Graph contrast learning space transcriptome clustering method based on reliable enhancement and redundancy reduction strategy

Abstract

The invention discloses a graph contrast learning space transcriptome clustering method based on a reliable enhancement and redundancy reduction strategy. The method comprises the following steps: firstly, preprocessing ST data, and establishing a spatial adjacency matrix through spatial context information; thirdly, establishing an undirected graph by integrating gene expression and an adjacent matrix, and performing smooth expression by using a graph Laplacian matrix; and then, generating a reliable enhanced view through a reversible network, carrying out embedded feature learning based on an encoder and a multi-head attention mechanism, reducing redundant information through a redundancy reduction strategy, and carrying out K-means clustering to cluster ST data. And finally, downstream analysis is carried out, including SVGs and DEGs identification and track and pseudo-time inference. According to the method, the model clustering capability is improved through the reversible network and the redundancy reduction strategy, and the problem that the clustering effect is reduced due to the fact that the noise edge is not distinguished and an intersection part exists between the identified spatial domains at present is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and more specifically, to a spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategies. Background Art

[0002] The continuous advancement of single-cell RNA sequencing (scRNA-seq) technology has helped to reveal the variability of gene expression in various cells and promoted the development of cell biology. However, scRNA-seq technology only uses the gene expression profile of the cell and ignores its spatial context, resulting in discontinuity of the identified cell clusters. Spatially resolved transcriptomics (ST) technology provides rich gene expression and spatial location information. Combining gene expression and spatial location information can more effectively decipher cell communication, explore tissue structure, and study disease pathology.

[0003] Current ST sequencing technologies mainly include two categories: (1) high-magnification RNA imaging technology, which locates mRNA transcripts through probes targeting specific genes, such as MERFISH, seqFISH, seqFISH+, and STARmap. (2) spatial labeling technology, which captures mRNA transcripts through spatial barcodes and then performs deep sequencing, such as Slide-seq, Slide-seqV2, and 10X GenomicsVisium (https: / / www.10xgenomics.com / ). They all provide rich gene expression and spatial location information, allowing us to more accurately divide tissue heterogeneity and cell function.

[0004] Spatial domain identification is an essential and important step in the ST data analysis process. Through spatial domain identification, we can better understand the cellular composition and specific transcriptomes in various complex tissues. In addition, we can also perform basic downstream analyses such as the discovery of new / rare cell types, spatially variable genes (SVG) and differentially expressed genes (DEG) detection, and trajectory and pseudo-time inference.

[0005] The current spatial domain recognition methods mainly include traditional clustering and spatial clustering. Traditional clustering methods use gene expression information as input to identify spatial domains, such as k-means, DBSCAN, spectral clustering, and hierarchical clustering. However, the ST clusters they obtain are almost inconsistent with tissue sections. In recent years, graph neural networks (GNNs) have been widely used in various fields such as spatial clustering due to their great potential in processing non-Euclidean structured data. For example, in CN202410347380.X, a spatial transcriptome spatial domain recognition method based on relationship graph attention is disclosed. Gene expression, histological image data and spatial location information are first collected through stRGAT, and then the data is reduced and normalized, and an undirected graph is constructed using standardized Euclidean distance to measure the similarity between nodes. The preprocessed data is then input into the relationship graph attention network to extract the final potential embedding, and finally an unsupervised clustering algorithm is used for spatial domain recognition. This patent improves the accuracy of existing spatial domain recognition algorithms. However, when learning features, they fail to distinguish noise edges, resulting in the existence of overlapping parts between the identified spatial domains, thereby reducing the clustering effect. Summary of the invention

[0006] The main technical problem to be solved by the present invention is to provide a spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategies in view of the deficiency that the prior art fails to distinguish noise edges, resulting in the existence of overlapping parts between the identified spatial domains, thereby reducing the clustering effect.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategies, comprising:

[0009] S1. Data preprocessing;

[0010] The collected spatially resolved transcriptomic data were screened, the expression of the remaining genes was log-transformed and normalized, and the dimensionality reduction of the highly variable genes was performed;

[0011] S2. Graph data construction;

[0012] The feature matrix and adjacency matrix are combined to construct an undirected graph, which is then smoothly represented using the graph Laplacian matrix.

[0013] S3. Spatial clustering;

[0014] S31. Constructing a smooth attribute matrix By perturbing the network p(·;θ), a perturbation attribute matrix is ​​established.

[0015] S32. From two attribute matrices and Learning latent features and

[0016] S33. Residual connection based on multi-head attention is enhanced and The features of and perturbation embedding

[0017]

[0018] S34. Construct a recovery network r(·;θ) to recover the perturbed embedding Get the restored embedding features For the original embedded features Perform disturbance operation to obtain

[0019] S35. Construct the original embedding matrix Separate and restored embedding matrices The perturbed embedding matrix The similarity matrix of :

[0020] Similarity Matrix calculate The i-th sample of The similarity between the j-th samples;

[0021] Similarity Matrix calculate The i-th sample of The similarity between the j-th samples;

[0022] ⊙ represents Hadamar product;

[0023] S36. Define a redundancy reduction module using mean square error Thereby reducing feature redundancy in the embedding space;

[0024] S37. Embed the learned original and perturbation embedding Integration, consensus embedding Cluster the ST data based on the calculated consensus embedding Z;

[0025] S4. Downstream analysis;

[0026] Detect spatially variable and differentially expressed genes and infer trajectories and pseudotime.

[0027] Furthermore, genes with an average expression level of less than 3 counts at each point in the spatially resolved transcriptomics data were filtered out.

[0028] Furthermore, the graph data construction step includes:

[0029] S21. Calculate the Euclidean distance through the image coordinates of the spot, and establish an adjacency matrix A based on the distance;

[0030] S22. Initialize the ST data into an undirected graph structure G = {X, ε, A}, where X∈R N×D is the feature matrix, A∈R N×N is the adjacency matrix, ε represents the set of edges;

[0031] S23. The adjacency matrix A is reconstructed into I represents the identity matrix.

[0032] Furthermore, in the adjacency matrix A, if i and j are neighbors then A ij =A ji =1, otherwise A ij =A ji =0.

[0033] Furthermore, the graph Laplacian matrix is ​​expressed as Where I represents the identity matrix, Denotes the degree matrix, let V = {v1,v2,...,v N} represents a set of N nodes that can be clustered into K categories,

[0034]

[0035] Furthermore, after constructing the similarity matrix in S35, the semantic loss function Ensure the perturbation of the original features and the restoration of the perturbed features.

[0036] Furthermore, the semantic loss function The semantic changes are captured for the perturbation recovery mechanism, expressed as:

[0037]

[0038] Furthermore, the objective loss function of the method is:

[0039]

[0040] is the contrast loss, is the semantic loss, is the redundancy loss, and α and β are hyperparameters.

[0041] Furthermore, the contrast loss For View and The contrast loss between them is expressed as:

[0042]

[0043] in sim(·) represents cosine similarity.

[0044] Furthermore, the redundancy loss is expressed as:

[0045]

[0046] in

[0047] Compared with the prior art, the beneficial effects are:

[0048] The present invention first preprocesses the ST data and establishes a spatial adjacency matrix through spatial context information. Then, an undirected graph is established by integrating gene expression and adjacency matrix, and a graph Laplacian matrix is ​​used for smooth representation. Subsequently, a reliable enhanced view is generated through a reversible network, embedded feature learning is performed based on an encoder and a multi-head attention mechanism, redundant information is reduced through a redundancy reduction strategy, and K-means clustering is performed to cluster the ST data. Finally, downstream analysis is also performed, including SVGs and DEGs identification and trajectory and pseudo-time inference. The present invention makes full use of the idea of ​​reversible networks, generates more reliable enhanced views through perturbation and recovery operations, ensures the integrity of the original semantics, and makes the embedded features we learn more reliable. At the same time, a multi-head attention mechanism combined with a residual network is introduced, which is conducive to adaptively capturing richer low-dimensional embedded features, making the learned features more robust and discriminative. In addition, a strategy to reduce redundancy in the latent space is used to minimize the redundancy between different embeddings, better retain more discriminative features, and further improve the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 : A graph-based spatial transcriptome clustering model description based on graph contrast learning with reliable enhancement and redundancy reduction strategies.

[0050] Figure 2: Spatial domain recognition on the DLPFC dataset; (A) Layer structure with manual annotations on 151,672 slices of DLPFC. (B) Box plot of ARI values ​​of STCGCar and other six methods on 12 DLPFC slices. (C) Spatial domain recognition results on 151,672 slices by Seurat, conST, CCST, STAGATE, DeepST, GraphST and STCGCar. (D) Trajectory inference map generated by conST, CCST and STCGCar on 151,672 slices of DLPFC. (E) Pseudo-time inference results on 151,672 slices of DLPFC by conST, CCST and STCGCar.

[0051] Figure 3 :Spatial expression patterns of genes in DLPFC; (A) Spatial expression patterns of SVGs detected by STCGCar in slice 151673. (B) Spatial expression patterns of SVGs detected by STCGCar in slice 151507. (C) Spatial expression patterns of metagenes detected by STCGCar on slice 151673. Figure 4 : Spatial domain identification of Human Breast Cancer (Block ASection1); (A) Manually annotated hierarchical structure based on pathological features. (B) Histograms of ARI and NMI calculated by STCGCar and six competing methods. (C) Domains identified by conST, CCST, DeepST, STAGATE, GraphST, and STCGCar. (D) Stacked violin plots depicting the expression of the top 3 DEGs on domains 0, 1, 6, 18, and 11. (E) Violin plots of the differentially expressed gene CCND1 in different domains. (F) Spatial expression patterns of svg captured by STCGCar.

[0052] Figure 5:Spatial domain recognition on Mouse Brain Serial Section 1 (Sagittal-Anterior) and STARmap datasets; (A) Spatial domain recognition on Mouse Brain Serial Section 1 (Sagittal-Anterior) by conST, GraphST, STAGATE, CCST and STCGCar. (B) Spatial domains recognized by conST, GraphST, STAGATE, CCST and STCGCar on STARmap and manually annotated domains. (C) Histograms of ARI and NMI calculated by STCGCar on Mouse Brain Serial Section 1 (Sagittal-Anterior) dataset compared with four other methods. (D) Histograms of ARI and NMI calculated by STCGCar on STARmap dataset compared with four other methods.

[0053] Figure 6 : Spatial domain identification on Adult Mouse Brain (FFPE) and Human Breast Cancer (DCIS) datasets; (A) Reference atlas of adult mouse brain annotation (FFPE) of Allen hippocampal formation. (B) Domains identified on Adult Mouse Brain (FFPE) by conST, GraphST, STAGATE, CCST, DeepST, and STCGCar. (C) H&E staining of Human Breast Cancer (DCIS) annotated by Agokos telepathology platform. (D) Domains identified on Human Breast Cancer (DCIS) by conST, GraphST, STAGATE, CCST, DeepST, and STCGCar.

[0054] Figure 7 : Supplementary file DLPFC figure. DETAILED DESCRIPTION

[0055] The present invention will be further explained and illustrated below in conjunction with the embodiments, but the specific embodiments do not limit the present invention in any form.

[0056] Example 1

[0057] This embodiment provides a spatial transcriptome clustering method (STCGCar) based on graph contrast learning with reliable enhancement and redundancy reduction strategy, the steps comprising:

[0058] S1. Data preprocessing;

[0059] The collected spatially resolved transcriptomics (ST) data were screened, and genes with an average expression of less than 3 measurements at each point were deleted. The expression of the remaining genes was logarithmically transformed and normalized. The first 3000 highly variable genes were screened using the SCANPY package of python, and the principal component analysis (PCA) was used to reduce the dimensionality of the highly variable genes, and the main 300-dimensional genes were selected as input.

[0060] S2. Graph data construction;

[0061] S21. Calculate the Euclidean distance using the image coordinates of the spot, then select the first 10 nearest neighbors of each point and build an adjacency matrix A based on the distance. In the adjacency matrix A, if i and j are neighbors, then A ij =A ji =1, otherwise A ij =A ji =0.

[0062] S22. Initialize the ST data into an undirected graph structure G = {X, ε, A}, where X∈R N×D is the feature matrix, A∈R N×N is the adjacency matrix, and ε represents the set of edges.

[0063] S23. Reconstruct the adjacency matrix A into The symmetric normalized graph Laplacian matrix is ​​expressed as

[0064] Where I represents the identity matrix, Denotes the degree matrix, let V = {v1,v2,...,v N} represents a set of N nodes that can be clustered into K categories, construct

[0065] S3. Spatial clustering;

[0066] S31. Considering that graph convolution will affect the robustness of the model, we use the t-layer graph Laplacian filter to construct a smooth attribute matrix

[0067] S32. Construct a perturbation recovery network to generate enhanced views for the attribute matrix Perform network perturbation and establish a perturbation attribute matrix by perturbing the network p(·;θ)

[0068] S33. Using the encoding network from two attribute matrices and Learning latent features and The residual connection method based on the multi-head self-attention module is enhanced and The features of and perturbation embedding Where Q, K, V are and Through three simple linear transformations f = W T X+b is obtained, where W is the weight matrix and b is the bias.

[0069] S34. In order to avoid semantic drift, a multi-layer perceptron is used to build a recovery network r(·;θ) to recover the perturbed embedding Get embedded features There are more semantic similarities with the original embedded features. Similarly, the constructed perturbation network p(·; θ) has more semantic similarities with the original embedded features. Perform disturbance operation to obtain In this way, the perturbed embedding features and the original embedding features have more semantic similarity.

[0070] S35. To obtain more reliable semantics on the enhanced view, construct a similarity matrix calculate The i-th sample of The similarity between the j-th samples.

[0071] Similarly, calculate the similarity matrix in represents the similarity between the original embedded features and the restored embedded features, Represents the similarity between the original embedded features and the perturbation embedded features. The similarity matrix can accurately capture the semantic changes based on the perturbation recovery mechanism.

[0072] S36. Through a loss function To ensure the perturbation of the original features and the recovery of the perturbation features, thereby ensuring the reliability of the enhanced view.

[0073] S37. To alleviate the information redundancy in ST data, a redundancy reduction module is defined using mean square error (MSE) in the latent embedding space. This reduces feature redundancy in the embedding space and effectively retains more discriminative features.

[0074] S38. Embed the learned original and perturbation embedding Integration, consensus embedding Based on the calculated consensus embedding Z, we cluster the ST data using K-means.

[0075] S39. Define the target loss function of the STCGCar method Where α and β are hyperparameters, by minimizing the semantic loss Contrastive loss and redundancy loss To jointly optimize the three objective functions, we cluster ST data accurately.

[0076] Contrastive loss in sim(·) represents cosine similarity. Contrastive loss It is used to amplify the similarity between points in the same cluster while minimizing the similarity between points in different clusters.

[0077] Redundancy loss in

[0078] S4. Downstream analysis;

[0079] Spatially variable genes and differentially expressed genes were detected, and trajectories and pseudotimes were inferred using Monocle3 and PAGA, respectively. Trajectory inference and pseudotime estimation help to elucidate the underlying dynamics of biological processes, including cell development and differentiation, and immune responses. Tracking cell transcriptome profiles over time and estimating their dynamics help to reconstruct their developmental processes.

[0080] Example 2

[0081] This embodiment is evaluated on six publicly available data sets according to the method described in Example 1: five 10X Visium (human DorsoLateral Prefrontal Cortex (DLPFC), Human Breast Cancer (Block ASection 1), Mouse Brain Serial Section 1 (Sagittal-Anterior), AdultMouse Brain (FFPE), Human Breast Cancer (Ductal Carcinoma In Situ,ie, DCIS)) data sets, and one STARmap (STARmap) data set. As shown in Table 1, the first four data sets provide labels, while the remaining two data sets have no label information. In addition, the DLPFC data set contains 12 parts, each of which contains 3,498 to 4789 points and 5 to 7 manually annotated areas. The STARmap data set includes 6 neocortical layers.

[0082] Table 1

[0083] Dataset Cell number Number of genes DLPFC 3498-4789 33538 Human Breast Cancer(Block ASection 1) 2518 19743 Mouse Brain Serial Section 1(Sagittal-Anterior) 2695 19112 STARmap 1027 1020 Adult Mouse Brain(FFPE) 2264 19465 Human Breast Cancer (DCIS) 3789 33601

[0084] This example uses six competing ST clustering algorithms as benchmarks, including Seurat, conST, STAGATE, CCST, GraphST, and DeepST. The experiment was repeated 10 times, and the average result of the 10 experiments was taken as the final result. The Adam optimizer was used to minimize the total loss. For labeled datasets, we use ARI and NMI as evaluation indicators. For unlabeled datasets, we use four internal indicators, Davies-Bouldin index (DB), Calinski-Harabasz index (CH), S_Dbw index, and silhouette coefficient (SC) for performance evaluation.

[0085] ARI is defined by the equation: Where n ij represents the number of shared points between i in the ground truth and j in the prediction result, a i =∑ j n ij and b j =∑ i n ij The NMI is defined as: where MI(U,V) represents the mutual information between the inferred cluster label U and the true label V, and H(·) is the entropy function that evaluates the uncertainty between different cluster labels.

[0086] For the specific hyperparameter settings of the model, we set n_top_genes = 3000, n_hidden_size = 300, and other parameters are shown in Table 2. The hyperparameter settings of other baseline methods are the same as those provided by the corresponding publications.

[0087] Table 2

[0088]

[0089] (1) Evaluate model performance on the DLPFC dataset:

[0090] like Figure 2 As shown in Figure B, compared with the other six clustering methods, STCGCar has the highest ARI. On 12 slices of DLPFC, the median ARI calculated by STCGCar is 0.531, which exceeds the other six methods. It is worth noting that, except for GraphST, the median ARI calculated by all the compared methods is less than 0.5.

[0091] like Figure 2As shown in Figure C, Seurat's ST clustering results are the worst. It can identify the spatial domain WM. However, it cannot aggregate other points in each layer, nor can it capture the obvious boundaries between different layers. The boundary features between these domains are not obvious. CCST divides the WM layer into two spatial domains. The clustering results obtained by DeepST, GraphST, and STCGCar are closer to the manual labels of the DLPFC tissue layers, but DeepST and GraphST divide the third layer into two domains. The ST clustering results of the remaining DLPFC slices are shown in Figure 2. Figure 7 shown.

[0092] In addition, the low-dimensional embedding learned by STCGCar is very helpful for spatial domain inference and downstream analysis. PAGA trajectory inference is performed based on the learned embedding. Figure 2 As shown in D, the trajectory inferred by STCGCar from PAGA is almost linear, showing a clear progression from WM to layer 3. However, the trajectories inferred by CCST and conST are disordered.

[0093] like Figure 2 E shows the visualization of the pseudo-time of the learned embeddings using UMAP. Compared with CCST and conST, the development trajectory obtained by STCGCar is more consistent. In addition, the pseudo-time trajectories obtained by CCST and conST do not develop in only one direction.

[0094] STCGCar achieves better clustering accuracy, which helps to detect SVGs more accurately. Figure 3 As shown in A and 3B, SVGs were identified on 151673 slices of DLPFC in order to further utilize the identified spatial domains. As a result, we identified 51 SVGs on 151673 slices. In particular, 1, 45, 1, and 4 SVGs were included in spatial domains 2, 3, 4, and 6, respectively. STCGCar detected two SVGs, PLP1 and HPCAL1, on slices 151673 and 151507, which were used to mark the neuronal layer and white matter layer, respectively.

[0095] Metagenes consisting of multiple genes are used to mark specific spatial domains. Figure 3 As shown in C, metagenes were captured in four spatial domains 0, 1, 2, and 3. It is noteworthy that the metagene in spatial domain 0 only appeared in layer 1. However, when NEFM, GFAP, and VIM-MBP were added to the metagenes in domain 0, the expression pattern became clearer.

[0096] (2) Evaluate model performance on the Human breast cancer (Block A Section 1) dataset:

[0097] The application of STCGCar in the Human breast cancer (Block A Section 1) dataset is studied. The dataset includes four main forms: DCIS / globular carcinoma in situ (DCIS / LCIS), healthy tissue (Healthy), invasive ductal carcinoma (IDC), and the tumor periphery with low malignant characteristics (tumor margin). Figure 4 As shown in B, STCGCar calculated the highest average ARI (i.e., 0.683) and NMI (i.e., 0.724), significantly exceeding other competing ST clustering methods.

[0098] like Figure 4 As shown in Figure C, the spatial domains identified by STCGCar are more consistent with the manually annotated domains. It accurately captures many domains (such as IDC2, Healthy1, etc.), and the segmentation of each domain is relatively smooth. However, conST has the worst results, and some points with the same cluster label are scattered in multiple domains, and the cluster boundaries are not clear. Compared with conST, GraphST, CCST, STAGATE, and DeepST can identify more distinguishable spatial domains, but they also have outliers and fuzzy boundaries. It is worth noting that STCGCar can describe the domains Healthy1, IDC2, DCIS / LCIS5, and DCIS / LCIS1 more accurately than the other five competing methods.

[0099] like Figure 4 As shown in D, DEG analysis was performed on domains 0, 1, 6, 18, and 11. The expression level of the AZGP1 gene determines the histological grade of breast cancer on domain 1, while the CD24 gene is a key indicator of triple-negative breast cancer. Figure 4 As shown in E, mutations and overexpression of CCND1 can be used to identify sites that may affect cancer development, determine changes in cancer drug resistance, and assess patients' cancer risk. Figure 4 As shown in F, the detected SVGs exhibited distinct spatial expression patterns.

[0100] (3) Evaluate model performance on the Mouse Brain Serial Section 1 (Sagittal-Anterior) and STARmap datasets:

[0101] like Figure 5As shown in A and 5B, the present invention analyzes the performance of STCGCar and other comparison methods on Mouse Brain Serial Section 1 (Sagittal-Anterior) and STARmap. The domains identified by STCGCar are highly consistent with manual annotations, and the segmentation of each identified domain is smoother. On these two datasets, conST and GraphST have the worst clustering results, with cluster points scattered in various fields and blurred cluster boundaries. However, STCGCar has learned more clusters that are the same as manual labels, although CCST and STAGATE still obtain very smooth results.

[0102] like Figure 5 As shown in Figure C, in Mouse Brain Serial Section 1 (Sagittal-Anterior), STCGCar calculated the highest average ARI (i.e., 0.457) and NMI (i.e., 0.735), showing strong clustering accuracy.

[0103] like Figure 5 As shown in Figure 2, on STARmap, STCGCar outperforms the other four state-of-the-art methods in terms of average ARI (i.e., 0.662) and NMI (i.e., 0.721).

[0104] (4) Evaluate model performance on Adult Mouse Brain (FFPE) and Human Breast Cancer (DCIS) datasets:

[0105] In order to verify the high clustering effect of the present invention, the performance of STCGCar on two unlabeled datasets, namely, adult mouse brain (FFPE) and human breast cancer (DCIS), was explored. Figure 6 A is an annotated map of the hippocampal structure of the adult mouse brain from the Allens reference map.

[0106] like Figure 6 As shown in B, conST and STAGATE can only roughly identify the outline of the hippocampal region, but cannot capture smaller spatial domains. They still did not detect the "rope-like" structure (Ammons horn) and "arrow-like" structure (dentate gyrus) in the hippocampus. DeepST and CCST clustered ST data relatively accurately. However, compared with them, STCGCar found more subclusters in the hippocampal structure. For example, STCGCar depicts clearer "rope-like" and "arrow-like" structures in the hippocampus, which is basically consistent with the annotation. Although both GraphST and STCGCar identified smaller spatial domains, STCGCar captured smoother boundaries than GraphST.

[0107] like Figure 6 As shown in D, on Human Breast Cancer (DCIS), STCGCar identifies more continuous and smooth regions than other models, and is more consistent with manually annotated clusters.

[0108] In addition, as shown in Table 3, the DB, CH, SC and S_Dbw values ​​of STCGCar and the other five methods on two unlabeled datasets were calculated. The results show that STCGCar has the smallest DB and S_Dbw and the highest CH calculated on adult mouse brain (FFPE); it has the smallest DB and S_Dbw and the highest CH and SC calculated on human breast cancer (DCIS), indicating that it has strong clustering performance on the unlabeled ST dataset.

[0109] Table 3

[0110]

[0111]

[0112] Example 3

[0113] This embodiment provides an ablation experiment to further evaluate the effects of the reversible network and redundancy reduction strategy on the clustering performance of slices 151510, 151670, and 151673 of Human Breast Cancer (Block A Section 1) and DLPFC. As shown in Table 4, l1, l2, and l3 represent semantic loss, contrast loss, and redundancy loss, respectively.

[0114] Table 4

[0115]

[0116] The results show that both the reversible network and the redundancy reduction strategy work well and greatly improve the ST clustering accuracy. In addition, STCGCar can cluster ST data more accurately through the enhanced view generated by the reversible network based on graph contrastive learning. The redundancy reduction strategy in the embedding space greatly reduces the redundant information in the embedded features, thereby improving the clustering performance of the model.

[0117] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy, characterized in that the steps include: S1. Data preprocessing; The collected spatially resolved transcriptomic data were screened, the expression of the remaining genes was log-transformed and normalized, and the dimensionality reduction of the highly variable genes was performed; S2. Graph data construction; The feature matrix and adjacency matrix are combined to construct an undirected graph, which is then smoothly represented using the graph Laplacian matrix. S3. Spatial clustering; S31. Constructing a smooth attribute matrix By perturbing the network p(·;θ), a perturbation attribute matrix is ​​established. S32. From two attribute matrices and Learning latent features and S33. Residual connection based on multi-head attention is enhanced and The features of and perturbation embedding S34. Construct a recovery network r(·;θ) to recover the perturbed embedding Get the restored embedding features For the original embedded features Perform disturbance operation to obtain S35. Construct the original embedding matrix Separate and restored embedding matrices The perturbed embedding matrix The similarity matrix of : Similarity Matrix calculate The i-th sample of The similarity between the j-th samples; Similarity Matrix calculate The i-th sample of The similarity between the j-th samples; ⊙ represents Hadamar product; S36. Define a redundancy reduction module using mean square error Thereby reducing feature redundancy in the embedding space; S37. Embed the learned original and perturbation embedding Integration, consensus embedding Cluster the ST data based on the calculated consensus embedding Z; S4. Downstream analysis; Detect spatially variable and differentially expressed genes and infer trajectories and pseudotime.

2. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 1, characterized in that: Genes with an average expression of less than 3 counts per point in the spatially resolved transcriptomics data were removed.

3. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 1, characterized in that: The steps of graph data construction include: S21. Calculate the Euclidean distance through the image coordinates of the spot, and establish an adjacency matrix A based on the distance; S22. Initialize the ST data into an undirected graph structure G = {X, ε, A}, where X∈R N×D is the feature matrix, A∈R N×N is the adjacency matrix, ε represents the set of edges; S23. The adjacency matrix A is reconstructed into I represents the identity matrix.

4. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 3, characterized in that: In the adjacency matrix A, if i and j are neighbors then A ij =A ji =1, otherwise A ij =A ji =0.

5. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 1, characterized in that: The graph Laplacian matrix is ​​expressed as Where I represents the identity matrix, Denotes the degree matrix, let V = {v1,v2,...,v N } represents a set of N nodes that can be clustered into K categories, 6. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 1, characterized in that: After constructing the similarity matrix in S35, the semantic loss function is used Ensure the perturbation of the original features and the restoration of the perturbed features.

7. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 6, characterized in that: Semantic loss function The semantic changes are captured for the perturbation recovery mechanism, expressed as:

8. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 1, characterized in that: The objective loss function of the method is: is the contrast loss, is the semantic loss, is the redundancy loss, and α and β are hyperparameters.

9. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 8, characterized in that: Contrastive loss For View and The contrast loss between them is expressed as: in sim(·) represents cosine similarity.

10. The spatial transcriptome clustering method based on graph contrast learning with reliable enhancement and redundancy reduction strategy according to claim 8, characterized in that: Redundancy loss It is expressed as: in

Images