhood analyses are all computational-level evidence and cannot directly demonstrate functional changes, causal relationships, or regulatory mechanisms. Supporting mechanism-level conclusions requires multi-sample replication, independent dataset validation, or experimental verification.

8 Methods Summary

All results above are based on the same preprocessed object. Manual layer annotations are used only for result evaluation and do not participate in unsupervised clustering or Tangram mapping.

Table 10: Main analysis parameter summary
Analysis step Main settings
Data input Sample 151673; filtered_feature_bc_matrix.h5; manual layer annotations
QC min_counts >= 500; min_genes >= 200; pct_counts_mt < 30; genes detected in ≥10 spots
Preprocessing target_sum=10,000; log1p; seurat_v3 HVG; 3,000 HVGs
PCA/kNN 50 PCs; kNN uses first 17 PCs
SVG Squidpy Moran’s I; 3,000 HVGs; 100 permutations
Leiden resolution=0.62
SpaGCN beta=49; n_clusters=7
BayesSpace q=7; gamma=2; nrep=50,000
Tangram Human MTG snRNA-seq reference; 10,743 nuclei; 24 subclasses; 1,279 marker genes; mode='clusters'; rna_count_based; 500 epochs
Evaluation metrics ARI; NMI; reference labels are manual layer annotations

9 Computational Environment

The table below lists only software directly used or directly affecting main results in this project, excluding dependencies.

Table 11: Software versions directly relevant to this project
Software Version Purpose
Python 3.12.13 Python analysis environment
R 4.5.3 BayesSpace runtime
Scanpy 1.12.1 Data loading, QC, normalization, PCA, Leiden
Squidpy 1.8.1 Spatial neighbor graph, Moran’s I, neighborhood analysis
SpaGCN 1.2.7 Spatial clustering
BayesSpace 1.20.2 Spatial clustering
Tangram 1.0.4 snRNA-seq reference mapping
scikit-learn 1.8.0 ARI, NMI, and other evaluation metrics

10 References

  • Maynard KR, et al. (2021). Transcriptome-scale spatial gene expression in the human dorsolateral prefrontal cortex. Nature Neuroscience, 24, 425–436.
  • Zhao E, et al. (2021). Spatial transcriptomics at subspot resolution with BayesSpace. Nature Biotechnology, 39, 1375–1384.
  • Hu J, et al. (2021). SpaGCN: Integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network. Nature Methods, 18, 1342–1351.
  • Biancalani T, et al. (2021). Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram. Nature Methods, 18, 1352–1362.
  • Palla G, et al. (2022). Squidpy: a scalable framework for spatial omics analysis. Nature Methods, 19, 171–178.