Abstract
Recent advances in spatial transcriptomics have enabled simultaneous
preservation of high‐throughput gene expression profiles and the
spatial context, enabling high‐resolution exploration of distinct
regional characterization in tissue. To effectively understand the
underlying biological mechanisms within tissue microenvironments, there
is a requisite for methods that can accurately capture external spatial
heterogeneity and interpret internal gene regulation from spatial
transcriptomics data. However, current methods for region
identification often lack the simultaneous characterizing of spatial
structure and gene regulation, thereby limiting the ability of spatial
dissection and gene interpretation. Here, stDCL is developed, a dual
graph contrastive learning method to identify spatial domains and
interpret gene regulation in spatial transcriptomics data. stDCL
adaptively incorporates gene expression data and spatial information
via a graph embedding autoencoder, thereby preserving critical
information within the latent embedding representations. In addition,
dual graph contrastive learning is proposed to train the model,
ensuring that the latent embedding representation closely resembles the
actual spatial distribution and exhibits cluster similarity.
Benchmarking stDCL against other state‐of‐the‐art clustering methods
using complex cortex datasets demonstrates its superior accuracy and
effectiveness in identifying spatial domains. Our analysis of the
imputation matrices generated by stDCL reveals its capability to
reconstruct spatial hierarchical structures and refine differential
expression assessment. Furthermore, it is demonstrated that the
versatility of stDCL in interpretability of gene regulation, spatial
heterogeneity at high resolution, and embryonic developmental patterns.
In addition, it is also showed that stDCL can successfully annotate
disease‐associated astrocyte subtypes in Alzheimer's disease and
unravel multiple relevant pathways and regulatory mechanisms.
Keywords: dual graph contrastive learning, gene regulation, graph
contrastive learning, spatial heterogeneity
__________________________________________________________________
stDCL, a dual graph contrastive learning method, captures spatial
heterogeneity and interprets gene regulation in spatial transcriptomics
data. Integrating spatial and gene expression data through graph
embeddings, stDCL provides robust spatial characterization and accurate
region identification, reconstructs spatial hierarchies, and identifies
disease‐associated cell subtypes, unveiling new insights into tissue
microenvironments and disease mechanisms.
graphic file with name ADVS-12-2410081-g009.jpg
1. Introduction
Spatial transcriptomics involves a series of emerging genomic
technologies that reveal gene expression in tissues with spatial
context, thereby providing a deeper understanding of the structure and
function of tissues at the spatial level.^[ [36]^1 ^] Several spatial
transcriptomics technologies have been developed, including RNA
sequencing‐based technologies such as 10x Visium,^[ [37]^2 ^]
Slide‐seq,^[ [38]^3 ^] and spatial transcriptomics (ST),^[ [39]^4 ^]
which enable genome‐wide analysis of gene expression by thousands
across tissue locations. In addition, single‐molecule fluorescence in
situ hybridization (smFISH)‐based technologies, such as MERFISH,^[
[40]^5 , [41]^6 , [42]^7 ^] seqFISH,^[ [43]^8 , [44]^9 ^] and seqFISH
plus,^[ [45]^10 ^] offer single‐cell resolution, bridging gene
expression and spatial position. These advancements allow studying
spatial gene expression patterns,^[ [46]^11 , [47]^12 ^] exploring the
transcriptomic landscape of tissues, and inferring cell‐to‐cell
communication.^[ [48]^13 ^] In general, spatial transcriptomics has
facilitated new discoveries in many areas of biology.^[ [49]^14 ,
[50]^15 ^]
The spatial heterogeneity of complex tissue architecture is
fundamentally related to the spatial domains of different regions.
These domains, characterized by their distinct gene expression patterns
and histological characteristics, underlie various biological functions
and regulatory mechanisms. They are characterized by distinct cellular
compositions, transcriptomic diversity, and cell‐to‐cell interactions,
all of which are critical to understanding physiology and pathology.^[
[51]^16 , [52]^17 , [53]^18 ^] Therefore, it is crucial to identify and
determine the spatial domains in tissues. Traditional clustering
algorithms, such as K‐means and Louvain, rely primarily on gene
expression data and ignore critical spatial information, resulting in
clustering results that fail to accurately represent the intricacies of
the actual tissue. To mitigate this issue, several methodologies
incorporating spatial information for enhanced identification of
spatial domains have been developed. Among these, stLearn^[ [54]^19 ^]
extracts morphological features from morphological images and spatial
locations using deep learning models and normalizes gene expression
using morphological distances. BayesSpace,^[ [55]^20 ^] a Bayesian
statistical clustering method, imposes a prior for ST data and
encourages spatially adjacent spots to belong to the same cluster.
SpaGCN^[ [56]^21 ^] employs graph convolution networks to integrate
gene expression with weighted graphs constructed from spatial distance
and histology to identify spatial domains. STAGATE^[ [57]^22 ^]
develops an adaptive graph attention autoencoder to learn
low‐dimensional latent embeddings of ST data and identify spatial
domains. CCST^[ [58]^23 ^] uses unsupervised GCNs to learn a cell
embedding representation based on graphs extracted from spatial
transcriptomics data. However, although these methods combine gene
expression and spatial information to identify spatial domains, they
are all based entirely on unsupervised learning, resulting in the
learned embedding representations often lacking the fidelity required
to accurately segment spatial domains, making it prone to generate
discrepancies between identified domains and pathology annotations.
In recent years, various methods rooted in self‐supervised learning
have emerged to enhance the optimization of embedding representations
and elevate the efficacy of spatial domain identification. SpaceFlow^[
[59]^24 ^] uses a spatially regularized deep graph network to combine
gene expression similarity with spatial information through a
contrastive learning strategy. conST^[ [60]^25 ^] explores multiple
patterns of ST data through a contrastive learning framework in two
training stages and learns to embed representations beneficial for
diverse downstream tasks. Similarly, GraphST^[ [61]^26 ^] introduces
self‐supervised graph contrastive learning to learn informative
representations of gene expression profiles alongside their spatial
coordinates, thus enriching latent representation learning. ConGI^[
[62]^27 ^] deciphers the spatial domains by adapting gene expression to
histopathology images through contrastive learning. However, these
methods typically lack ability to fully utilize the spatial information
within their self‐supervised contrastive learning designs, resulting in
inadequate guidance of model training by spatial information, thus
perpetuating the difficulty of accurately identifying spatial domains.
Moreover, existing methods are unable to optimize the latent embedding
representations at both spatial and gene levels, thereby ignoring many
key gene regulations.
Here, we develop stDCL, a graph convolutional autoencoder framework
that uses dual contrastive learning to identify spatial domains from
spatial transcriptomics data. stDCL integrates spatial information and
gene expression information through graph convolutional networks (GCNs)
to learn the latent embedding representations and allow imputation of
ST data. Then, stDCL employs dual graph contrastive learning
(space‐aware contrastive learning and cluster‐level feature contrastive
learning) to enable the embedding representation and imputation matrix
to preserve the important spatial heterogeneity and gene regulation.
The space‐aware contrastive learning improves the similarity between
spots of close spatial location while reducing the similarity between
spots of distant spatial location within the representation space. The
cluster‐level feature contrastive learning brings the intra‐cluster
spots closer and further away from the inter‐cluster spots in terms of
feature dimensionality. Evaluation of stDCL's clustering performance on
the human dorsolateral prefrontal cortex (DLPFC) 10x Visium dataset and
frontal cortex MERFISH dataset demonstrated its superiority over eight
state‐of‐the‐art methods for identifying spatial domains. In
particular, stDCL exhibited robust performance even on complex tissue
structures and detected substructures not detected by other methods.
Additionally, we analyzed the interpretable functional genes in the
latent embedding representation of stDCL and aggregated their spatial
expression information to provide a reliable comprehension of multiple
regions. Furthermore, the imputation of stDCL could effectively
reconstruct the hierarchical structure of the spatial transcriptomics
data, enhancing layer‐specific differential gene identification while
mitigating noise. We also demonstrated that stDCL can reveal spatial
heterogeneity between different regions at high spatial resolution and
unravel patterns of brain development in mouse embryos. Moreover, stDCL
proved valuable in identifying cellular phenotypes associated with
Alzheimer's disease, elucidating the complex regulatory processes
underlying its pathology.
2. Results
2.1. Overview of stDCL
stDCL learns the low‐dimensional latent embedding representation from
spatial information and gene expression information by dual graph
contrastive losses. As depicted in Figure [63] 1 , the workflow of
stDCL comprises four principal stages: (1) constructing a
transcriptomics profile‐based spatial graph that merges gene expression
and spatial data from spatial transcriptomics (ST) data; (2) deriving
the low‐dimensional latent representations using a graph convolutional
autoencoder; (3) bringing spatially neighboring spots closer and
separating non‐neighboring spots via space‐aware contrastive learning;
and (4) utilizing cluster‐level feature contrastive learning to augment
the correlation between homologous dimensional genes based on
cluster‐level representations derived from the readout function.
Figure 1.
Figure 1
[64]Open in a new tab
Overview of stDCL. a) The graph construction process for spatial
transcriptomics data via spatial location and gene expression. b) stDCL
learns the low‐dimensional latent representation and the imputation
matrix with spatial information and gene expression information via the
graph convolutional autoencoder. c) stDCL uses a weight‐shared GCN
encoder to encode the corrupted gene expression matrix generated by
data augmentation and obtains the latent embedding representations in
different views. The space‐aware contrastive learning is used to
further learn the relationships between spots on the latent
representation. d) stDCL uses the cluster‐level feature contrastive
learning to strengthen the correlation of the same dimension features
in two views and improve the clustering performance of the model. e)
The application capabilities of latent embedding representations
trained and imputation matrices by stDCL.
stDCL first constructs a transcriptomics profile‐based spatial graph by
integrating gene expression information and spatial information to
jointly characterize and balance the relationship between spots at
gene‐specific and spatial levels (Figure [65]1a). Subsequently, the
process continues with the use of a preprocessed gene expression matrix
and the neighborhood graph as inputs into graph convolutional networks
(GCNs), serving both as encoder and decoder to effectively preserve the
underlying topological structure inherent in the latent embedding
representations (Figure [66]1b). At the end, through optimizing the
expression matrix and the loss function associated with graph
reconstruction, stDCL learns to derive low‐dimensional latent embedding
representations and imputation matrices, leveraging the spatial and
gene expression information.
In addition, we propose a dual graph contrastive learning model in
stDCL to learn graph‐level discriminative representations efficiently.
We first generate a corrupted gene expression matrix by adding Gaussian
noise to the original gene expression matrix to obtain different views
of genetic characterization for data augmentation. Then, stDCL feeds
the corrupted gene expression matrix into the shared encoder, yielding
cross‐view latent characterizations. Further, to preserve the
underlying spatial similarities and topological relationships between
cellular spots, we propose a space‐aware contrastive learning approach
that harnesses spatial information to ensure alignment between the
spots in the embedding representations and their true spatial
distribution (Figure [67]1c). Specifically, we calculate the mean
squared error (MSE) of the cosine similarity matrix derived from the
cross‐view representations utilizing the neighborhood graph,
effectively assigning higher similarity scores to spots in close
spatial proximity and lower scores to those in distant spatial
locations during the training process. On this basis, we also propose a
cluster‐level feature contrastive learning strategy to enhance the
spatial clustering proficiency of our model (Figure [68]1d), which
utilizes a readout function to calculate cluster‐level embedding
representations, achieved by averaging the cross‐view representations
for each cluster. Subsequently, we compute the mean squared error (MSE)
between the identity matrix and the cosine similarity matrix of the
cluster‐level embedding representations to draw representations of the
identical dimensional features closer together. Finally, the
cluster‐level feature contrastive learning approach is proposed to
strengthen intra‐cluster connectivity and weakens inter‐cluster
connectivity from the perspective of gene dimension.
2.2. stDCL Accurately Identifies Spatial Heterogeneity from Spatial
Transcriptomic Profiling of DLPFC
To evaluate the spatial clustering performance of stDCL, we used it on
the human dorsolateral prefrontal cortex (DLPFC) 10x Visium dataset^[
[69]^28 ^] from brain tissue, which contains spatially resolved
transcriptomic profiles of 12 DLPFC slices. The cortical layers (L1‐L6)
and the white matter (WM) regions within each slice have been carefully
identified by manual annotation, relying on morphological
characteristics and genetic markers, as described by Maynard et al. ^[
[70]^28 ^] To ensure a comprehensive comparison, we conducted a
comparative analysis of stDCL against eight state‐of‐the‐art methods
(Seurat,^[ [71]^29 ^] SpaGCN,^[ [72]^21 ^] stLearn,^[ [73]^19 ^]
BayesSpace,^[ [74]^20 ^] SEDR,^[ [75]^30 ^] CCST,^[ [76]^23 ^]
STAGATE,^[ [77]^22 ^] and GraphST^[ [78]^26 ^]) on the DLPFC dataset.
Figure [79] 2a shows the results using the Adjusted Rand Index (ARI)
metric to compare performance between stDCL and the other methods on
the 12 tissue slices. We observe that stDCL achieved the highest ARI
score on 8 of the 12 datasets. We depict the overall clustering
performance of the different methods on the 12 tissue slices using
boxplots, where the middle line and white diamond represent the median
and mean ARI values, respectively. On all 12 slices, stDCL showed
superior performance with the highest mean (ARI = 0.61) and median (ARI
= 0.61) scores, exhibiting the least variance among the methods
evaluated, as depicted in Figure [80]2b. GraphST also demonstrated
remarkable spatial clustering performance, with a median score reaching
0.59. In contrast, the mean and median scores of the other methods were
not above 0.50, with both BayesSpace and CCST displaying high variance,
with significant performance fluctuations across the several tissue
slices. For a more detailed assessment, we further utilized the
normalized mutual information (NMI) and homogeneity score (HS) metrics
to evaluate the clustering performance of stDCL (Figure [81]S1,
Supporting Information). Notably, the experimental results indicated
that stDCL outperformed the other methods with both metrics.
Figure 2.
Figure 2
[82]Open in a new tab
stDCL accurately identifies spatial heterogeneity from spatial
transcriptomic profiling of DLPFC. a) Comparison of ARI values between
stDCL and 8 spatial transcriptomics clustering methods on 12 DLPFC
slices (n = 1 in each group). b) The box plot of clustering accuracy on
all 12 DLPFC slice datasets in terms of ARI score for 9 methods (n = 12
in each group; center line, median; box limits, upper and lower
quartiles; whiskers, 1.5× interquartile range). c) The manually
annotated layer structure compared to spatial domains detected by
Seurat, SpaGCN, stLearn, BayesSpace, SEDR, CCST, STAGATE, GraphST, and
stDCL on slice 151508.
We next take tissue slice 151508 as an example, which contains 4384
spots, both stDCL and STAGATE achieved superior clustering performance,
and successfully identified the spatial domains that closely match the
manually annotated tissue layers (Figure [83]2c). In contrast, methods
such as Seurat and SEDR encountered difficulty in clearly defining
layer boundaries, leading to numerous outliers that compromised the
spatial structuring in the brain region. SpaGCN failed to separate the
spots within layers L4‐L6 based on the hierarchical structure, whereas
for stLearn and CCST, cluster most spots in layers L4‐L6 into a single
expansive cluster. Overall, stDCL can effectively identify the cortical
layer structures in the DLPFC dataset with stability and robustness
across several slices. Figures [84]S2– [85]S4 (Supporting Information)
detail the outcomes for the additional slices.
2.3. stDCL Provides Better Clustering Performance on Complex Spatial Domains
and Tissue Structures
In the next benchmark experiment, we applied stDCL to two 10x Visium
datasets with more complex tissue structures (the mouse anterior brain
dataset and the human breast cancer dataset), and compared it with
other state‐of‐the‐art clustering methods. Long et al. ^[ [86]^26 ^]
provided manual annotations for this mouse dataset in accordance with
the Allen Mouse Brain Reference Atlas, identifying 52 clusters, as
depicted in Figure [87] 3a. The experimental results revealed that
stDCL outperformed the other methods in terms of clustering performance
on this dataset (Figure [88]3b; Figure [89]S5, Supporting Information).
From the spatial clustering results, it becomes evident that only stDCL
succeeded in accurately capturing the ‘CPu’ domain and fully
identifying the “St” domain, while other methods produced aberrant
tissue structures or disorganized spot distributions. Furthermore, the
clusters identified by Seurat, SpaGCN, and SEDR showed spatial
fragmentation, and the boundaries defined by stLearn and BayesSpace
were indistinct. Of the remaining three methods, CCST and STAGATE
introduced unnecessary layer structures, while GraphST formed
discontinuous clusters. In summary, on spatial transcriptomics datasets
with intricate organizational structures, stDCL demonstrated its
capacity to effectively cluster spatial domains that align with
biological annotations and identify several substructures.
Figure 3.
Figure 3
[90]Open in a new tab
Spatial clustering performance on datasets with complex tissue
structures.) The spatial domains of the mouse anterior brain dataset
manually annotated by the Allen Mouse Brain Reference Atlas. b) The
spatial domains detected by different methods on the mouse brain
dataset. c) H&E images and manual annotation on the human breast cancer
dataset. d) Comparisons of ARI values between stDCL and 8 spatial
transcriptomics clustering methods on the human breast cancer dataset
(n = 1 in each group). e) Spatial domains detected by different methods
on the human breast cancer dataset.
For the human breast cancer dataset, the dataset was manually annotated
with 20 regions by pathologists based on the H&E image and reported
breast cancer marker genes (Figure [91]3c). stDCL still maintained the
best clustering performance on this dataset. The ARI value of stDCL
improved by 7% compared with other methods (Figure [92]3d), and it also
outperformed other methods in the evaluation metrics of NMI and HS
(Figure [93]S6, Supporting Information). The spatial domains identified
by stDCL were the closest to manual annotation (Figure [94]3e).
Specially, for regions “IDC 5” and “Healthy 1,” only stDCL clearly
captures most of the spots as a cluster, while other methods failed.
Besides stDCL, CCST also had higher ARI value and better spatial
clustering performance on the human breast cancer dataset, while
Seurat, SpaGCN and STAGATE performed poorer clustering results on this
dataset. In summary, on spatial transcriptomics datasets with complex
organizational structures, scDCL was able to cluster spatial domains
consistent with biological annotations and identify some substructures.
2.4. stDCL Efficiently Demarcates Tissue Structures in Spatial Transcriptomic
Profiling Across Different Platforms
To test the spatial clustering performance of stDCL using various
spatial transcriptomics technologies, we used it on different platforms
to analyze the MERFISH dataset,^[ [95]^31 ^] with 9 tissue slices from
the frontal cortex and striatum of mouse brain of diverse ages. From
the 9 slices, slices 4‐0, 4‐1, 4‐2, 6‐0, 6‐1, and 6‐2 correspond to
mouse brain at 4 weeks, and slices 8‐0, 8‐1, and 8‐2 are mouse brain at
90 weeks. Distinct tissue regions were manually delineated on each
slice, indicating the corpus callosum, pia mater, striatum, olfactory
region, brain ventricle, cortical layer V, cortical layer VI, and
cortical layer II. We evaluated the performance of the algorithm on
different platforms and further conducted a comparative analysis
comparing stDCL with four well‐known spatial clustering methods,
including GraphST, STAGATE, CCST, and SEDR. As depicted in Figure [96]
4a, stDCL demonstrated superior spatial clustering performance on all 9
tissue slices compared to other methods, with stDCL achieving the
highest ARI on 7 of them. In terms of overall clustering performance,
stDCL exhibited the highest mean score (ARI = 0.63) and median (ARI =
0.62) of all methods across all 9 slices (Figure [97]4b). Notably,
stDCL attained an ARI of 0.78 on slice 8‐0, significantly outperforming
the other methods. The mean and median scores of the other methods were
below 0.55, with SEDR proving ineffective for spatial clustering on the
MERFISH dataset. Additionally, we employed NMI and HS metrics to
provide a comprehensive assessment of the performance of stDCL on the
MERFISH dataset (Figure [98]S7, Supporting Information). The
experimental findings indicated that stDCL surpasses the other methods
measured by both metrics, and notably, stDCL achieved the highest HS
scores on all slices.
Figure 4.
Figure 4
[99]Open in a new tab
stDCL efficiently demarcates tissue structures in spatial
transcriptomic profiling across different platforms. a) Comparisons of
ARI values between stDCL and 4 spatial transcriptomics clustering
methods on 9 slices from the MERFISH dataset (n = 1 in each group). b)
The box plot of clustering accuracy on all 9 slice datasets in terms of
ARI score (n = 9 in each group; center line, median; box limits, upper
and lower quartiles; whiskers, 1.5× interquartile range). The manually
annotated tissue structure compared to spatial domains detected by
stDCL, GraphST, STAGATE, CCST and SEDR on c slice 4‐0, d slice 6‐0 and
e slice 8‐0.
In the young mouse brain slices 4‐0 and 6‐0, and in an old brain slice
8‐0, stDCL accurately identified the spatial domains, aligning closely
with the manually annotated tissue structures (Figure [100]4c–e),
distinguishing regions like the corpus callosum, pia mater, and
striatum in both ages, which was challenging for the other methods.
Specifically, the pia mater identified by GraphST tended to be
conflated with other tissue structures, and the clarity of the striatum
identification in old slice 8‐0 was markedly compromised compared to
stDCL. STAGATE and CCST, failed to identify the pia mater, and the
corpus callosum identified in young brain slice 6‐0 and striatum in old
brain slice 8‐0 were incomplete. The spatial domains found by SEDR are
relatively disordered and the spatial clustering performance poor.
Moreover, Figures [101]S8–[102]S10 (Supporting Information) provide
results for all other slices for an exhaustive evaluation. We aslo
applied stDCL to the slices from Adult mouse central nervous system,
obtained using the STARMap Plus technique.^[ [103]^32 ^] Distinct
tissue regions were manually delineated on three slice, indicating the
CB 1, CB 2, CTX 1, CTX 2, DG, and so on. As depicted in
Figure [104]S11a (Supporting Information), stDCL demonstrated superior
spatial clustering performance on all 3 tissue slices compared to other
methods. In mouse central nervous system slices, stDCL accurately
identified spatial domains that were tightly aligned with manually
annotated tissue structures (Figure [105]S11b, Supporting Information).
Notably, stDCL is able to successfully detect HY regions, whereas other
methods either failed to detect or incorporated HY regions into other
regions. For the other methods, GraphST is also able to identify most
of the spatial domains, while CCST and SEDR detected regions that are
more ambiguous. In conclusion, stDCL consistently exhibits effective
performance in spatial clustering on several datasets utilizing various
spatial transcriptomics technologies, and maintaining robust
performance with different age of mouse brain.
We also compared the runtime of stDCL with other clustering algorithms.
For the test datasets, we adopted 19 datasets with 1200–30 000 cells
(Figure [106]S12, Supporting Information). As shown in Figure [107]S12
(Supporting Information), stDCL, GraphST, and SEDR demonstrate
comparable execution times, while CCST exhibits the longest runtime
among the methods. Notably, CCST's runtime escalates sharply once the
cell count surpasses 5000, indicating a steep increase in computational
demand at higher cell densities. For STAGATE, the growth rate in
running time is particularly pronounced between 10 000 and 20 000
cells, highlighting its sensitivity to increased cell numbers. Although
stDCL's runtime is generally longer than that of GraphST, this is
attributed to the added computational complexity introduced by its dual
graph contrastive learning mechanism, which enhances model performance
at the cost of increased processing time. In addition, we compared the
memory usage of these methods at different numbers of cells
(Figure [108]S13, Supporting Information). In Figure [109]S13
(Supporting Information), we can observe that stDCL and GraphST
maintain relatively consistent memory usage, even as the number of
cells scales up, suggesting their efficiency in managing computational
resources. However, CCST has significantly higher memory requirements,
especially when the number of cells exceeds 5000, suggesting that its
algorithmic structure may need to allocate a large amount of memory for
larger datasets. SEDR maintains a lower memory usage. As the number of
cells increases from 20 000 to 25 000, STAGATE's memory usage also
increases significantly, reflecting the method's growing demand on
resources for larger data sizes. In summary, despite incorporating dual
graph contrastive learning, stDCL maintains runtime and memory usage
within a manageable range.
2.5. The Latent Embedding Characterization of stDCL Aggregates Interpretable
Spatial Expression Information
To investigate whether stDCL can capture critical spatial
transcriptomics information, we conducted genome interpretability
analysis on the latent embedding representation of stDCL using the
mouse forebrain 10x Visium dataset. Initially, we identified two
regions based on the manual annotation, namely MOB::ONL and VL
(Figure [110] 5a). We then utilized an interpretable gene selection
method^[ [111]^33 ^] informed by the standard deviation ranking of GNN
weight matrices to select the top 1000 most highly expressed genes from
the latent embedding representations of stDCL, GraphST, and STAGATE.
Subsequently, we visualized the top 5 most highly expressed genes in
regions MOB::ONL and VL using the spatial clustering methods and
compared with the highly expressed genes obtained by manual annotation
(Figure [112]5b). It can be observed that stDCL correctly identified
the highly expressed genes in the VL region without mistaking genes
from other spatial domains. Furthermore, stDCL detected highly
expressed genes, such as Igf2, which were not identified by the other
methods or even the manual annotation. For the olfactory bulb (MOB)
olfactory nerve layer (ONL) MOB::ONL region, both stDCL and manual
annotation found the highly expressed gene S100a5, while the other
methods failed. To verify the significance of S100a5 expression in the
MOB::ONL region, we visualized the expression of S100a5 in the mouse
olfactory bulb datasets profiled by Stereo‐seq^[ [113]^34 ^] and
Slide‐seqV2,^[ [114]^35 ^] respectively (Figure [115]5c). As
illustrated in Figure [116]5c, the S100a5 gene is clearly highly
expressed in the olfactory nerve layer. Recent studies have also
demonstrated that S100a5 is mainly expressed in the olfactory bulb and
olfactory sensory neurons (OSNs), with expression significantly
upregulated in response to olfactory stimulation.^[ [117]^36 , [118]^37
^]
Figure 5.
Figure 5
[119]Open in a new tab
The latent embedding characterization of stDCL aggregates interpretable
spatial expression information. a) The spatial domains of the mouse
anterior brain dataset manually annotated by the Allen Mouse Brain
Reference Atlas (left). b) The dotplot of the top 5 highly expressed
genes in MOB::ONL and VL in manual annotation, stDCL, GraphST and
STAGATE. c) The visualization of expression of S100a5 on mouse
olfactory bulb datasets, profiled by Stereo‐seq and Slide‐seqV2,
respectively. d) Visualization of the spatial expression of genes Igf2
and S100a5 denoised by stDCL.
To further evaluate the ability of stDCL to aggregate and characterize
the spatial expression information of the data, we computed the 15
nearest neighbors for each spot based on the Euclidean distance in the
latent embedding representation of stDCL and then averaged the gene
expression data for these neighbors to yield a denoised gene expression
matrix. As depicted in Figure [120]5d, compared to the raw gene
expression data, the expression levels of the interpretable genes Igf2
and S100a5, identified through the latent embedding representation,
were elevated in their respective regions, and their enrichment levels
were higher. Overall, these experimental results demonstrate that the
latent embedding characterization of stDCL can uncover important
interpretable genes and aggregate their spatial expression information,
offering insights into the molecular architecture and function
of tissue.
2.6. stDCL Efficiently Recovers Gene Expression to Characterize Spatial
Hierarchy and Differential Gene Expression in the Mouse Visual Cortex
To investigate the reconstruction performance of stDCL in spatial
hierarchy structure and denoising and enhancement of differential gene
expression, we used mouse primary visual cortex (V1) profiled by
STARmap.^[ [121]^38 ^] The dataset comprises 1,207 cells and 1,020
genes, and the layer structure and cell types of the tissue slices are
depicted in Figure [122] 6a. We first applied STAGATE, GraphST, and
stDCL to the STARmap dataset and compared their clustering performance
(Figure [123]S14, Supporting Information). The results demonstrated
that stDCL surpassed the other methods. To evaluate the performance on
reconstructing the hierarchical structure through the original spatial
information, we first projected the imputation matrix generated by the
three methods onto two dimensions as the imputation space using the
scanpy package. We proceeded to compare the correlation of pairwise
distances in the imputation space generated by stDCL, GraphST, and
STAGATE with the original space, as depicted in Figure [124]6b. We see
that the imputation space of stDCL exhibited a significantly higher
correlation with the original space compared to the other two methods,
with a Pearson correlation coefficient (PCC) of 0.502, while STAGATE
and GraphST achieved PCC values of 0.298 and 0.345, respectively. We
also used other correlation metrics for systematic comparison, and
stDCL is the best in all four metrics (Figure [125]S15, Supporting
Information). In addition, the imputation space of stDCL preserved the
relative position between the L1‐L6 layers, that is, the distance to
the L1 layer gradually increased from the inside to the outside,
whereas the hierarchies reconstructed by STAGATE and GraphST did not
show this tendency (Figure [126]S16, Supporting Information).
Figure 6.
Figure 6
[127]Open in a new tab
stDCL can efficiently recover gene expression to characterize spatial
hierarchy and differential gene expression in the mouse visual cortex.
a) The layer structure and cell types of the tissue slices of the mouse
primary visual cortex (V1) dataset. b) The density plot of the
correlation of pairwise distances between spots in the original space
and the imputation space generated by STAGATE, GraphST and stDCL. c)
The spatial expression of layer‐specific genes. d) The expression trend
of differential genes in each layer along the X‐axis in the original
space and the imputation spaces generated by STAGATE, GraphST and
stDCL, respectively.
We also assessed the performance of stDCL imputation on denoising and
enhancing differential gene expression results. Specifically, we
focused on four layer‐specific differential genes, Pcp4 (L6), Rprm
(L6), Tmsb10 (L5) and Mapk3 (L1), and compared their spatial expression
levels in both the raw and imputed expression matrices (Figure [128]6c;
Figure [129]S17, Supporting Information). These genes were specifically
expressed at the corresponding spatial positions of the imputation
matrix, whereas their raw spatial expression was inconspicuous and
dispersed. We further investigated the expression trends of these
layer‐specific genes along the X‐axis in the imputation space and the
original space. As illustrated in Figure [130]6d, the imputation space
of stDCL effectively preserved most of the expression patterns of genes
along the layer axis, closely resembling the patterns observed in the
original space. For other methods, GraphST retained only a part of the
gene expression trend, while STAGATE displayed a broader trend. In
particular, all three methods substantially preserved the expression
pattern of the gene Tmsb10. Overall, imputation of stDCL not only
successfully reconstructs the hierarchical structure of the spatial
transcriptomics data, but also clarifies and strengthens the
identification of layer‐specific differential genes to unravel spatial
intricacies in the mouse visual cortex.
2.7. Evaluation of Hyperparameter Selection and Ablation Analysis
We evaluated the effect of the number of selected genes, the number of
neighbors and their metric distance used to construct the neighborhood
graph, the loss weights, the network framework and other parameters on
the clustering performance of stDCL on 12 DLPFC slice datasets. In
addition, we investigated the effectiveness of each component of stDCL
on the same dataset.
The clustering performance of stDCL was evaluated using varying numbers
of selected genes and different numbers of neighbors, as depicted in
Figure [131] 7a and Figure [132]S18a (Supporting Information). The
results revealed that stDCL achieved optimal clustering when utilizing
3000 genes as input and constructing the neighborhood graph with three
neighbors. Subsequently, stDCL was trained by three loss functions
simultaneously, namely reconstruction loss, space‐aware contrastive
loss, and cluster‐level feature contrastive loss. To ensure a
harmonious balance of the overall loss function, we assigned weights to
three parameters (γ[1], γ[2], and γ[3]) and explored various
combinations of weights. Specifically, in accordance with our previous
experiences, we selected γ[1], γ[2], and γ[3] in the range [1.0, 5.0,
10.0], [0.2, 0.5, 0.8], or [0.2, 0.5, 0.8]. On this basis, we
enumerated weights to obtain 27 distinct loss weight assignments and
compared the clustering performance in each case (Figure [133]7b).
Notably, among the 27 configurations, composition 24 yielded the
highest average ARI value, where γ[1], γ[2], and γ[3] were set to [10,
0.5, 0.8], respectively. We also employed NMI and HS metrics for
further comparisons (Figure [134]S19, Supporting Information). We also
systematically analyze the parameters alpha [1] and alpha [2] of the
reconstruction loss and give the best settings as [1.0, 0.5]
(Figure [135]S20, Supporting Information). To construct the
neighborhood graph, we tested the clustering performance using
different methods and distance metrics. As shown in Figure [136]7c and
Figure [137]S18b (Supporting Information), stDCL achieved optimal
clustering performance when constructing a KNN graph using the
Euclidean distance metric. Regarding the network architecture, we
evaluated the clustering performance of stDCL with varying numbers of
hidden layers, as depicted in Figure [138]7d and Figure [139]S18c
(Supporting Information). Experimental results indicated that stDCL was
most effective in clustering with a single hidden layer. In addition,
we also analyzed the number of nodes in the hidden layer, as
illustrated in Figure [140]S21 (Supporting Information).
Figure 7.
Figure 7
[141]Open in a new tab
Evaluation of hyperparameter selection and ablation analysis on 12
DLPFC slices. a) The clustering performance was evaluated using ARI
metric for different numbers of genes and numbers of neighbors (n = 12
in each group; center line, median; box limits, upper and lower
quartiles; whiskers, 1.5× interquartile range). b) The ARI bubble plot
under different loss weight assignments. c) The radar plot of
clustering performance for different ways and distance metrics of
constructing the neighborhood graph. d) The ridgeline plot of
clustering performance for different numbers of hidden layers in the
graph autocoder (n = 12 in each group). e) The ablation study for the
neighborhood graph and dual graph contrastive learning (n = 1 in each
group).
We conducted an ablation analysis on stDCL to assess the effectiveness
of the neighborhood graph and dual graph contrastive learning
mechanisms. Specifically, we compared the spatial clustering
performance of stDCL with its partially ablated counterpart using the
ARI metric. In the neighborhood graph module construction, we removed
two components: spatial location (“Without Spatial”) and the gene
expression matrix (“Without Gene”). In the contrastive learning module,
we evaluated three ablation cases: (1) replacing the space‐aware
contrastive loss function with the InfoNCE loss function (“InfoNCE”);
(2) removing the cluster‐level feature contrastive loss function
(“Without Clu”): (3) using only the InfoNCE loss function (“Without
Clu(Info)”). As shown in Figure [142]7e and Figure [143]S22 (Supporting
Information), the spatial clustering performance of stDCL outperformed
the other arrangements. In addition, the clustering performance of
stDCL exhibited greater stability, whereas the other arrangements
displayed some degree of fluctuation. In particular, the spatial
clustering performance of stDCL after removing gene expression
information during the construction of the neighborhood graph was
significantly higher than that of the other arrangements on slices
151671 and 151672. This implies that spatial information plays a
predominant role in the correlation of spots on these two slices, and
the addition of gene expression information has an adverse impact on
clustering performance. In summary, these experimental results
highlight the effectiveness of both the neighborhood graph construction
and the dual graph contrastive learning components in stDCL. To provide
a more comprehensive analysis, we investigated how dual graph
contrastive learning with stDCL contributes to the performance of
various tasks, including spatial heterogeneity dissection, spatial
hierarchy characterization and gene expression recovery. We evaluated
three ablation cases: (1) using only the space‐aware contrastive
learning (stDCL(Spa)); (2) using only the cluster‐level feature
contrastive learning (stDCL(Clu)): (3) using only the InfoNCE loss
function (stDCL(Info)). We study the performance of these cases on
various tasks in more detail and provided the analysis and experimental
results in Note [144]S8 (Supporting Information).
2.8. stDCL Reveals Molecular and Spatial Heterogeneity in Mouse Brain at
Single‐Cell Resolution
To evaluate whether stDCL can explore molecular and spatial
heterogeneity, we applied stDCL to analyze spatial transcriptomics data
from the mouse brain at single‐cell resolution. We collected a spatial
dataset obtained from the mouse somatosensory cortex acquired with
osmFISH,^[ [145]^39 ^] which contains 4839 cells and 33 genes. The real
annotations of this dataset, as illustrated in Figure [146] 8a, exhibit
its origin at the Pia Layer 1 and encompass 11 cell‐wise regions. We
compared the spatial clustering performance of STAGATE, GraphST and
stDCL on this dataset using NMI and ARI metrics, and then visualized
the clustering results (Figure [147]8b,c, Supporting Information). The
experimental results indicate that stDCL significantly outperformed
other methods, and the identified spatial domains are closer to the
real annotation. To elaborate, in comparison to the other two methods,
stDCL could distinguish Layer 4 and White matter more effectively,
rendering the layer boundaries more distinctly defined. Furthermore, we
validated that stDCL revealed heterogeneity in these spatial layers
through the expression of marker genes^[ [148]^39 ^] native to osmFISH
data (Figure [149]8d). For instance, Rorb, a marker gene for Layer 4
pyramidal neurons, exhibited the expected high expression within Layer
4 identified by stDCL. Plp1, a marker gene for oligodendrocytes,
displayed elevated expression within the white matter, where various
types of oligodendrocytes are known to reside. In addition, Flt1, a
marker gene for endothelial cells in Pia Layer 1, and Slc32a1, a marker
gene for interneurons in IC CP, were also found to be expressed as
anticipated within their respective spatial locations. These findings
collectively emphasize the capacity of stDCL to elucidate the molecular
and spatial heterogeneity, further revealing spatial gene expression
patterns. In addition, we also conducted trajectory inference on the
relatively well‐defined domains identified by stDCL, such as Layer 4,
Layer 6, and white matter (Figure [150]8e). The inferred trajectory
projects from White matter to Layer 4, aligning with the established
pattern of cortical development from the inside out,^[ [151]^40 ,
[152]^41 ^] where new neurons migrate vertically along the radial glial
fibers to the marginal area of the cortical periphery to form a new
cortex on top of the existing layers.
Figure 8.
Figure 8
[153]Open in a new tab
stDCL reveals molecular and spatial heterogeneity in mouse brain at
single‐cell resolution. a) The ground truth of regional annotation for
the osmFISH dataset. b) The spatial domains identified by STAGATE,
GraphST and stDCL. c) The clustering performance comparison of the
three methods. d) The visualization of domains identified by stDCL and
the corresponding marker genes. e) The visualization of the trajectory
inferred by stDCL.
Subsequently, we conducted an analysis of mouse hypothalamic preoptic
area tissue slices (Bregma‐0.14) obtained using the MERFISH Spatial
Profiling Technology.^[ [154]^42 ^] This dataset contains 5926 cells
and 155 genes, annotated across eight structural regions: bed nuclei of
the strata terminalis (BST), medial preoptic area (MPA), medial
preoptic nucleus (MPN), periventricular hypothalamic nucleus (PV),
paraventricular hypothalamic nucleus (PVH), and paraventricular nucleus
of the thalamus (PVT), third ventricle (V3), and columns of the fornix
(fx) (Figure [155] 9a). We examined the spatial domain clustering
results of STAGATE, GraphST, and stDCL (Figure [156]9b). It can be
observed that stDCL surpassed the other methods in its ability to
detect the majority of spatial domains, exhibiting a closer resemblance
to the histological annotations. Specifically, stDCL was adept at
identifying and delineating the BST region, achieving smoother
boundaries than those produced by alternative approaches. In addition,
stDCL was able to successfully identify the V3 region, a task that
STAGATE failed to perform in the lower part of the V3 region, and
GraphST incorrectly merged some V3 cells with other regions. Then, we
performed enrichment analysis on the top 20 differentially expressed
genes detected in the V3 region identified by stDCL (Figure [157]9c)
and detected marker genes for this region (Figure [158]S26, Supporting
Information). These genes are mainly enriched in response to glucose
and monosaccharide; the reason being there are basically ependymal
cells in the V3 region (Figure [159]9d), suggesting their potential
involvement in glucose and monosaccharide transport.^[ [160]^43 ,
[161]^44 ^] Moreover, ependymal cells, as a type of glial cell found in
the central nervous system, are an indispensable component of the
central nervous system.^[ [162]^45 ^] Therefore, the V3 region is also
enriched for Gene Ontology (GO) terms associated with glial cells and
the nervous system, such as glial cell projections and regulation of
nervous system processes. In summary, stDCL demonstrated an exceptional
capacity to accurately identify and characterize the complex molecular
and spatial heterogeneity embedding within the mouse brain at the
single‐cell resolution.
Figure 9.
Figure 9
[163]Open in a new tab
stDCL reveals molecular and spatial heterogeneity in mouse brain at
single‐cell resolution. a) The manual annotation of mouse hypothalamic
regions with reference to the Allen Mouse Brain Atlas. b) The
identified spatial domains are shown for STAGATE, GraphST and stDCL. c)
The GO enrichment analysis (two‐sided Wilcoxon test and adopt BH to
adjust p‐values for multiple comparisons). d) The true cell annotation
of mouse hypothalamic MERFISH dataset.
2.9. stDCL Unravels Developmental Patterns in the Mouse Embryo Brain
To investigate developmental patterns in the mouse embryonic brain, we
applied stDCL to analyze multiple mouse embryo Stereo‐seq datasets^[
[164]^34 ^] acquired at the E9.5 and E10.5 time points. We first
performed clustering analysis on the E1S1 slices of E9.5 embryos using
STAGATE, GraphST, and stDCL, followed by a comparison with the ground
truth annotations (Figure [165] 10a,b). The reason for setting the
number of clusters higher than the number of annotated tissue regions
was to obtain higher resolution identification of tissue regions. As
depicted in Figure [166]10b, the clustering results of stDCL exhibited
a closer alignment with the annotations, enabling the clear detection
of specific regions such as heart, liver, and dermatomyositis. In
addition, we analyzed the correlation between the clusters identified
by stDCL. Notably, stDCL uncovered a group of highly correlated
clusters concentrated in the brain region (Figure [167]10c). To explore
developmental patterns in the mouse embryonic brain, we integrated this
group of clusters as a significant brain region detected by stDCL and
investigated its distinctions from other regions. We first calculated
the differentially expressed genes within this brain region and
visualized the expression levels of the top 15 up‐regulated genes
(Figure [168]10d,e; Figure [169]S27, Supporting Information). These
genes are clearly highly expressed in the brain region compared to
other regions. We further constructed a Protein–Protein Interaction
(PPI) network for the up‐regulated differentially expressed genes via
STRING^[ [170]^46 ^] and extracted the top‐scoring gene modules from
the network using Molecular Complex Detection (MCODE)^[ [171]^47 ^]
(Figure [172]10f). Additionally, we conducted Gene Ontology (GO)
enrichment analysis, revealing that the highest‐scoring gene modules
were associated with neurodevelopment (Figure [173]10g). In the early
stages of mouse brain development, many neurons are generated, thus
nervous system development plays an important role at this time.^[
[174]^48 ^] Moreover, several pathways related to brain development,
including neuronal development, neuronal differentiation, and forebrain
development, were also enriched. These findings indicate that stDCL
reveals the developmental patterns inherent in the mouse embryo.
Figure 10.
Figure 10
[175]Open in a new tab
stDCL unravels developmental patterns in the mouse embryo brain. a) The
ground truth of regional annotation for the E1S1 section of E9.5
embryos. b) The spatial domains identified by STAGATE, GraphST and
stDCL. c) The correlation matrix between clusters identified by stDCL.
A group of highly correlated clusters concentrated in the brain region
is marked by a black box. d) The violin plot of expression levels of
the top 15 up‐regulated genes. e) The visualization of expression
levels of the top up‐regulated genes in the brain region identified by
stDCL. f) The PPI network of up‐regulated differential genes. The
purple marked part is the highest scoring gene module in the network
calculated by MCODE. g) The heat map of GO enrichment analysis.
2.10. stDCL Uncovers Biological Connections Across Multiple Mouse Embryo
Slices
We performed an extensive clustering analysis on four distinct
embryonic slices, denoted as E2S1, E2S2, E2S3, and E2S4, from E9.5
embryos (Figure [176] 11a; Figure [177]S28, Supporting Information).
The experimental results show that the clustering performance of stDCL
was better than other clustering methods overall and can correctly
identify most regions. On this basis, we examined the correlation
between the clusters identified by stDCL on these embryo
sections (Figure [178]11b). Remarkably, stDCL revealed a consistent
correlation between the regions identified within the brain and those
present within the spinal cord on all four slices (Figure [179]11c). To
investigate the underlying biological significance, we aggregated these
regions for KEGG pathway analysis of differentially expressed genes
within the integrated regions (Figure [180]11d; Figure [181]S29,
Supporting Information). It can be observed that up‐regulated
differentially expressed genes are enriched for axon guidance and the
Wnt signaling pathway in all four sections. This observation highlights
their pivotal roles in axon guidance and Wnt signaling in the complex
process of neuronal development, particularly in forming commissural
circuits connecting the brain and spinal cord within the CNS.^[
[182]^49 ^] The axon guidance mechanism is key to accomplishing this
task.^[ [183]^50 , [184]^51 ^] In addition, the Wnt signaling pathway
is also essential for the formation of neuronal circuits.^[ [185]^52 ^]
To elucidate further, the Wnt signaling controls neuronal polarity,
promotes axon growth, and regulates axon navigation to the final
synaptic target.^[ [186]^53 , [187]^54 ^]
Figure 11.
Figure 11
[188]Open in a new tab
stDCL uncovers biological connections across multiple mouse embryo
slices. a The ground truth of regional annotation for the E1S1, E2S1,
E2S2, E2S3, and E2S4 section of E9.5 embryos. The spatial domains
identified by stDCL, GraphST, STAGATE, CCST and SEDR. b The correlation
matrix between clusters identified by stDCL. Marked by the black box is
a group of highly correlated clusters concentrated in the brain region.
c The spatial domains identified by stDCL on four sections of E2S1,
E2S2, E2S3, and E2S4 in E9.5 embryos (left). The regions of correlation
between the brain and spinal cord identified by stDCL on all four
sections (right). d The KEGG pathway enrichment analysis of the
selected region on the E2S1 section. e The ground truth of regional
annotation for the E1S1 section of E10.5 embryos (left). The regions of
correlation between the brain and spinal cord identified by stDCL on
the E1S1 section (right). f The GSVA enrichment pathway diagram of the
selected region.
To provide a more comprehensive analysis, we extended our investigation
by conducting analogous clustering and correlation analyses on the E1S1
section of E10.5 embryos, which has a denser spinal cord region
distribution (Figure [189]11e; Figure [190]S30, Supporting
Information). We also found the consistent correlation between brain
and spinal cord regions on this slice, and the area of the correlated
region was larger. We performed gene set variation analysis (GSVA) on
the integrated regions and found that the Wnt signaling pathway was
predominantly enriched (Figure [191]11f). This demonstrates that
embryos at this stage, which have a denser spinal cord region, are
important for the formation of commissural circuits connecting the
brain and spinal cord within the CNS. In particular, we also enriched
for various Interleukin (IL) signaling pathways, including IL‐2, IL‐3,
IL‐6, and IL‐9. The IL signaling pathways highlight the importance of
immune‐neural interactions, especially in response to injury or
inflammation. This is particularly relevant for areas such as the
hypothalamus and spinal cord, which are involved in both neuroendocrine
regulation and immune responses. For example, in the spinal cord, IL‐6
is involved in the regulation of chronic pain, especially neuropathic
pain. It can promote inflammatory responses in the spinal cord by
activating microglia and astrocytes, thereby affecting pain
transmission.^[ [192]^55 ^] Collectively, these findings revealed that
stDCL uncovers the biological connections between the brain and spinal
cord in the mouse embryo.
2.11. stDCL Illustrates Underlying Regulatory Mechanisms in Alzheimer's
Disease
Alzheimer's disease (AD), a progressive neurodegenerative disorder, is
marked by distinct pathologies in brain tissue, including reactive
astrocyte changes. To illustrate the underlying regulatory mechanisms,
we applied stDCL to the STARmap PLUS dataset,^[ [193]^56 ^] which
comprises AD and control mouse brains from 8‐month and 13‐month
specimens (Figure [194] 12a). stDCL identifies several cell clusters,
including astrocytes, endothelial cells, microglia, oligodendrocytes,
etc (Figure [195]12b and Figure [196]S31, Supporting Information). We
integrated all astrocytes identified by stDCL and performed
subclustering analysis to obtain three distinct subpopulations Astro1,
Astro2, and Astro3 (Figure [197]12c). Notably, the proportion of Astro3
was higher in AD than in controls, and Astro3 was significantly
increased in AD from 8 to 13 months, suggesting a dynamic cellular
response in AD progression (Figure [198]12d). Compared to other
astrocyte subtypes, Gfap, Vim, Cd9, and Igfbp5 were significantly
upregulated in Astro3, which aligns with the disease‐associated
astrocyte (DAA) phenotype reported in literature.^[ [199]^57 , [200]^58
^] Gfap is primarily expressed in astrocytes and significantly elevated
in the brains of Alzheimer's disease (AD) patients, where it correlates
with gliosis and the presence of A β; plaques.^[ [201]^59 , [202]^60 ^]
As a potential blood biomarker for AD,^[ [203]^61 ^] Gfap could aid in
diagnosing patients and tracking disease progression and therapeutic
efficacy.^[ [204]^62 ^] Vim is expressed in response to damage in
brains, and its upregulation in Alzheimer's disease may play a role in
neuronal repair following synaptic disruption and dendrite
retraction,^[ [205]^63 ^] although its precise function in this context
remains to be clarified. To further validate the association of Astro3
with the DAA‐like cell population, we introduced single nuclei RNA
sequencing (sNuc‐seq) datasets of 10‐month‐old WT and AD (5xFAD) mouse
brains ([206]GSE143758),^[ [207]^64 ^] including 1179 WT astrocytes and
2079 disease‐associated astrocytes. We performed multimodal
intersection analysis (MIA)^[ [208]^64 ^] using astrocytes from the
sNuc‐seq dataset with astrocytes identified by stDCL in the STARmap
PLUS dataset. MIA showed that these AD astrocytes were significantly
enriched in the Astro3 subpopulation, WT astrocytes were predominantly
enriched in the Astro2 subpopulation, and these cells were not
significantly enriched in the Astro1 subpopulation (Figure [209]12g).
In addition, gene expression of Astro3 was also more relevant to AD
astrocytes compared to WT astrocytes. These results demonstrate the
strong association of the Astro3 subpopulation with AD, affirming its
alignment with the DAA‐like population and suggesting its proliferation
within AD cells as the pathology advances.
Figure 12.
Figure 12
[210]Open in a new tab
stDCL illustrates the underlying regulatory mechanisms in Alzheimer's
disease. a) The spatial domains identified by stDCL on the STARmap PLUS
dataset. b) The dot plot of marker gene expression levels in different
cell populations identified by stDCL. c) The UMAP visualization of
astrocyte subtypes. d) The comparison of different astrocyte subtypes
proportions in AD and control samples at different times. e) The violin
plot of expression levels of top marker genes in Astro3. f) The UMAP
visualization of expression levels of top marker genes in Astro3. g)
The MIA map of WT and AD astrocytes in sNuc‐seq and stDCL‐identified
astrocyte subtypes in ST (left). The violin plot of expression levels
of Astro3 gene in WT and AD astrocytes (right). h) The gene ontology
enrichment analysis results of DAA‐like populations in astrocytes. i)
The visualization of pseudotime trajectories of different astrocyte
subtypes.
Then, to further explore the AD‐related gene regulation in the Astro3
subpopulation, we performed GO enrichment analysis (Figure [211]12h) on
the differentially expressed genes. It can be observed that
differentially expressed genes in Astro3 were involved in glial cell
development and proliferation, amyloid fibril formation, and protein
import, all of which are relevant to pathological processes. In
addition, these genes were mainly enriched in the regulation of tau
protein kinase activity, the imbalance of tau kinase and phosphatase
activity is believed to lead to tau hyperphosphorylation in the
disease,^[ [212]^65 ^] which is a major pathological characteristic of
AD. Additionally, we also performed pseudotime trajectory inference
across the three astrocyte subtypes, unveiling that Astro1 and Astro2,
which are not associated with disease, represent the initial state of
astrocytes, and progressively merge into the branch of the DAA‐like
subtype Astro3, which is consistent with the trend of disease
progression (Figure [213]12i).
3. Discussion
Spatial transcriptomics enables the quantification of gene expression
while simultaneously capturing spatial gene locations. The integration
of gene expression and spatial information aids in the identification
of spatial gene expression clusters, offering a pathway to elucidate
spatial heterogeneity. Nevertheless, the effective integration of
spatial data remains an ongoing challenge. Here, we have developed a
graph‐based contrastive learning method called stDCL to identify
spatial domains and uncover spatial heterogeneity. In particular, dual
graph contrastive learning has been proposed to synergize both the
space‐aware contrastive learning and cluster‐level feature contrastive
learning for the single‐cell spatial dissections of tissues.
Specifically, stDCL accurately identifies spatial domains in the human
dorsolateral prefrontal cortex tissue, outperforming eight other
state‐of‐the‐art methods in spatial clustering performance. In
particular, stDCL can identify spatial domains that closely match with
human annotations consistently on the datasets generated from different
spatial transcriptomics techniques, while maintaining stable
performance across different age groups. Based on the latent embedding
representation, stDCL can uncover important interpretable gene
candidates in tissue and aggregate their spatial expression
information. Additionally, we have demonstrated stDCL's effectiveness
in reconstructing hierarchical structures and enhancing differential
expression analysis on the STARmap dataset. Compared to STAGATE and
GraphST, the imputation matrix generated by stDCL preserves the
relative position between layers in the original space and accurately
reconstructs the expression patterns of individual layer‐specific
differential genes. Furthermore, stDCL successfully characterized the
complicated molecular and spatial heterogeneity within the mouse brain
at single‐cell resolution.
Notably, stDCL can unravel developmental patterns in the brain of mouse
embryos. We applied stDCL to identify spatial domains in the E1S1 slice
of E9.5 embryos. Interestingly, stDCL found a cluster of highly
correlated regions concentrated within the brain area, elucidating gene
regulation relevant to early stages of mouse brain development. In the
downstream cluster analysis of additional slices, stDCL revealed the
correlations between identified regions within the brain and spinal
cord, highlighting the co‐enrichment of crucial axon guidance
mechanisms involved in forming commissural circuits connecting these
two regions. To provide a comprehensive analysis, we employed the E1S1
slice of E10.5 embryos for validation, confirming the significant
pathways implicated in the formation of neuronal circuits.
Furthermore, stDCL can reveal the underlying regulatory mechanisms in
Alzheimer's disease. We used stDCL to perform spatial clustering on
Alzheimer's disease (AD) datasets, annotating astrocytes using marker
genes. From the analysis of the astrocyte subgroup, it was found that
the Astro3 subtype represented a higher proportion of diseased cells
compared to control cells. Therefore, we further verified that subtype
Astro3 was a DAA‐like population through the sNuc‐seq dataset and
explored its AD‐related biological processes. In addition, we conducted
pseudotime trajectory analysis to infer trends in astrocyte changes
during disease progression.
In summary, stDCL is a versatile approach based on dual graph
contrastive learning to jointly analyze spatial transcriptomics
datasets across various regions, considering both gene expression and
spatial distribution. In the complex spatial structure, stDCL
demonstrates the ability to identify spatial domains, enhance the
interpretability of genes, recover spatial hierarchy, elucidate spatial
heterogeneity, unveil developmental patterns, and elucidate disease
regulatory mechanisms. We will continue to advance and refine the
capabilities of stDCL with open‐source availability and reproducibility
in the future.
4. Experimental Section
Data processing
In this study, stDCL performs spatial clustering by utilizing both the
gene expression matrix and spatial location information derived from
spatial transcriptomics data. To elucidate this methodology further, a
logarithmic transformation was initially applied to the raw gene
expression matrix and subsequently normalized it based on library size
using the SCANPY software package.^[ [214]^66 ^] After normalization,
the gene expression counts were further scaled to unit variance and
zero mean. For 10x Genomics and Stereo‐seq datasets, the top 3000
highly variable genes as input were carefully selected. In the case of
datasets with a smaller number of genes, such as STARmap, osmFISH, and
MERFISH, it was opted to utilize all available genes as input without
any additional selection process. This comprehensive approach ensured
the suitability of this analysis for a wide range of spatial
transcriptomic datasets.
Transcriptomics Profile‐Based Spatial Graph Construction
A transcriptomics profile‐based spatial graph was developed to
characterize the relationships between spots. In previous studies, most
methodologies relied either on spatial information alone or on prior
clustering outcomes to capture spot‐to‐spot relationships. However,
this kind of graph constructed only through spatial location distance
and pre‐trained information cannot effectively describe the
relationship between spots, especially when they were spatially distant
but belong to the same domain. To comprehensively characterize the
relationships between spots, an undirected graph was designed that
integrates both gene expression information and spatial information.
Specifically, the Euclidean distance between each spot and other spots
was calculated through their spatial location information and gene
expression profiles, respectively. For spatial location information,
the top k [1] nearest spots were selected to each spot as its
neighbors, while opting for the top k [2] nearest spots for gene
expression information. Then, connections (edges) between spots that
share neighborhood status were established, yielding two adjacency
matrices A ^spatial and A ^gene respectively, which is defined as
follows:
[MATH:
Aij<
mi>spatial=1,j∈N
k1(i)0,j∉N
k1(i)Aijgene=1,j∈N
k2(i)0,j∉N
k2(i) :MATH]
(1)
where
[MATH:
Nk1(i) :MATH]
and
[MATH:
Nk2(i) :MATH]
represent the neighborhood of the spot i for spatial information and
gene expression information, respectively. Afterwards, the parameter α
∈ (0, 1) was introduced to adjust the weight between spots within A
^gene . Its primary purpose is to establish a relationship framework
where spatial location information takes precedence, is supplemented,
and refined by gene expression information. Finally, A ^spatial and A
^gene were integrated to yield a neighborhood graph A as follows:
[MATH: A=minAspat
ial+αAgene
,1 :MATH]
(2)
Among them, the reason for limiting the weights between spots to 1 is
to prevent over‐correction of gene expression information.
Graph Convolutional Autoencoder
To better integrate and capture gene expression information and spatial
location information, a graph convolutional autoencoder was employed to
learn the latent embedding representation for spatial transcriptomics
data. Specifically, both the encoder and the decoder utilize a
single‐layer graph convolutional network (GCN), taking the normalized
gene expression matrix X and the adjacency matrix A constructed from
spatial information as input. Therefore, for spatial transcriptomics
data, the latent embedding representation Z generated by the encoder
and the reconstructed gene expression matrix H generated by the decoder
can be defined as:
[MATH: Z=σ(D−12AD−12XWe) :MATH]
(3)
[MATH: H=σ(D−12AD−12ZWd) :MATH]
(4)
where σ(·) denotes the nonlinear activation function; D = diag{(I + A)1
[n ]} (n is the total number of spots) represents the degree matrix; W
[e ]and W [d ]are the parameter matrices in the encoder and decoder,
respectively. Further, an inner product decoder was designed to compute
the reconstructed adjacency matrix
[MATH: A^ :MATH]
, which is defined as follows:
[MATH: A^=sigmoid(ZTZ) :MATH]
(5)
where sigmoid(x) = 1/(1 + e ^−x ) represents the nonlinear activation
function. To jointly extract critical gene expression information and
spatial location information, the reconstruction loss of the gene
expression matrix was minimized and the adjacency matrix as follows:
[MATH: Lrec<
/msub>=1n∑i=1nα1||Hi−Xi||2+α2||A^i−<
msub>Ai||2
:MATH]
(6)
where i denotes the i‐th spot, α[1] and α[2] are weight coefficients.
Space‐Aware Contrastive Learning
A space‐aware contrastive learning framework that leverages spatial
information was proposed to ensure alignment between the spots in the
embedding representations and their true spatial distribution. A
space‐aware contrastive learning framework was proposed to improve the
quality of the latent embedding representation through spatial
information. First, data augmentation was performed to generate a
corrupted gene expression matrix by adding Gaussian noise, which
defined as:
[MATH: Z^=X⊙Ng :MATH]
(7)
where N [g ]represents the random noise matrices obeying the Gaussian
distribution, ⊙ is the Hadamard product.
[MATH: X^ :MATH]
was inserted into the encoder that shares parameters with X to generate
the latent embedding representation
[MATH: Z^ :MATH]
under a different perspective as follows:
[MATH: Z^=σ(D−12AD−12X^We) :MATH]
(8)
Contrastive learning endeavors to optimize the likeness between
positive pairs while concurrently minimizing the resemblance among
negative pairs. Inspired by this principle, a space‐aware contrastive
learning approach tailored was proposed to spatial transcriptomics
data. The space‐aware contrastive loss is established on a reasonable
assumption that adjacent spots are similar and non‐adjacent spots are
dissimilar. Therefore, the key idea is to enhance the similarity of
spatially adjacent spots and weaken the similarity of non‐adjacent
spots. Specifically, the cosine similarity matrix S of cross‐view
latent embedding representations was calculated, which is defined as
follows:
[MATH:
Sij=(Zi)(Z^j)T||Zi||||Z^j|| :MATH]
(9)
where S [ij ]denotes the cosine similarity between the embedding
representation of the i‐th spot in the first view and the embedding
representation of the j‐th spot in the second view. To align the spot
similarities with the spatial coordinate distribution, the mean square
error (MSE) was employed between S and the adjacency matrix A as the
contrastive loss, formulated as follows:
[MATH: Lcon<
/msub>=1n2<
/msup>∑i=1n∑j=1nSij−
Aij
2 :MATH]
(10)
It is important to emphasize that during model training, as we
minimized the contrastive loss, the similarity of spot pairs
characterized by high weight in matrix A is increased while
simultaneously observing a reduction in the similarity of pairs with
lower weights. Obviously, this process clearly results in our
space‐aware contrastive loss effectively bringing spatially adjacent
spots into closer proximity and pushing non‐adjacent spots further
apartx, ensuring that the embedding representation can effectively
capture the spatial similarity inherent in the underlying
organizational structure.
Cluster‐Level Feature Contrastive Learning
Apart from adjusting the correlation of spots across views through
spatial information, the correlation of features in the cluster‐level
embedding representation was further optimized. The embedding
representations of different views were equally divided into k groups
(k is the number of clusters), calculated the average embedding within
each group, and finally combined them to obtain the cluster‐level
embedding representations. Specifically, the readout function was
introduced
[MATH: R(·):Rn×d→Rk×d :MATH]
(d is the latent embedding dimension) formulated as:
[MATH: R(Z)=Z¯1,<
msub>Z¯2,<
mtext>…,Z¯k<
/mrow> :MATH]
(11)
[MATH: Z¯i=1|ni|∑j∈niZj,i∈1,2,…,
k :MATH]
(12)
where
[MATH: Z¯i
:MATH]
denotes the average embedding of group i; n [i ]represents the set of
spots in group i; |n [i ]| is the number of spots in group i. Likewise,
still calculated the cosine similarity matrix
[MATH: S∼∈Rd×d :MATH]
of cross‐view cluster‐level embedding representations from the aspect
of feature dimension, which is defined as follows:
[MATH: S∼ij=(R(Z)i)T(R(Z^)j)||R(Z)i||||R(Z^)j|| :MATH]
(13)
where
[MATH: S∼ij :MATH]
denotes the cosine similarity between the average embedding of the i‐th
dimension feature in all groups in the first view and the average
embedding of the j‐th dimension feature in all groups in the second
view. Representations of the same dimension feature across views were
taken as positive pairs and maximize their similarity. Therefore, the
MSE of
[MATH: S∼ :MATH]
and an identity matrix I ∈ R ^d × d were used as the cluster‐level
contrastive loss, which is defined as follows:
[MATH: Lclu<
/msub>=1d2<
/msup>∑i=1d∑j=1dS∼ij−Iij<
/msub>2=1d2<
/msup>∑i=1dS∼ii−12+1d
2−d∑<
mrow>i=1d∑j≠iS∼ij2 :MATH]
(14)
The cluster‐level feature contrastive learning strengthens the
correlation between the same dimension features in two views and
reduces the correlation between different dimension features. This can
filter out the redundant information of latent embedding
representation, preserve more discriminative features well, and improve
the clustering performance of the model.
Overall Loss Function
stDCL learns the latent embedding representation by minimizing the
reconstruction loss
[MATH: Lrec<
/msub> :MATH]
, the space‐aware contrastive loss
[MATH: Lcon<
/msub> :MATH]
, and the cluster‐level feature contrastive loss
[MATH: Lclu<
/msub> :MATH]
, which can be defined as:
[MATH: L=γ1Lrec<
/msub>+γ2Lcon<
/msub>+γ3Lclu<
/msub> :MATH]
(15)
where γ[1], γ[2] and γ[3] are weight coefficients assigned to each loss
to control the balance of the overall loss function. γ[1], γ[2] and
γ[3]was set to [10, 0.5, 0.8] after parametric analysis. For model
training, it was divided it into two parts: pre‐training and training,
and both are optimized using the Adam optimizer. The learning rate and
epoch for these two parts were 0.001, 0.005 and 500,
1000, respectively.
Clustering and Refinement
To identify the spatial domains in our study, clustering techniques
were employed on the latent embedding representations generated by
stDCL. First, the dimension of the learned embedding representations of
stDCL was reduced to 20 by PCA, and then used clustering methods to
perform clustering. When the data is manually annotated, the mclust
clustering algorithm was employed. For cases without manual annotation,
the Leiden algorithm was used. To enhance the clustering performance,
data augmentation were employed with two different Gaussian
distributions and selected clustering results with relatively low
Davies–Bouldin scores. Furthermore, stDCL incorporates a refinement
process for the clustering results. Specifically, the neighborhood of
each spot was defined as a circle with a radius of r (r is set to 50)
and reassigned the spots to domains with the most common labels within
their respective neighborhoods.
Implementation Details and Comparisons with Baseline Methods
stDCL was implemented in Python, and its core model is built on the
PyTorch framework, which is publicly available at
[215]https://github.com/Philyzh8/stDCL. To evaluate the performance of
stDCL in spatial clustering, stDCL was compared with the following
methods:
* Seurat. Seurat is a comprehensive software package written in R
that facilitates the analysis of single cell data, including
modules dedicated to processing spatial transcriptomics data. For
data preprocessing, the default parameters in Seurat were utilized
. To perform clustering, the appropriate resolution based on the
desired number of clusters was determined and the “FindClusters”
function was employed to assign cells to their respective clusters.
* SpaGCN. SpaGCN is a method based on graph convolutional networks
that integrates three types of information including gene
expression, spatial location, and histological images to identify
spatial domains. The adjacency matrix in SpaGCN is constructed with
parameters alpha' and beta' set to 1 and 49, respectively. During
the training process, we set the learning rate to 0.05 and trained
the model for 200 epochs.
* stLearn. stLearn uses the deep learning model to extract
information from tissue morphology to normalize gene expression and
improve spatial clustering. In this study, stLearn was employed for
data normalization and spatial clustering, using the default
parameter settings. The clustering method used in stLearn is the
K‐means method.
* BayesSpace. BayesSpace is a Bayesian statistical clustering method
that leverages information about spatial neighborhoods to enhance
the resolution of spatial transcriptomics data and perform
clustering analysis. The parameters were configured as follows: “d”
was set to 15, “nrep” to 50 000, and “gamma” to 3.
* SEDR. SEDR (Spatial Embedding with Deep Representation) combines
both an autoencoder and a variational autoencoder to generate
embedding representations of transcripts and spatial information.
During training, the number of epochs were set to 200, the learning
rate to 0.01, and the dropout rate to 0.2.
* CCST. CCST is an unsupervised cell clustering method that uses GCNs
to learn cell embedding representations from graphs derived from
spatial transcriptomics data. For parameter settings, the default
parameter settings were used in their code. The training process
consisted of 5000 epochs, and the number of hidden layer nodes was
set to 256. Additionally, the parameter “lambda_I” was set to 0.3.
* STAGATE. STAGATE is a graph attention autoencoder framework
designed to learn a low‐dimensional latent embedding by combining
spatial information and gene expression profiles. For data
preprocessing, 3000 highly variable genes were specifically
selected as input for analysis. The parameters “rad_cutoff” and
“alpha” were set to 150 and 0, respectively.
* GraphST. GraphST is a graph self‐supervised comparative learning
method that effectively utilizes spatial information and gene
expression profiles for spatial clustering. The parameters set for
model training were adopted from the recommended parameters in
their code. For loss function weight assignment, the parameters
“alpha” and “beta” were set to 10 and 1, respectively. The training
process consisted of 600 epochs, with a learning rate of 0.001 and
a dropout rate of 0.2.
Benchmarking Metrics for stDCL Clustering
Three common external metrics were adopted to evaluate clustering
performance, namely adjusted rand index (ARI), normalized mutual
information (NMI), and homogeneity score (HS). ARI can measure the
similarity between the cluster labels predicted by the algorithm and
the real cluster labels. NMI quantifies the similarity of between two
clusters by considering the mutual information between the predicted
and true cluster labels, and normalizing it based on the entropy of the
individual clusters. These two metrics can be defined as follows:
[MATH: ARI=∑ijnij2−∑iai2
∑jbj2
/n2
12∑iai2
+∑jbj2
−∑iai2
∑jbj2
/n2 :MATH]
(16)
[MATH: NMI=∑ijnijlogni
jnaibj∑iailogain∑jbjlogbjn :MATH]
(17)
where n is the total number of samples, n [ij ]is the number of samples
belonging to both the i‐th cluster in the true label and the j‐th
cluster in the predicted label, a [i ]and b [j ]are the total number of
samples in the i‐th cluster in the true label and the j‐th cluster in
the predicted label, respectively.
HS can measure the homogeneity of the clustering algorithm for sample
division, which means that only samples of the same category are
included in the cluster. The HS ranges from 0 to 1, where a higher
value indicates a higher level of homogeneity in the clusters, which
can be defined as:
[MATH: h=1−H(C|K)H(C) :MATH]
(18)
where H(C|K) is the conditional entropy of the true label partition
under the predicted label partition condition.
In addition, the internal metric Davies‐Bouldin Index (DBI) was also
used to improve the clustering performance of stDCL. DBI can calculate
the average of the maximum ratio of the within‐cluster distance and the
between‐cluster distance for each cluster, which can be formulated as:
[MATH: DBI=1n∑i=1nmaxj≠iSi¯+Sj¯<
/mover>Eij<
/mfrac> :MATH]
(19)
where n is the total number of clusters,
[MATH:
Si¯
:MATH]
denotes the average distance from the sample in the i‐th cluster to the
cluster centroid, and E [ij ]represents the Euclidean distance between
the i‐th cluster and the j‐th cluster. The lower the DBI, the better
the clustering quality.
Benchmarking Metrics for Stdcl Imputation
Four common metrics were adopted to evaluate the imputation performance
of stDCL, namely Pearson correlation coefficient (PCC), Mean Squared
Error (MSE), Spearman Rank Correlation Coefficient (SRCC), and Cosine
Similarity (CS). These metrics evaluate the correlation and consistency
of stDCL's imputation space with the original space. PCC was used to
measure the linear correlation between the imputed and the true values.
It captures the strength and direction of a linear relationship, with a
value ranging from –1 to 1, where 1 indicates a perfect positive linear
correlation and –1 indicates a perfect negative linear correlation,
which can be defined as:
[MATH: PCC=∑i<
/mi>=1n(xi−x¯)(yi−y¯)∑<
/mo>i=1n(xi−x¯)2∑i=1n
(yi−y¯)2 :MATH]
(20)
where x [i ]and y [i ]represent the true and imputed values,
respectively, and
[MATH: x¯ :MATH]
and
[MATH: y¯ :MATH]
denote the mean of the true and imputed values.
MSE is used to measure the average squared difference between the true
and imputed values. MSE penalizes larger deviations more heavily, which
makes it sensitive to outliers. The lower the MSE, the better the
imputation performance, which can be defined as:
[MATH: MSE=1n∑i=1n(xi−yi)2 :MATH]
(21)
where n is the number of spots, and x [i ]and y [i ]are the true and
imputed values, respectively.
SRCC measures the rank‐based correlation between the true and imputed
values. Unlike PCC, SRCC is non‐parametric and assesses how well the
relationship between the two variables can be described by a monotonic
function, which can be formulated as:
[MATH: SRCC=∑i<
/mi>=1n(rxi−rx¯)(ryi−ry¯)∑<
/mo>i=1n(rxi−rx¯)2∑i=1n
(ryi−ry¯)2 :MATH]
(22)
where
[MATH:
rxi
:MATH]
and
[MATH:
ryi
:MATH]
are the rank values of x [i ]and y [i ], respectively.
CS quantifies the similarity between two non‐zero vectors by measuring
the cosine of the angle between them. It is particularly useful when
comparing high‐dimensional data and ranges from 0 to 1, with 1
indicating perfect similarity, which can be defined as:
[MATH: CS=∑i<
/mi>=1nx
iyi∑i=1nxi2∑i=1nyi2 :MATH]
(23)
where x [i ]and y [i ]are the true and imputed values.
Differential Gene Expression Analysis
We performed differential gene expression (DE) analysis using the
function “FindMarkers” in the R package Seurat (v4.0.2).^[ [216]^29 ^]
The Wilcoxon rank sum test was applied to identify differentially
expressed genes, and a Benjamini‐Hochberg correction was used to adjust
the p‐values. Specifically, we selected the genes with adjusted p‐value
⩽ 0.05 as differentially expressed genes in the tested regions.
Functional Enrichment
Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG)
enrichment analyses was performed on the differentially expressed genes
using the R package ClusterProfiler (v3.18.1).^[ [217]^67 ^] The
function “enrichGO” and “enrichKEGG” were applied to GO analysis and
KEGG analysis. In both analyses, the parameters were set as follows:
“pAdjustMethod = BH” and “pvalueCutoff = 0.05”. Subsequently, the
results of the GO analysis were visually presented using the “heatplot”
function from the R package enrichplot (v1.14.2). These analyses
provided valuable insights into the enriched biological processes and
pathways associated with the differentially expressed genes.
PPI Network
STRING^[ [218]^46 ^] was utilized to generate a protein–protein
interaction (PPI) network for the up‐regulated differentially expressed
genes in the brain regions of mouse embryos. The PPI network was
constructed using interactions with scores greater than 0.4. To
visualize the network, Cytoscape was employed,^[ [219]^68 ^] which
allowed us to gain a visual representation of the network and explore
its underlying structure. Furthermore, MCODE^[ [220]^47 ^]was adopted
to identify significant coherent modules within the PPI network and
showed the highest scoring modules.
Trajectory Inference
The R package Slingshot (v2.2.1)^[ [221]^69 ^] was used for trajectory
inference. Slingshot is a powerful tool designed to infer the structure
and order of cell lineage differentiation, which consists of two main
stages: inference of lineage structure, and inference of pseudotemporal
changes in cells of each lineage. In this analysis, Slingshot was
utilized on the latent representations obtained from stDCL. The
predicted cluster labels were used as input for the “slingshot”
function in Slingshot. This allowed us to leverage the learned latent
representations to infer the trajectory and pseudotime information for
the cells. Following trajectory inference, Slingshot assigned a
pseudo‐time value to each location. In addition, Slingshot needed to
set a domain as the starting domain in trajectory inference, and the
White matter was set as the starting domain when analyzing the osmFISH
dataset. This choice determined the initial reference point from which
the trajectory was inferred, aiding in the interpretation of the
differentiation process in the context of the specific dataset.
Moreover, R package Monocle3 (v1.3.6)
([222]http://cole‐trapnell‐lab.github.io/monocle‐release/monocle3) was
used to infer pseudotime trajectories of astrocyte subtypes in the
STARmap PLUS data. Monocle3 used the “learn_graph” function to learn
principal graph in the UMAP space. Pseudotime was computed by
calculating the geodesic distance from the specified root node to the
cell node along the principal graph, and adopted the function
“plot_cells” to visualize the trajectory inference.
Gene Set Variation Analysis (GSVA)
Gene set variation analysis were performed using the R package GSVA
(v1.38.2). In the analysis, the “gsva” function from the GSVA package
was utilized. The GSVA enrichment scores were estimated using the
“method = ssgsea” parameter setting. This method, based on the Single
Sample Gene Set Enrichment Analysis (ssGSEA) approach, allowed us to
evaluate the activity or enrichment levels of gene sets or pathways in
each sample. The related pathway enrichment was used to assess the
activation of WikiPathways pathways, which is available in the MSigDB
database.^[ [223]^70 ^]
Conflict of Interest
The authors declare no conflict of interest.
Supporting information
Supporting Information
[224]ADVS-12-2410081-s001.pdf^ (1.5MB, pdf)
Acknowledgements