Skip to main content
NCBI home page
Briefings in Bioinformatics logoLink to Briefings in Bioinformatics
. 2024 Apr 16;25(3):bbae173. doi: 10.1093/bib/bbae173

Attention-guided variational graph autoencoders reveal heterogeneity in spatial transcriptomics

Lixin Lei 1, Kaitai Han 2, Zijun Wang 3, Chaojing Shi 4, Zhenghui Wang 5, Ruoyan Dai 6, Zhiwei Zhang 7, Mengqiu Wang 8, Qianjin Guo 9,✉
PMCID: PMC11021349  PMID: 38627939

Abstract

The latest breakthroughs in spatially resolved transcriptomics technology offer comprehensive opportunities to delve into gene expression patterns within the tissue microenvironment. However, the precise identification of spatial domains within tissues remains challenging. In this study, we introduce AttentionVGAE (AVGN), which integrates slice images, spatial information and raw gene expression while calibrating low-quality gene expression. By combining the variational graph autoencoder with multi-head attention blocks (MHA blocks), AVGN captures spatial relationships in tissue gene expression, adaptively focusing on key features and alleviating the need for prior knowledge of cluster numbers, thereby achieving superior clustering performance. Particularly, AVGN attempts to balance the model’s attention focus on local and global structures by utilizing MHA blocks, an aspect that current graph neural networks have not extensively addressed. Benchmark testing demonstrates its significant efficacy in elucidating tissue anatomy and interpreting tumor heterogeneity, indicating its potential in advancing spatial transcriptomics research and understanding complex biological phenomena.

Keywords: spatially resolved transcriptomics, spatial clustering, graph deep learning, variational graph autoencoder, attention-guided

INTRODUCTION

In spatially resolved transcriptomics (SRT) studies [1–4], identifying spatial domains [5–8] is a critical task. Currently, techniques for delineating spatial domains can be classified into two primary categories: non-spatial approaches [9–13] and spatial clustering methods [8, 12, 14–16]. Traditional non-spatial methods primarily utilize clustering algorithms [17] such as K-means [18] and Louvain [19] However, their progress is limited due to technological fluctuations, resulting in either too few or overly sparse spots [20, 21]. Another approach involves defining cell type features through convolutional single-cell RNA-seq [22], but experiments demonstrate their inadequacy for achieving cell- or subcellular-level resolution in ST data. These approaches utilize gene expression data as input. However, since the cells must be scrambled to prepare the sample in the experiment, the location information of the cells is lost, and as a result, the clustering results are often not closely related to the actual structure of the tissue slice.

Therefore, considering the relationship between gene expression [23], spatial location and tissue morphology [24] becomes particularly important to better connect gene expression and spatial positional information. Some relevant studies have made attempts in this direction. For example, BayesSpace [8] adopts Bayesian methods and introduces spatial priors to encourage nearby points to belong to the same cluster. stLearn [16] utilizes morphological distance and spatially neighboring structures to standardize gene expression data. SpaGCN [25] integrates gene expression, spatial location and morphological data to discern spatial domains through the creation of an undirected weighted graph. SEDR [26] employs a deep autoencoder network to embed spatial information. DeepST [27] enhances data representation by using graph neural networks (GNNs) to construct a denoising autoencoder. The spatial-ID [28] method leverages transfer learning techniques to construct probability distributions from learned single-cell RNA sequencing (scRNA) data, forming pseudo-labels. These are then combined with a variational autoencoder for cell identity characterization on a reference dataset with clearly defined cell types. The PAST [29] method integrates self-attention mechanisms [30] with residual networks and graph convolutional networks to extract latent features from SRT data. The STAGATE [31] method incorporates graph attention mechanisms into encoder and decoder operations, interpreting spatial domains in spatial transcriptomics and showcasing functionality in three-dimensional visualization. STAGATE demonstrates the potential application capabilities of attention mechanisms in understanding the similarity between adaptively adjacent points and representing spatial domain boundaries.

Among the methods discussed, the application of GNN methods [32–39], particularly utilizing the variational graph autoencoder (VGAE), demonstrates notably superior ARI scores in benchmark testing. In methods like PAST and STAGATE, additional attention layers are incorporated to further enhance the clustering performance of the model. However, limitations persist regarding the balance between attention to local and global structures. Additionally, research on tasks involving clustering without prior knowledge of the number of clusters remains relatively limited, and the performance of traditional clustering algorithms is often impeded as they typically rely on user-provided prior information to determine the number of clusters. Lastly, effectively leveraging slice data, spatial information and raw gene expression from SRT data to construct enhanced gene expression matrices for calibrating low-quality gene expression remains a challenge.

Here, we introduce AttentionVGAE (AVGN), which leads the VGAE [39] by incorporating modules containing multi-head attention (MHA blocks) [40–42] to identify spatial domains. Specifically, AttentionVGAE utilizes pre-trained MaxVit networks to extract latent features from slice images and combines them with raw expression data and positional information of spatially adjacent points to construct an enhanced gene expression matrix. Subsequently, latent representations are obtained from the enhanced gene expression matrix using a graph convolutional encoder with MHA blocks. Finally, the reparameterization and graph deconvolution decoder are employed to reconstruct unstructured information. Notably, the introduction of MHA blocks in AttentionVGAE enables the network to concentrate attention on specific regions or features while also focusing on multiple key features in spatial data, thereby effectively capturing the complex relationships in gene expression within tissues. This facilitates learning more general and abstract feature representations, enhancing the network’s generalization capability. Additionally, AttentionVGAE attempts to adaptively adjust its focus on potential clustering structures in the data with the assistance of MHA blocks, thereby improving the flexibility and efficacy of clustering algorithms. Consequently, AttentionVGAE not only calibrates low-quality gene expression by integrating slice images, spatial dependencies and gene expression but also enhances the balance between attention to local and global structures, thus improving the performance of identifying spatial domains.

Through benchmark testing against six state-of-the-art methods on SRT datasets from 10x Visium, AttentionVGAE demonstrates higher clustering accuracy. It elucidates fine-grained anatomical regions in the adult mouse brain dataset from 10x Visium and the Slide-seq mouse hippocampus dataset. Moreover, AttentionVGAE is employed to decipher tumor heterogeneity in adult Human Glioblastoma Multiforme and infiltrative ductal carcinoma datasets [43, 44], providing deeper biological insights into cancer-related genes and serving as a valuable reference for subsequent analysis combining other clinical data [45–47].

MATERIALS AND METHODS

Dataset

In this manuscript, to comprehensively assess the performance of AttentionVGAE compared to state-of-the-art spatial detection methods, we utilized a highly regarded dataset of the human dorsolateral prefrontal cortex (DLPFC) [43] as a benchmark. This dataset was obtained from the 10x Visium platform and includes data from three subjects, totaling 12 sections. The anatomical coverage of each section spans all six layers of neuronal cells and the cortex, resulting in a total of seven layers.

Additionally, diverse datasets were employed, including coronal slices of the adult mouse brain (fresh-frozen). There are 2702 spots under tissue with an average of 115 569 reads per spot and a median of 6018 genes per spot.

Furthermore, data from adult human glioblastoma cells were used, comprising 3468 spots under tissue with an average of 16 383 reads per spot. Notably, 87.4% of reads confidently mapped to the targeted transcriptome. Another dataset involved invasive ductal carcinoma, with 4727 spots under tissue, averaging 40 795 reads per spot and a median of 2964 genes per spot. All these datasets originated from the 10x Visium platform.

Samples from the Stereo-seq platform were obtained, including the first four sagittal slices from the early developmental stages of C57BL/6 mouse embryos, selected from the MOSTA database [44]. These slices correspond to E9.5 (~7.1 mm2), E10.5 (~11.5 mm2), E11.5 (~18.8 mm2) and E12.5 (~32.1 mm2), chosen to capture crucial developmental milestones. Also from the MOSTA database are the Dorsal_midbrain_celand data and Adult mouse hemi-brain data, providing molecular insights into the gene expression dynamics during early mouse embryonic development.

Finally, samples from the Slide-seq platform [45], including the Hippocampus dataset, were included. This dataset contributes to a further evaluation of AttentionVGAE’s effectiveness in handling high-spatial-resolution data, demonstrating its adaptability across different platforms.

Preprocessing

To fully leverage image information for spatial domain detection tasks, AVGN employs a pre-trained MaxVit [46] model. Initially, AVGN performs image segmentation on each data point in the spatial transcriptomics data, dividing the images into small blocks. The size of the segmented image blocks is set to 224 pixels × 224 pixels, encompassing three channels (RGB). These small image blocks serve as inputs to the MaxVit model, which has an input layer dimension of (1, 3, 224, 224), where 1 represents batch size, 3 represents the number of channels and 224x224 denotes the image dimensions. The MaxVit model extracts features from each image block, and the resulting feature matrix, MaxVit(input), has dimensions equal to the predefined number of PCA [47] principal components, typically set to 50.

To obtain an augmented gene expression matrix, the enhanced expression (Inline graphic) for the gene expression (Inline graphic) at any point a in the spatial domain is calculated using the following formula:

graphic file with name DmEquation1.gif (1)

The b represents the set of n neighbors associated with point a (in AVGN, the default is to calculate the radius a based on the mean and variance of the distances to the six nearest neighboring points, considering all points within a distance less than r as neighbors of point a). Inline graphic is the cosine similarity between the gene expression profiles of points a and b, Inline graphic indicates whether the two points are neighbors (equals 1 if they are neighbors, 0 otherwise) and Inline graphic represents the gene expression weight between the two points, calculated as follows:

graphic file with name DmEquation2.gif (2)

AVGN constructs a KDTree [48] using enhanced gene expression data and point coordinate information. KDTree is a tree-like data structure used for efficient searching of nearest neighbors in multi-dimensional space. Utilizing the constructed KDTree, a nearest neighbor search is performed for each point in the dataset, returning the distances and indices of the nearest k + 1 neighbors for each point. Subsequently, by iterating over each data point, edges are added to the graph structure connecting it with its k nearest neighbors, and these edges are stored in a graphList. The inner loop iterates over each row corresponding to a data point in the dataset, adding the index of the current point and its k nearest neighbors to the graphList. This process establishes a graph in which each data point is connected to its k nearest neighbors.

Finally, the data are normalized [49] to ensure that the sum of gene expression values for each cell is equal to 1. Logarithmic transformation and standardization are applied to scale down the magnitude of the data and ensure that the expression values for each gene have similar scales across all cells. Principal Component Analysis (PCA) is employed to reduce the dimensionality of the data to a specified number of principal components (default set to 128 dimensions in AVGN). This series of processing steps helps eliminate differences between cells, providing more comparable data for subsequent analysis.

Encoder components

For each element Inline graphic, the normalization operation involves the following two steps:

Calculate the meanInline graphic and standard deviationInline graphic for each feature:

graphic file with name DmEquation3.gif (3)
graphic file with name DmEquation4.gif (4)

where Inline graphicrepresents the sample size.

Standardize each element Inline graphic to obtain the standardized element Inline graphic:

graphic file with name DmEquation5.gif (5)

This normalization process standardizes the distribution of each data feature, giving it a mean of 0 and a variance of 1. This standardized format enhances the ease of subsequent analysis and modeling.

Multi-head attention mechanism components

This powerful technique enhances a neural network’s ability to process data by employing parallel computations and multiple heads. This improves the model’s representational capacity and its ability to capture diverse relationships. The original input is a vector of dimensions Inline graphic. This input is linearly transformed by a linear layer, mapping it to Inline graphic different subspaces, each with a dimension Inline graphic . This operation can be seen as segmenting the input features. Independent attention calculations are then performed for each subspace in parallel, improving processing efficiency. Subsequently, the information from different subspaces is fused together. Finally, the integrated attention vectors undergo another linear layer to map them back to the original Inline graphic dimension. In AVGN, we introduced the MAH block, which is composed of a linear layer stacked on top of a multi-head attention layer. The objective is to allow the linear layer to learn the latent representation of input features. This is achieved by parallel computation through the multi-head attention layer, enhancing the model’s information capture capabilities. The input dimension of the linear layer is 128, with an output dimension of 32, while the input and output dimensions of the multi-head attention layer are both set to 32.

graphic file with name DmEquation6.gif
graphic file with name DmEquation7.gif (6)

Each attention head can learn distinct relationships, empowering the model to more effectively capture various facets of information within the input sequence. Consequently, this enhances the model’s generalization and representational capabilities.

VGAE components

The main difference between VGAE and VAE lies in the use of a multidimensional Gaussian distribution to replace the graph embedding in VAE. This avoids being confined to a single point in the latent space. If the input is Inline graphic and the parameters of the multivariate normal distribution are Inline graphic, then in the encoder, the probability of Inline graphic given Inline graphic is denoted as Inline graphic, and in the decoder, the probability of Inline graphic given Inline graphic is denoted as Inline graphic. In the encoder, the input Inline graphic is known, and the variable is Inline graphic, while in the decoder, the parameters of the multivariate normal distribution Inline graphic are known. VGAE aims to make Inline graphicas close as possible to Inline graphic. Inline graphic. In the VGAE framework, the encoder is responsible for determining the probability distribution of the latent variable Inline graphic given the input Inline graphic. The goal is to capture the latent representation that best represents the input data. The subsequent steps in VGAE involve the decoder and the optimization process to align the distributions Inline graphic and Inline graphic for effective graph reconstruction and generation. To achieve this, the process begins with the encoder: Please note that there are two encoders here. GCN1 is used to generate a low-dimensional embedding vector Inline graphic, GCN2 is used to generate Inline graphic and Inline graphic，where Inline graphic and Inline graphic(Figure 1D, b).

Figure 1.

Figure 1

Open in a new tab

The workflow of AttentionVGAE. (A) Data required for the process includes hematoxylin and eosin (H&E)-stained histological images (optional), spatial gene expression, and spatial coordinates. (B) AttentionVGAE computes the adjacency matrix and utilizes a pre-trained MaxVit [46] to learn features present in the images, followed by appropriate normalization [47]. (C) The preprocessed feature matrix and adjacency matrix are input into the optimized VGAE to extract latent embeddings. (D) Based on the latent embeddings, a cell adjacency network is constructed for further clustering operations, and the results of spatial domain detection are visualized [50]. (a) Details of the MaxVit block. (b) Data flow display in encoder. (c) The calculation process in reparameterization. (d) Figure deconvolution decoder working details.

Therefore, the GCN layer is represented as:

graphic file with name DmEquation8.gif (7)

Next is the use of reparameterization (Figure 1D, c):

graphic file with name DmEquation9.gif
graphic file with name DmEquation10.gif (8)

Finally, the decoder part (Figure 1D, d):

graphic file with name DmEquation11.gif (9)

In AVGN, the preprocessed input features X undergo encoding through the encoder (which includes the MHA block), resulting in a latent representation of dimension 20. This representation is then concatenated with the re-parameterization output, yielding an embedding Z of dimension 28. Finally, the decoder is employed to reconstruct features with a dimension of 128.

Performance evaluation

In this manuscript, we employ the Adjusted Rand Index (ARI) as a key evaluation metric to measure the similarity between two partitions. It calculates the number of data points assigned the same or different labels in the two partitions, providing a quantitative assessment of their resemblance.

graphic file with name DmEquation12.gif (10)

Differing from Normalized Mutual Information (NMI), ARI takes into account the influence of random assignments, ensuring a more precise outcome. ARI values span from −1 to 1. Positive values signify superior clustering performance compared to random assignment, while negative values indicate subpar clustering results in contrast to random assignment. A value of 0 suggests similarity to random assignment.

NMI quantifies the degree of information shared between the clustering result and the reference partition. NMI values generally fall within the range of 0 to 1, where increased values signify greater similarity between the clustering outcome and the reference partition. NMI is less sensitive to the size of partitions and the number of categories, making it robust when dealing with different partition sizes and category numbers.

graphic file with name DmEquation13.gif (11)

RESULTS

In order to comprehensively illustrate the performance, versatility and potential applications of AttentionVGAE across various datasets, we designed and conducted experimental work.

Overview of AttentionVGAE workflow

Firstly, AttentionVGAE integrates and reduces the dimensionality of histology, gene expression and spatial location information (Figure 1A and B), utilizing a pre-trained MaxVit model for learning image features (Figure 1D, a). Subsequently, this integrated information is input into a VGAE guided by a multi-head attention mechanism (Figure 1C). The latent embeddings obtained from the encoder are employed to construct a cell adjacency network, which undergoes Leiden clustering. Detailed explanations for this section can be found in the VGAE component of the Materials and Methods section in this manuscript. Following that, refinement is applied to enhance the clustering results. Finally, the clustered results are visualized in the spatial domain by incorporating spatial location information (Figure 1D).

AttentionVGAE underwent benchmark testing on the DLPFC dataset

To evaluate the performance difference of AttentionVGAE relative to state-of-the-art techniques in spatial domain detection, we conducted a comparative analysis involving AttentionVGAE and seven other detection methods: Scanpy, stLearn, SEDR, SpaGCN, BayesSpace, DeepST and STAGATE. It’s worth noting that Leiden, implemented in Scanpy, is categorized as a non-spatial clustering method, while the others belong to spatial clustering methods.

In the Uniform Manifold Approximation and Projection (UMAP) visualizations, noteworthy observations come to light. The clustering results of stLearn reveal a dissonant and indistinct pattern in Figure 2D, indicating a lack of clearly defined hierarchical contours within the spatial domain visualization (Figure 2C). SEDR, SpaGCN and BayesSpace exhibit relatively clear hierarchical boundaries, but there are more outliers near the edges of the layers, leading to a sense of discontinuity in the layers (Figure 2C). DeepST shows significant deficiencies in defining the boundaries of the overall smaller cortical layer [53] and larger cortical layer. STAGATE improves on this issue, particularly evident in the visualizations of Layer_0 to Layer_4 (Figure 2C). In the visualizations of AttentionVGAE’s detection results, the clustering performance of Layer_2, Layer_3 and Layer_5 is notably superior to other spatial clustering methods. The edge of the clustering visualization results is more accurate, clear and smooth, demonstrating excellent clustering performance (Figure 2C). In the UMAP visualizations of the 151 673 samples, data processed by AttentionVGAE exhibit outstanding results in the scatter plot. The clustering points of the seven cortical layers are more clearly concentrated. Compared to the results of other detection methods, AttentionVGAE’s UMAP results have fewer outliers. This indicates that AttentionVGAE has a stronger ability to perceive structural heterogeneity [54]. Lastly, a temporal analysis visualization of the six key detection methods tested in this manuscript for the 151 673 samples reveals their differentiation trajectories, which is crucial for studying cell differentiation trajectories (from dark colors to light colors) (Figure 2E). Notably, AttentionVGAE’s average ARI is significantly higher than other spatial domain detection methods, with an average ARI of 0.592 (Figure 2B).

Figure 2.

Figure 2

Open in a new tab

The benchmark test results of AttentionVGAE compared to state-of-the-art spatial domain detection methods on the DLPFC dataset. (A) presents the ground truth benchmark of 151 673 samples manually annotated in the DLPFC dataset. (B) exhibits boxplots that compare the ARI metrics among different spatial domain detection methods applied to the DLPFC dataset. Each detection method underwent 12 tests. (C) showcases the spatial domain detection results of AttentionVGAE corresponding to ARI = 0.633, along with visualizations of the UMAP [51] and PAGA graphs [52]. (D) exhibits the test results of five selected other state-of-the-art spatial domain detection methods (choosing the highest ARI among the 12 tests) and their corresponding visualizations of UMAP and PAGA graphs.

AttentionVGAE demonstrates exceptional capabilities in detecting fine structures on coronal sections of the adult mouse brain (fresh-frozen)

In the spatial domain visualization of AttentionVGAE’s results in detecting fine structure positions from (Figure 3C), clear and distinct boundaries can be observed, forming an overall pattern that is both elegant and smooth. Upon consulting the average expression levels of highly variable genes and the proportion of cells harboring these genes in Figure 3B, a discernible correlation emerges, showcasing the intricate relationship between the finely identified spatial positions by AttentionVGAE and the expression patterns of differentially expressed genes. This correlation is visually substantiated through the depictions in Figure 3D and E. Based on these results, the incorporation of the multi-head attention mechanism in GVAE directs the focus toward essential aspects of the data, facilitating fine-grained modeling. In clustering tasks, this implies that the model can better distinguish between different categories or clusters and more accurately capture the inherent structure of the data. Therefore, AttentionVGAE demonstrates effective capabilities in spatial domain detection of data with fine structure positions.

Figure 3.

Figure 3

Open in a new tab

The spatial domain detection conducted by AttentionVGAE on coronal sections of the adult mouse brain (fresh-frozen). (A) includes a high-resolution tissue image of the coronal section of the adult mouse brain, the ground truth of spatial domain detection for adult mouse samples and the visualization of AttentionVGAE’s spatial domain detection results on experimental data. (B) displays the average expression levels of genes Ddn, Kcnab2 and Cplx2 in Clusters 0, 8, 10 and 12. These three genes were selected through differential gene analysis as highly variable genes. The proportion of cells containing these genes in the total cell population of the corresponding cell clusters is also shown. (C) provides detailed views of fine structures in the spatial domain detection results. (D) shows the spatial distribution maps of genes Ddn, Kcnab2 and Cplx2 at fine structure locations. (E) includes violin plots of gene expression levels for Ddn, Kcnab2 and Cplx2 in the clusters.

AttentionVGAE accurately performs spatial domain detection on adult human glioblastoma multiforme

From (Figure 4A), it can be observed that AttentionVGAE’s detection results exhibit the best average ARI performance. However, by examining the NMI metric, it is noted that AttentionVGAE’s NMI is slightly lower than the results from stLearn (Figure 4B). Employing both NMI and ARI contributes to a more comprehensive clustering evaluation, leveraging their distinct advantages in handling partitions of varying sizes and different numbers of categories. NMI demonstrates robustness in dealing with partitions of different sizes and category numbers, while ARI considers the impact of random assignment, potentially offering more accurate results in certain situations. The combined use of these two metrics allows for a more comprehensive assessment of clustering algorithm performance, reducing subjective bias. Careful examination of the UMAP plots in Figure 4C reveals that AttentionVGAE’s detection results also exhibit superior performance.

Figure 4.

Figure 4

Open in a new tab

The comparative results of AttentionVGAE and the latest spatial domain detection methods on the GBM dataset. (A) includes a high-resolution tissue image of GBM, the ground truth of spatial domain detection for GBM and visualizations of spatial domain detection results for six methods. (B) presents bar charts depicting the ARI and NMI metrics for the spatial domain detection results of the six methods. (C) showcases UMAP plots and PAGA analysis graphs for the detection results of the six spatial domain detection methods. (D) displays spatial visualizations of four differentially expressed genes.

Analyzing the spatial domain detection results of AttentionVGAE concerning highly variable differentially expressed genes within the tumor [55], we identified several genes contributing to the heterogeneity of tumor tissue. In this manuscript, we focused on the spatial visualization of gene expression for HLA-B [56], SPP-1 [57], VEGF-A [58] and TIMP-1 [59, 60]. In the spatial domain detection visualization results of AttentionVGAE, significant expression of HLA-B is observed in regions 2, 3, 5 and 10 (Figure 4D-1). HLA-B is a human leukocyte antigen involved in immune system recognition and antigen presentation. It assumes a vital function in immune responses, specifically in identifying and eradicating infections or aberrant cells. Elevated expression of HLA-B may contribute to enhancing the immune system’s recognition and attack against GBM cells, potentially exerting a positive effect on inhibiting tumor growth.

Differential expression of SPP-1 is distributed in regions 0, 4, 5 and 7 (Figure 4D-2). SPP-1, a secreted phosphoprotein, participates in various biological processes, including cell adhesion, migration, invasion and inflammation. In tumors, elevated expression of SPP-1 is associated with processes such as tumor growth, invasion and angiogenesis. High expression of SPP-1 in GBM may be linked to increased invasiveness and poor prognosis. It may influence disease progression by promoting tumor cell migration and invasion and regulating the tumor microenvironment. VEGF-A, a crucial factor in angiogenesis, shows heightened expression in regions 1 and 3 (Figure 4D-3).

In tumors, elevated levels of VEGF-A are typically associated with neoangiogenesis (angiogenesis). In GBM, high expression of VEGF-A may promote the formation of blood vessels around the tumor, providing the necessary oxygen and nutrients for tumor growth. This may contribute to tumor growth and invasion and may be associated with poor prognosis. On the other hand, different levels of TIMP-1 are also distributed in regions 0, 6 and 10 (Figure 4D-4). In GBM, the expression level of TIMP-1 may affect the activity of MMPs, thereby regulating cell migration and invasion. Elevated levels of TIMP-1 could be linked to heightened tumor invasiveness and an unfavorable prognosis.

These findings highlight AttentionVGAE’s potential to enhance our biological understanding by delving into the potential heterogeneity within the tumor. This is apparent from both spatial domain detection and gene expression patterns. The attention-guided mechanism provides VGAE with enhanced flexibility, enabling adaptive focus on various features or relationships within the data. This adaptive capability allows the model to better capture potential patterns based on the actual structure of the data.

AttentionVGAE further reveals heterogeneity in infiltrating ductal carcinoma data

Through Figure 5A and B, we observed that AttentionVGAE’s detection results have the highest ARI but a slightly lower NMI compared to stLearn’s performance. This situation, rarely encountered in previous research and other articles, suggests the importance of considering multiple metrics for a comprehensive and rigorous analysis. stLearn has historically exhibited poorer performance in recent studies, including our comparative experiments on the DLPFC dataset (Figure 2C). However, in this study of tumor and cancer cell datasets, stLearn exhibited robust performance, highlighting its potential for spatial domain detection in these datasets.

Figure 5.

Figure 5

Open in a new tab

Compares the spatial domain detection results of AttentionVGAE with the latest methods on the infiltrating ductal carcinoma (IDC) dataset. (A) includes high-resolution tissue images of IDC, the ground truth of spatial domain detection for IDC and visualizations of spatial domain detection results for six methods. (B) presents bar charts depicting the ARI and NMI metrics for the spatial domain detection results of the six methods. (C) Showcases UMAP plots for the detection results of the six spatial domain detection methods. (D) Displays spatial expression visualization of genes specific to differential expression analysis. (E) Visualizes the results of differential expression analysis in a volcano plot.

Additionally, in this manuscript, we investigated four differentially expressed genes: MUC1 [61], MGP [62], C1QA [63] and IGFBP5 [64] (Figure 5D). MUC1 encodes an adhesion protein that typically forms a protective mucous layer on the surface of epithelial cells. In cancer, abnormal expression of MUC1 is associated with cell adhesion, proliferation and invasion. MUC1 is often overexpressed in breast cancer, correlating with cancer cell infiltration and migration. MGP encodes a vitamin K-dependent protein involved in calcium metabolism in bone and soft tissue. It is also thought to contribute to the progression of tumors. In the latest research, MGP’s role in infiltrating ductal carcinoma (IDC) has yet to be fully explored, but it is related to its potential role in cell adhesion, proliferation and migration. C1QA encodes a C1q protein subunit in the immune system, participating in the formation of immune complexes. The specific role of C1QA in IDC may be related to the interaction between the immune system and the tumor microenvironment. IGFBP5 encodes an insulin-like growth factor-binding protein that regulates cell growth and apoptosis. IGFBP5 may play a regulatory role in cell growth and invasion in breast cancer. Analysis through AttentionVGAE provides essential evidence for further studying the impact of differentially expressed genes in infiltrating ductal carcinoma on specific regions. It contributes to understanding the heterogeneity of gene expression in spatial terms, subsequent clinical analyses and research on potential prognostic risk genes.

AttentionVGAE excels in UMAP analysis (Figure 5C), indicating that the AttentionVGAE method with an added attention mechanism is more adept at capturing the intrinsic structure and relationships within the data for this specific task. In the presence of complex features and patterns, a model needs adaptability and flexibility, attributes provided by the attention mechanism. The success of AttentionVGAE may result from its ability to dynamically focus on crucial features or relationships within the data, showcasing superior handling of heterogeneity and noise. This capability enhances the accuracy of clustering and spatial domain detection.

AttentionVGAE (adaptive) performs spatial domain detection on high- spatial-resolution samples from the Stereo-seq and Slide-seq platforms

The spatial position of cells in tissues strongly influences their function, but the lack of high-throughput, whole-genome gene expression readings with cellular resolution is a challenge. The Slide-seq technique’s high spatial resolution is crucial for mapping cellular atlases. Aggregating data into larger feature sizes poses challenges in resolving cell types within heterogeneous tissue regions. AttentionVGAE effectively detects fine-scale structures in mouse brain data from the Stereo-seq and Slide-seq platforms. In the data from Stereo-seq’s adult mouse hemi-brain sections, AttentionVGAE with adaptive clustering not only detects fine-scale structures clearly but also filters out noise to a large extent (Figure 6A). When performing spatial domain detection on Hippocampus data from the Slide-seq platform, AttentionVGAE’s results are highly consistent with the Slide-seq platform, especially in capturing fine-scale structures (Figure 6E).

Figure 6.

Figure 6

Open in a new tab

AttentionVGAE performs spatial domain detection with adaptive clustering on high=spatial-resolution samples from the Stereo-seq and Slide-seq platforms. (A) AttentionVGAE (adaptive) detects spatial domain information in the adult mouse hemi-brain data from the Stereo-seq platform. (B) Annotation of sample E12.5E1S1. (C) AttentionVGAE (adaptive) detects spatial domain information in the Dorsal_midbrain_cell sample from the Stereo-seq platform. (D) AttentionVGAE (adaptive) demonstrates spatial domain detection performance on mouse embryo section data. (E) AttentionVGAE (adaptive) identifies intricate structures in the Slide-seq Hippocampus dataset.

AttentionVGAE aims to conduct spatial domain detection by utilizing continuous data from mouse embryo section data obtained from the Stereo-seq platform, revealing the dynamic differentiation process of developmental structures. The results indicate that AttentionVGAE (adaptive) significantly identifies some continuous differentiation processes in the embryo (Figure 6D). Furthermore, AttentionVGAE (adaptive) also demonstrates considerable spatial domain detection capabilities on high-spatial-resolution datasets (Figure 6E). This is closely associated with the excellent performance of the multi-head attention mechanism in handling high-dimensional data, enabling the model to simultaneously focus on multiple aspects of the data, thereby enhancing its ability to capture both local and global relationships in the data. Consequently, AttentionVGAE achieves remarkable spatial domain detection capabilities for high-spatial-resolution datasets.

Performing ablation experiments on AttentionVGAE

The purpose of this section of ablation experiments is to further explore the contribution of the MAH block to the overall AVGN method. Conducting a more in-depth analysis will shed light on the value information of VGAE and the attention module in spatial domain recognition.

In the comparative experiment on the 151 673 slices of the DLPFC, the selected result represents the network outcome with the highest ARI value among 10 runs after removing the MAH block (Figure 7A, right), ARI = 0.582. It can be observed that, for the results after removing the MAH block, AVGN exhibits a more excellent performance in spatial domain detection for Layer_2 in the region marked by the red dashed box, while subtle fractures appear in the results after removing the module. After observing the boundary lines between Layer_6 in the detection results, it is evident that AVGN with the MAH block has smoother and more continuous lines for inter-layer boundaries. We believe this is closely related to the attention mechanism’s ability to represent spatial domain boundaries. After conducting 20 comparative experiments, it was observed that AVGN achieved higher ARI and NMI scores compared to the models without the MHA block (Figure 7C, a).

Figure 7.

Figure 7

Open in a new tab

Results of ablation experiments on three datasets. (A) results of ablation experiments on the DLPFC dataset (151 673 slices) are displayed. From left to right: true labels, spatial domain detection results using AVGN and spatial domain detection results after removing the MHA block from AVGN. The artificially marked dashed boxes and dashed outlines are included for reference. (B) results of spatial domain detection on adult mouse brain slices are presented, emphasizing fine-grained comparisons of locations indicated by dashed boxes. (C) shows the results of ablation experiments focusing on the heterogeneity of infiltrating ductal carcinoma. The upper left and lower left display expression distributions of C1QA and MGP genes, respectively, along with dashed lines indicating manually annotated regions of gene heterogeneity. The four images on the right demonstrate the comparative effects of ablation experiments, also marked with the same dashed lines. (a) The results of the ablation experiment were identified and quantified for 20 times on the sample data of 151 673. On the left is the line box diagram of ARI score, and on the right is the NMI score. (b) Quantitative results of 20 spatial domain identification ablation experiments were performed on invasive ductal carcinoma data. On the left is a wireframing of the ARI score and on the right is the NMI score.

To explore the value of VGAE and the attention module in the spatial domain detection of fine positions, we chose the adult mouse hippocampus as the spatial domain detection object for ablation experiments. In Figure 7B (right), the spatial domain of the mouse hippocampus (highlighted by the border) shows obvious fractures in the results after removing the MAH block, while AVGN detects only extremely subtle fractures, providing an overall coherent and complete detection of the hippocampus. Additionally, the part within the dashed box clearly demonstrates AVGN’s higher sensitivity to structural heterogeneity compared to the network without the MAH block.

Finally, we attempted to reveal the impact of the attention mechanism on capturing gene expression differences in infiltrating ductal carcinoma data (Figure 7C). Through the experiment, we found that the attention mechanism has a minor impact on the general area of gene expression. However, upon closer observation of the region marked by the line, we observed that attention mechanism appear to have a compacting effect on spatial domain detection results for locally high gene expression. This effect is faintly evident in the marked area of the control group for the MGP gene. In the results of the C1QA gene control group, we found that AVGN detected more details compared to the network without the MAH block. After conducting 20 comparative experiments, it was observed that both ARI and NMI scores improved when the MHA block was added to AVGN (Figure 7C, b).

DISCUSSION

As deep learning technology progresses, attention mechanisms and GNNs are increasingly prevalent in spatial transcriptomics research. Our contribution, AttentionVGAE, innovatively integrates these advancements into a variational graph neural network encoder, guided by attention mechanisms. This unique combination allows for adaptive focus on different nodes or edge information within the graph, enhancing model performance, particularly in scenarios lacking prior knowledge of cluster quantity.

One key innovation of AttentionVGAE lies in its ability to address the calibration of low-quality gene expression data and balance attention between local and global structures. This dual focus ensures accurate representation of gene expression patterns while capturing fine-grained local variations and broader tissue structural relationships.

In rigorous benchmark testing, AttentionVGAE demonstrated exceptional performance across various tasks, including detecting fine structures in coronal sections of the adult mouse brain, identifying spatial domains in human glioblastoma multiforme and infiltrating ductal carcinoma datasets and conducting spatial domain detection on high-spatial-resolution samples from platforms like Stereo-seq and Slide-seq. These results underscore the practical utility of AttentionVGAE in advancing spatial transcriptomics research.

Overall, AttentionVGAE has made progress in spatial domain recognition by correcting low-quality gene expression, utilizing multi-head attention modules, graph convolutional encoders and graph deconvolution decoders. However, due to the sparsity of SRT data, AttentionVGAE still faces limitations in correcting low-quality gene expression. Therefore, the next critical step should involve considering the integration of single-cell sequencing data, which offers single-cell resolution, to further calibrate SRT data with high-resolution data. This advancement can drive improvements in the performance of tissue spatial domain recognition.

Key Points

  • AttentionVGAE is a novel approach in spatial transcriptomics research, adept at capturing spatial domain information while addressing challenges like low-quality gene expression calibration and balancing attention between local and global structures.

  • Our architecture integrates multi-head attention with variational graph autoencoder, offering an adaptive solution for scenarios lacking prior cluster quantity knowledge.

  • Leveraging a graph convolutional neural network framework, AttentionVGAE precisely models tissue microenvironment structural relationships using histological images, genetic data and spatial positions.

  • Test results show AttentionVGAE’s superiority in spatial domain detection, advancing our understanding of tissue gene expression patterns and genomic research.

Author Biographies

Lixin Lei is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. His expertise is bioinformatics and machine learning.

Kaitai Han is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. His expertise is deep learning.

Zijun Wang is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. His expertise is deep learning.

Chaojing Shi is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. His expertise is deep learning.

Zhenghui Wang is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. Her expertise is bioinformatics.

Ruoyan Dai is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. Her expertise is bioinformatics.

Zhiwei Zhang is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. His expertise is deep learning.

Mengqiu Wang is a graduate student at the Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology. Her expertise is bioinformatics.

Qianjin Guo is a professor at the Academy of Artificial Intelligence, Beijing University of Petrochemical Technology. His main research fields include intelligent equipment and systems, machine learning, biomedical imaging and ultrafast spectroscopy.

Contributor Information

Lixin Lei, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Kaitai Han, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Zijun Wang, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Chaojing Shi, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Zhenghui Wang, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Ruoyan Dai, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Zhiwei Zhang, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Mengqiu Wang, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

Qianjin Guo, Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing 102617, China.

FUNDING

This work was supported by NSFCs (Nos. 52361145714, 21673252) and the Beijing Municipal Education Commission, China, under grant number 2019821001, and the fund of Climbing Program Foundation from Beijing Institute of Petrochemical Technology (Project No. BIPTAAI-2021007).

CODE AVAILABILITY

The code for AVGN can be obtained at https://github.com/Listen-lei/AttentionVGAE-main.

DATA AVAILABILITY

The paper utilizes publicly available datasets for its analysis. These include the following:

(1) Human DLPFCs within the spatialLIBD at http://spatial.libd.org/spatialLIBD.

(2) Adult mouse brain (fresh-frozen), Adult Human Glioblastoma Multiforme and Infiltrating ductal carcinoma datasets at https://support.10xgenomics.com/spatial-gene-expression/datasets.

(3) Mouse embryo data, Dorsal_midbrain_celand and Adult mouse hemi-brain data at https://db.cngb.org/stomics/mosta.

(4) Hippocampus dataset at https://portals.broadinstitute.org/single_cell/study/slide-seq-study.

References

  • 1. Dries R, Chen J, Del Rossi N, et al. Advances in spatial transcriptomic data analysis. Genome Res 2021;31:1706–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Williams CG, Lee HJ, Asatsuma T, et al. An introduction to spatial transcriptomics for biomedical research. Genome Med 2022;14:68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Moses L, Pachter L. Museum of spatial transcriptomics. Nat Methods 2022;19:534–46. [DOI] [PubMed] [Google Scholar]
  • 4. Tian L, Chen F, Macosko EZ. The expanding vistas of spatial transcriptomics. Nat Biotechnol 2023;41:773–82. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Shi X, Zhu J, Long Y, Liang C. Identifying spatial domains of spatially resolved transcriptomics via multi-view graph convolutional networks. Brief Bioinform 2023;24:bbad278. [DOI] [PubMed] [Google Scholar]
  • 6. Dries R, Zhu Q, Dong R, et al. Giotto: a toolbox for integrative analysis and visualization of spatial expression data. Genome Biol 2021;22:78. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Zhang C, Dong K, Aihara K, et al. STAMarker: determining spatial domain-specific variable genes with saliency maps in deep learning. Nucleic Acids Res 2023;51:e103–3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Zhao E, Stone MR, Ren X, et al. Spatial transcriptomics at subspot resolution with BayesSpace. Nat Biotechnol 2021;39:1375–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Sadrishojaei M, Jafari Navimipour N, Reshadi M, et al. An energy-aware clustering method in the IoT using a swarm-based algorithm. Wirel Netw 2022;28:125–36. [Google Scholar]
  • 10. Gao Y, Wang Z, Xie J, Pan J. A new robust fuzzy c-means clustering method based on adaptive elastic distance. Knowl-Based Syst 2022;237:107769. [Google Scholar]
  • 11. Heryati A, Herdiansyah MI. The application of data mining by using K-means clustering method in determining new students’ admission promotion strategy. Int J Eng Adv Technol 2020;9:824–33. [Google Scholar]
  • 12. Li Y, Hu P, Liu Z, et al. Contrastive clustering. Proc AAAI Conf Artif Intell 2021;35:8547–55. [Google Scholar]
  • 13. Sinaga KP, Yang M-S. Unsupervised K-means clustering algorithm. IEEE. Access 2020;8:80716–27. [Google Scholar]
  • 14. Ezugwu AE, Ikotun AM, Oyelade OO, et al. A comprehensive survey of clustering algorithms: state-of-the-art machine learning applications, taxonomy, challenges, and future research prospects. Eng Appl Artif Intel 2022;110:104743. [Google Scholar]
  • 15. Rajan R, Rajest S, Singh B. Spatial data mining methods databases and statistics point of views. Innov Inf Commun Technol Ser 2021;103–9. [Google Scholar]
  • 16. Pham D, Tan X, Balderson B, et al. Robust mapping of spatiotemporal trajectories and cell–cell interactions in healthy and diseased tissues. Nat Commun 2023;14:7739. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Teng H, Yuan Y, Bar-Joseph Z. Clustering spatial transcriptomics data. Bioinformatics 2022;38:997–1004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Hamerly G, Elkan C. Learning the k in k-means. Adv Neural Inf Process Syst 2003;16. [Google Scholar]
  • 19. Traag VA, Waltman L, Eck NJ. From Louvain to Leiden: guaranteeing well-connected communities. Sci Rep 2019;9:5233. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Kleshchevnikov V, Shmatko A, Dann E, et al. Cell2location maps fine-grained cell types in spatial transcriptomics. Nat Biotechnol 2022;40:661–71. [DOI] [PubMed] [Google Scholar]
  • 21. Cable DM, Murray E, Zou LS, et al. Robust decomposition of cell type mixtures in spatial transcriptomics. Nat Biotechnol 2022;40:517–26. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Zhang S, Li X, Lin J, et al. Review of single-cell RNA-seq data clustering for cell-type identification and characterization. RNA 2023;29:517–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Armingol E, Officer A, Harismendy O, Lewis NE. Deciphering cell–cell interactions and communication from gene expression. Nat Rev Genet 2021;22:71–88. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Palla G, Fischer DS, Regev A, Theis FJ. Spatial components of molecular tissue biology. Nat Biotechnol 2022;40:308–18. [DOI] [PubMed] [Google Scholar]
  • 25. Hu J, Li X, Coleman K, et al. SpaGCN: integrating gene expression, spatial location and histology to identify spatial domains and spatially variable genes by graph convolutional network. Nat Methods 2021;18:1342–51. [DOI] [PubMed] [Google Scholar]
  • 26. Xu H, Fu H, Long Y, et al. Unsupervised spatially embedded deep representation of spatial transcriptomics. Genome Medicine 2024;16:12. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Xu C, Jin X, Wei S, et al. DeepST: identifying spatial domains in spatial transcriptomics by deep learning. Nucleic Acids Res 2022;50:e131. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Shen R, Liu L, Wu Z, et al. Spatial-ID: a cell typing method for spatially resolved transcriptomics via transfer learning and spatial embedding. Nat Commun 2022;13:7640. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Li Z, Chen X, Zhang X, et al. Latent feature extraction with a prior-based self-attention framework for spatial transcriptomics. Genome Res 2023;33:1757–73. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Niu Z, Zhong G, Yu H. A review on the attention mechanism of deep learning. Neurocomputing 2021;452:48–62. [Google Scholar]
  • 31. Dong K, Zhang S. Deciphering spatial domains from spatially resolved transcriptomics with an adaptive graph attention auto-encoder. Nat Commun 2022;13:1739. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Liu T, Fang Z-Y, Zhang Z, et al. A comprehensive overview of graph neural network-based approaches to clustering for spatial transcriptomics. Comput Struct Biotechnol J 2024;23:106–28. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Dwivedi VP, Joshi CK, Luu AT, et al. Benchmarking graph neural networks. J Mach Learn Res 2023;24:1–48. [Google Scholar]
  • 34. Li J, Chen S, Pan X, et al. Cell clustering for spatial transcriptomics data with graph neural networks. Nat Comput Sci 2022;2:399–408. [DOI] [PubMed] [Google Scholar]
  • 35. Liu M, Gao H, Ji S. Towards deeper graph neural networks. Proc. 26th ACM SIGKDD Int. Conf Knowl Discov Data Min 2020;338–48. [Google Scholar]
  • 36. Wu L, Cui P, Pei J, et al. Graph Neural Networks: Foundation, Frontiers and Applications. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2022;4840–1.
  • 37. Xie Y, Xu Z, Zhang J, et al. Self-supervised learning of graph neural networks: a unified review. IEEE Trans Pattern Anal Mach Intell 2022;45:2412–29. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Zeng Y, Wei Z, Yu W, et al. Spatial transcriptomics prediction from histology jointly through transformer and graph neural networks. Brief Bioinform 2022;23:bbac297. [DOI] [PubMed] [Google Scholar]
  • 39. Yoo J, Jeon H, Jung J, et al. Accurate Node Feature Estimation with Structured Variational Graph Autoencoder. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2022;2336–46.
  • 40. Wang D, Zhang Z, Jiang Y, et al. DM3Loc: multi-label mRNA subcellular localization prediction and analysis based on multi-head self-attention mechanism. Nucleic Acids Res 2021;49:e46–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Chen S, Li Q, Zhao J, et al. NeuroPred-CLQ: incorporating deep temporal convolutional networks and multi-head attention mechanism to predict neuropeptides. Brief Bioinform 2022;23:bbac319. [DOI] [PubMed] [Google Scholar]
  • 42. Liu J, Chen S, Wang B, et al. Attention as relation: learning supervised multi-head self-attention for relation extraction. Proceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence2021;3787–93.
  • 43. Maynard KR, Collado-Torres L, Weber LM, et al. Transcriptome-scale spatial gene expression in the human dorsolateral prefrontal cortex. Nat Neurosci 2021;24:425–36. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Chen A, Liao S, Cheng M, et al. Spatiotemporal transcriptomic atlas of mouse organogenesis using DNA nanoball-patterned arrays. Cell 2022;185:1777–1792.e21. [DOI] [PubMed] [Google Scholar]
  • 45. Rodriques SG, Stickels RR, Goeva A, et al. Slide-seq: a scalable technology for measuring genome-wide expression at high spatial resolution. Science 2019;363:1463–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Tu Z, Talebi H, Zhang H, et al. MaxViT: multi-axis vision transformer. Comput Vis – ECCV 2022;2022:459–79. [Google Scholar]
  • 47. Maćkiewicz A, Ratajczak W. Principal components analysis (PCA). Comput Geosci 1993;19:303–42. [Google Scholar]
  • 48. Ram P, Sinha K. Revisiting kd-tree for Nearest Neighbor Search. Proceedings of the 25th acm sigkdd international conference on knowledge discovery & data mining2019;1378–88.
  • 49. Vitányi PMB, Balbach FJ, Cilibrasi RL, et al. Normalized information distance. Inf Theory Stat Learn 2009;45–82. [Google Scholar]
  • 50. Wolf FA, Angerer P, Theis FJ. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol 2018;19:15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51. McInnes L, Healy J, Melville J. UMAP: uniform manifold approximation and projection for dimension reduction. arXiv 2020.
  • 52. Wolf FA, Hamey FK, Plass M, et al. PAGA: graph abstraction reconciles clustering with trajectory inference through a topology preserving map of single cells. Genome Biol 2019;20:59. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53. Kolk SM, Rakic P. Development of prefrontal cortex. Neuropsychopharmacology 2022;47:41–57. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54. Miller BF, Bambah-Mukku D, Dulac C, et al. Characterizing spatial gene expression heterogeneity in spatially resolved single-cell transcriptomic data with nonuniform cellular densities. Genome Res 2021;31:1843–55. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 55. Nieto P, Elosua-Bayes M, Trincado JL, et al. A single-cell tumor immune atlas for precision oncology. Genome Res 2021;31:1913–26. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 56. Michelakos T, Kontos F, Kurokawa T, et al. Differential role of HLA-A and HLA-B, C expression levels as prognostic markers in colon and rectal cancer. J Immunother Cancer 2022;10:e004115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 57. Ding C, Zhang K, Chen Y, Hu H. LncRNA TP73-AS1 enhances the malignant properties of colorectal cancer by increasing SPP-1 expression through miRNA-539–5p sponging. Pathol-Res Pract 2023;243:154365. [DOI] [PubMed] [Google Scholar]
  • 58. Wiszniak S, Schwarz Q. Exploring the intracrine functions of VEGF-A. Biomolecules 2021;11:128. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 59. Schoeps B, Frädrich J, Krüger A. Cut loose TIMP-1: an emerging cytokine in inflammation. Trends Cell Biol 2023;33:413–26. [DOI] [PubMed] [Google Scholar]
  • 60. Brusa S, Terracciano D, Bruzzese D, et al. Circulating tissue inhibitor of metalloproteinases 1 (TIMP-1) at COVID-19 onset predicts severity status. Front Med 2022;9:1034288. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 61. Supruniuk K, Radziejewska I. MUC1 is an oncoprotein with a significant role in apoptosis (review). Int J Oncol 2021;59:68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 62. Caiado H, Cancela ML, Conceição N. Assessment of MGP gene expression in cancer and contribution to prognosis. Biochimie 2023;214:49–60. [DOI] [PubMed] [Google Scholar]
  • 63. Liang Z, Pan L, Shi J, Zhang L. C1QA, C1QB, and GZMB are novel prognostic biomarkers of skin cutaneous melanoma relating tumor microenvironment. Sci Rep 2022;12:20460. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 64. Dittmer J. Biological effects and regulation of IGFBP5 in breast cancer. Front Endocrinol 2022;13:983793. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The paper utilizes publicly available datasets for its analysis. These include the following:

(1) Human DLPFCs within the spatialLIBD at http://spatial.libd.org/spatialLIBD.

(2) Adult mouse brain (fresh-frozen), Adult Human Glioblastoma Multiforme and Infiltrating ductal carcinoma datasets at https://support.10xgenomics.com/spatial-gene-expression/datasets.

(3) Mouse embryo data, Dorsal_midbrain_celand and Adult mouse hemi-brain data at https://db.cngb.org/stomics/mosta.

(4) Hippocampus dataset at https://portals.broadinstitute.org/single_cell/study/slide-seq-study.

Articles from Briefings in Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES