Single cell space transcriptome data analysis method, device and system and storage medium

Abstract

The invention discloses a single cell space transcriptome data analysis method, device and system, and a storage medium. The method comprises the following steps: acquiring single cell space transcriptome data; the spatial transcriptome data comprises a gene expression map and spatial site coordinates; constructing a graph structure considering gene expression similarity and spatial position continuity on the basis of the data, and performing data representation on a gene expression graph by using an adversarial auto-encoder; in combination with a graph neural network, robust potential characterization is learned, training loss is constructed through reconstruction of a gene expression map, and meanwhile, a clustering prediction result of spatial sites is obtained in combination with mcluster clustering. By adopting the technical scheme of the invention, the spatial clustering analysis with higher precision can be realized, and the key spatial functional region in the biological tissue can be identified.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data analysis, and particularly relates to a single-cell spatial transcriptome data analysis method and device, system and storage medium. BACKGROUND

[0002] The rapid development of single-cell omics technology provides a new perspective for bioinformatics research. Single-cell sequencing technology can provide single-cell level RNA information at different sequencing depths, thereby providing key technical support for biological analysis. For example, single-cell RNA sequencing (scRNA-seq) captures gene expression profiles at the resolution of a single cell, enabling researchers to explore the regulatory mechanisms between different cell types or tissues. However, the spatial location of cells within the tissue often affects their functional intensity and interaction with other cells, so spatial information is also crucial for understanding biological processes. The emergence of spatial transcriptomics (ST) provides a new way to solve this challenge.

[0003] In recent years, a series of spatial transcriptome technologies have emerged, including 10xVisium, Slide-seq, Slide-seqv2, MERFIS, SeqFISH and STARmap. 10xVisium, as an early technology representative, has limitations in spatial resolution. Technologies such as Slide-seq and Slide-seqV2 significantly improve this limitation. Methods such as MERFISH, SeqFISH+ and STARmap achieve subcellular level spatial resolution. Spatial resolution defined by the physical size of each point is a key parameter for evaluating spatial transcriptome technology. Higher resolution technology can generate smaller micro-points containing fewer cells. As a research field of great interest, spatial transcriptomics not only provides gene expression information in the spatial context of the tissue, but also reveals functional associations from the spatial dimension.

[0004] In the early stage, machine learning techniques were widely applied to spatial transcriptome data analysis. For example, Giotto employs a hidden Markov model to estimate the spatial similarity between nodes and identifies spatial functional domains through spatial clustering. BayesSpace clusters spatial points with similar expression profiles based on a Bayesian framework, thereby improving the resolution of spatial functional domains. With the successful application of deep learning in various fields, researchers began to introduce it into spatial transcriptome data analysis. SpaGCN is the first method to apply graph neural networks to spatial functional domain identification, capturing features between spatially adjacent nodes through graph convolution networks and performing clustering analysis. SEDR combines autoencoders with graph convolution networks to learn latent embedding representations containing both transcriptome and spatial structure information by jointly modeling gene expression matrices and spatial adjacency matrices, and introduces a clustering loss function to achieve smoother spatial functional domain identification. STAGATE, another graph neural network-based method, uses an adaptive attention mechanism to dynamically allocate adjacent node weights during training, effectively enhancing the consistency of spatial functional domain boundaries. SpaceFlow introduces contrastive learning into spatial transcriptome data analysis by constructing negative samples through random perturbation of spatial graph structures and training the model to distinguish them from real graph structures, thereby improving spatial functional domain identification performance. GraphST further combines self-supervised contrastive learning with graph convolution networks, making full use of the context information between spatial nodes, enabling it to adaptively optimize feature representations during training and enhance spatial functional domain identification capabilities. SiGra performs single-cell spatial analysis through an image-enhanced graph transformer, using imaging features to identify spatial functional domains and enhance sparse noisy transcriptome data.

[0005] Each of the above-mentioned methods has achieved certain results, but still has several limitations: the difference between sequencing depth and spatial resolution generates complex noise, leading to inconsistencies in gene detection results between different datasets, which affects model stability and hinders robust spatial transcriptome analysis. Therefore, these methods face challenges in accurately identifying specific spatial functional domains, especially in areas with sparse coverage or strong expression pattern heterogeneity. In addition, existing methods mainly rely on local adjacency information, limiting their ability to capture long-range spatial information. SUMMARY

[0006] To solve the problems existing in the prior art, the present application provides a single-cell spatial transcriptome data analysis method and device, system, and storage medium, which can achieve higher precision spatial clustering analysis and help identify key spatial functional regions in biological tissues.

[0007] To achieve the above-mentioned purpose, the present application provides the following solutions: A single-cell spatial transcriptome data analysis method, comprising: Obtaining single-cell spatial transcriptome data; the spatial transcriptome data comprises gene expression profiles and spatial site coordinates; According to the single-cell spatial transcriptome data, a graph structure is constructed by taking into account the gene expression similarity and spatial position continuity, and an adversarial autoencoder is used to perform data representation on the gene expression profiles; A robust latent representation is learned in combination with a graph neural network, and a training loss is constructed by reconstructing the gene expression profiles, while a clustering prediction result of the spatial sites is obtained by jointly clustering.

[0008] As a preferred, the data representation process of the adversarial autoencoder adopts an improved hybrid encoder and generator for generative adversarial training, and the improved hybrid encoder integrates a multi-head attention mechanism and a residual connection strategy.

[0009] As a preferred, the loss calculation of the adversarial autoencoder is defined as follows: ; ; Wherein, and represent the latent embedding of real data and simulated data respectively; represents a discriminator configured to assign a predicted label to real and synthetic latent embeddings; the discriminator loss and the generator loss is calculated by a binary cross-entropy function between the real label and the predicted label.

[0010] As a preferred, the improved hybrid encoder is defined as follows: ; ; ; Wherein, is a feature matrix of the spatial site, obtained from the preprocessed gene expression matrix, with a dimension of 200; represents the number of attention heads in the multi-head attention layer; the feature matrix of each spatial site is divided into segments, and each segment is responsible for feature extraction by the corresponding attention head; is a feature subset processed by attention heads, with a dimension of 200 / ; three projection matrices , , convert the input spatial representation into query, key and value vectors respectively; is used to and scaled by the dot product of the attention head output and the input feature; denotes the spatial site low-dimensional embedding extracted by the i-th attention head; denotes the spatial site low-dimensional embedding extracted by the i-th attention head; is a concatenation function used to aggregate the spatial site low-dimensional embeddings of all attention head outputs; parameters is a weight factor that controls the fusion ratio of input and output features in the residual connection; denotes a fully connected layer used to implement linear transformation; denotes the latent embedding representation of the spatial site.

[0011] The application also provides a single-cell spatial transcriptome data analysis device, comprising: A first processing module is configured to obtain single-cell spatial transcriptome data; the spatial transcriptome data comprises gene expression profiles and spatial site coordinates; A second processing module is configured to construct a graph structure that takes into account both gene expression similarity and spatial position continuity according to the single-cell spatial transcriptome data, and use an adversarial autoencoder to perform data representation on the gene expression profiles; A third processing module is configured to learn a robust latent representation in combination with a graph neural network, and construct a training loss through reconstruction of the gene expression profiles, and obtain a clustering prediction result of the spatial sites in combination with mcluster clustering.

[0012] Preferably, the data representation process of the adversarial autoencoder uses an improved hybrid encoder and generator to perform generative adversarial training, and the improved hybrid encoder integrates a multi-head attention mechanism and a residual connection strategy.

[0013] The application also provides a single-cell spatial transcriptome data analysis system, comprising a memory and a processor, wherein the memory stores a computer program that is run by the processor, and the computer program performs a single-cell spatial transcriptome data analysis method when run by the processor.

[0014] The application also provides a storage medium that stores a computer program, and the computer program performs a single-cell spatial transcriptome data analysis method when run.

[0015] Compared with the prior art, the application has the following beneficial effects: This invention preprocesses spatial transcriptome data, including gene expression profiles and spatial site coordinates. Based on this data, a graph structure is constructed that balances gene expression similarity and spatial continuity. An adversarial autoencoder is then used to represent the gene expression map, handling sparse noise. Subsequently, a graph neural network is used to learn robust latent representations, and a training loss is constructed through gene expression map reconstruction. Finally, mcluster clustering is combined to obtain spatial site clustering predictions. Specifically, the adversarial autoencoder data representation process employs an improved hybrid encoder and generator for generative adversarial training. The hybrid encoder integrates multi-head attention and residual connection strategies to enhance feature learning capabilities. The graph convolutional neural network representation process further integrates spatial coordinate information to capture spatial neighborhood structural features. This invention enables higher-precision spatial clustering analysis, aiding in the identification of key spatial functional regions in biological tissues. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the single-cell spatial transcriptome data analysis method according to an embodiment of the present invention; Figure 2 This is a performance comparison chart for evaluating DACN in the recognition spatial domain in this embodiment; Figure 3 This is a performance comparison chart of DACN on the high-resolution dataset (BaristaSeq) in this embodiment; Figure 4 This example illustrates the application of DACN in the subcellular resolution spatial transcriptome dataset (osmFISH). Figure 5 For parameter sensitivity analysis in this embodiment: DACN performance variation graph under different numbers of attention heads (n); Figure 6 This embodiment presents an ablation experiment analysis: a performance comparison chart of DACN and DACN with AAE removal module (DACNw / oAAE). Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1 like Figure 1 As shown, this invention provides a method for analyzing single-cell spatial transcriptome data. Taking the following application scenario as an example, in this scenario, the spatial transcriptome data graph is defined as G=(V,E), where V represents all spatial site nodes in the graph. E represents spatial site edges in the graph where gene expression is similar and adjacent. The target spatial domain is considered as the primary clustering target, and the relationships between other spatial sites are considered as auxiliary information. The input is the spatial transcriptome data graph G, and the output is the spatial domain partitioning result based on spatial clustering.

[0021] To address the aforementioned application scenarios, this invention proposes a single-cell spatial transcriptome data analysis method, also referred to as DACN in this embodiment, used to identify the spatial domain of spatial transcriptome data. In general, DACN first preprocesses spatial transcriptome data containing gene expression profiles and spatial site coordinates; based on the data, it constructs a graph structure integrating gene expression similarity and spatial location continuity; it uses an adversarial autoencoder to perform representation learning on the gene expression data to effectively handle sparse noise in the data; it employs a graph neural network to learn robust latent representations and constructs a training loss function through gene expression reconstruction; finally, it combines a clustering algorithm to achieve spatial site clustering prediction. Specifically, the adversarial autoencoder uses an improved hybrid encoder and generator for generative adversarial training. The hybrid encoder integrates a multi-head self-attention mechanism and a residual connection structure to enhance feature extraction capabilities; the graph neural network further integrates spatial coordinate information during representation learning to effectively capture spatial neighborhood structural features. Finally, the optimized feature representation is used to identify the spatial domain.

[0022] like Figure 1As shown, DACN can be divided into several stages. The information encoding stage encodes graph node features of spatial sites based on gene expression profiles and spatial site coordinates. The feature generation adversarial stage performs generative adversarial learning between real spatial site representations and simulated representations with introduced noise to explore subtle differences in gene expression profiles. The graph node feature learning stage uses a graph convolutional neural network to perform adjacency aggregation on the graph node features of spatial sites, further mining the consistency and difference information of neighboring spatial nodes. The feature optimization stage optimizes the features based on the feature reconstruction loss of the gene expression profile.

[0023] The single-cell spatial transcriptome data analysis method of this invention specifically includes: Step 1. Information Encoding.

[0024] A spatial locus refers to a single spatial unit captured by spatial transcriptome sequencing technology, typically corresponding to a specific microregion or single cell in a tissue section. Each locus contains expression information for all genes at that location, i.e., a gene expression profile. Spatial location refers to the actual physical coordinates of the spatial locus within the tissue, usually represented in two or three dimensions. It describes the spatial distribution structure of cells or microregions and defines the spatial proximity relationships between loci. A weighted undirected graph can be constructed based on the gene expression information and spatial location relationships of spatial loci. , where the vertex set Represents spatial locations and edge sets. The weight matrix represents the connectivity between sites. This indicator reflects the similarity of gene expression patterns and spatial proximity between spatial loci. Gene expression similarity refers to the degree of similarity in gene expression patterns between different spatial loci, quantified by calculating the correlation of the expression levels of all genes at two loci. This indicator reflects the similarity of transcriptional states between loci; higher similarity suggests that the two loci may have similar cell types or functional states. Spatial proximity refers to the physical distance between spatial loci, usually calculated using Euclidean or Manhattan distance. In spatial transcriptomics analysis, proximity reflects the actual spatial distribution of cells or microregions within tissues; loci that are closer together are more likely to have direct cell-cell interactions or signal transduction. For any two spatial loci... and Let their gene expression vectors be respectively and The spatial coordinates are respectively and The combined similarity weight between the two points is then... The calculation formula is: ; in, The Euclidean distance representing gene expression vectors. Represents the Euclidean distance between spatial coordinates. and These are two parameters used to adjust the influence of expression differences and spatial distance on the weight calculation, and are dynamically adjusted according to the loss function. according to The value, when Greater than the preset threshold or belong When constructing a spatial nearest neighbor set (such as k-nearest neighbors), establish connecting edges. This forms the aforementioned weighted undirected graph structure that balances local spatial continuity with consistency in expression patterns.

[0025] An adversarial autoencoder is constructed by combining a hybrid encoder with a generative adversarial network (GAN). The core function of the hybrid encoder is to map gene expression profiles into low-dimensional vectors. Its structure includes a multi-head attention layer and fully connected layers. The multi-head attention layer is used to characterize complex tissue regions that often exhibit highly heterogeneous and disordered expression patterns: each attention head captures different local gene expression relationships between spatial sites. Local gene expression relationships refer to the local correlation between spatial sites in a specific gene expression pattern, reflecting the cooperative expression pattern of neighboring or functionally related sites on a specific subset of genes within the tissue microenvironment. Simultaneously, this invention introduces residual connections within the multi-head attention layer to prevent overfitting or excessive bias towards local feature representations while preserving key information about the original spatial gene expression. The hybrid encoder is defined as follows: ; ; ; in, The feature matrix of spatial sites is obtained from the preprocessed gene expression matrix and has a dimension of 200. This represents the number of attention heads in a multi-head attention layer. The feature matrix of each spatial location is divided into... There are 3 segments, and each segment is responsible for feature extraction by a corresponding attention head. for Each attention head processes a subset of features, with a dimension of 200 / Three projection matrices , , The spatial representation of the input is converted into query, key, and value vectors, respectively. Used for and The dot product is scaled to prevent the result from becoming too large and to reduce the risk of gradient explosion. Indicates the first Low-dimensional embeddings of spatial sites extracted by attention heads. This is the concatenation function, used to aggregate the low-dimensional embeddings of spatial locations from all attention head outputs. Parameters This is a weighting factor that controls the fusion ratio of input and output features in the residual connection. This indicates a fully connected layer, used to implement linear transformations. Latent embedding representation of spatial sites Step 2. Feature generation adversarial.

[0026] In this step, the invention adds a discriminator layer to the hybrid encoder, enabling it to function as a discriminator within the generative adversarial network framework. During training, the generator optimizes gene expression profiles by progressively generating simulated data, effectively compensating for the lack of high-expression signals. After training, the discriminator gains the ability to distinguish between real and simulated data. Through adversarial training between the generator and the discriminator, the discriminator's ability to differentiate between real and generated data continuously improves. In the initial training phase, the discriminator may converge prematurely due to its underdeveloped recognition capabilities, which would hinder the effective updating of the generator's gradients and lead to training instability. To balance the training dynamics of the generator and the discriminator, hyperparameters are introduced. This strategy, used to control the update frequency of the discriminator in a single iteration, significantly improves overall training stability. The loss calculation for the adversarial autoencoder is defined as follows: ; ; in, and These represent the potential embeddings of gene expression profiles from real spatial transcriptome data and gene expression profiles from faked spatial transcriptome data, respectively. The representative discriminator is configured to assign predicted labels to both real and synthetic latent embeddings; the discriminator loss... and generator loss It is calculated using the binary cross-entropy function between the true label and the predicted label.

[0027] Step 3. Graph node feature learning.

[0028] After obtaining low-dimensional embeddings from AAE and a graph structure that balances local spatial continuity and consistency of representation patterns, a three-layer graph neural network is designed to extract spatial features from the graph. The first layer of the graph convolutional network embeds the low-dimensional features (…). The graph is mapped to a dimension matching the spatial feature matrix $x$ mentioned in claim 3; the second and third graph convolutional networks are used to estimate the mean and variance, respectively. To further improve performance, this invention employs a reparameterization technique, enabling the graph embedding to resample from the latent space based on the learned mean and variance. The definition of the graph convolutional network is as follows: ; ; ; ; in, This represents the graph embedding features obtained through the first layer of the graph convolutional network; , and These represent three independent trainable weight matrices; This is an adjacency matrix constructed based on the coordinates of each spatial point; and These correspond to the mean vector and normalized log-variance of the latent space, respectively. This is a random noise vector sampled from a standard normal distribution; This is a low-dimensional vector generated using the reparameterization technique.

[0029] Step 4. Feature optimization.

[0030] A decoder is constructed using two graph convolutional layers. This decoder is used to reconstruct features based on low-dimensional embeddings extracted by an adversarial autoencoder and a graph convolutional network, and the reconstruction loss is calculated through this reconstruction process. The decoder is defined as follows: ; ; in, Indicates cascading and The final low-dimensional vector obtained; This is a technique for applying dropout operations to reduce overfitting; and These represent the adjacency matrix and the trainable weight matrix, respectively. This represents the reconstructed gene expression profile matrix.

[0031] Step 5. Cluster Prediction The mcluster clustering method is used to obtain clustering prediction results for spatial sites. Specifically, a similarity matrix W is first established between the low-dimensional vectors of all spatial sites. Then, based on the similarity matrix, the Laplacian matrix is ​​further constructed and decomposed to obtain a lower-dimensional feature representation. Based on this feature representation, the spatial sites are optimally clustered. The mcluster clustering method is defined as follows: ; ; This is a low-dimensional feature matrix of spatial sites, where rows represent spatial sites and columns represent the features of each spatial site. The bandwidth parameter of the low-dimensional feature similarity kernel controls the decay rate of similarity. The output is the spatial site clustering label vector; To divide the site into K clusters, The fusion similarity between sites i and j This means optimizing the objective function to obtain the optimal cluster partition that maximizes the similarity of expressions within each cluster.

[0032] The following embodiments of the present invention will be combined with specific experimental verification to demonstrate the feasibility and progressiveness of the present invention.

[0033] During experimental verification, the DACN model will be evaluated from the following aspects.

[0034] •RQ1: Does DACN outperform other benchmark spatial clustering models? • Does the design of the RQ2:AAE module contribute to improving the performance of DACN? •RQ3: Do hyperparameters affect the performance of DACN? •RQ4: Does the spatial transcriptome dataset at different resolutions affect the performance of DACN? 1. Dataset and Experiment Configuration Dataset Description: To validate the performance of the DACN model of this invention, the Human Dorsolateral Prefrontal Cortex (DLPFC), BaristaSeq, and osmFISH datasets were used. Data details are shown in Table 1. The DLPFC dataset was generated using 10xGenomicsVisium technology and contains 12 slices from human dorsolateral prefrontal cortex tissue. Each slice covers approximately 4,000 spatial sites and has manually labeled cortical structures. This dataset is used as a reliable dataset for evaluating spatial domain recognition performance due to its well-defined spatial structure and broad applicability in benchmark studies. The BaristaSeq dataset is derived from mouse primary brain cortex tissue and was generated using BaristaSeq high-throughput technology, providing high spatial resolution. The osmFISH dataset was obtained through osmotic pressure-based fluorescence in situ hybridization and provides single-cell resolution spatial transcriptome data of the mouse somatosensory cortex, containing a total of 4,839 sites.

[0035] Parameter configuration: Consistent training parameter settings were used across all datasets, including a learning rate of 0.01 and 550 training epochs. An early stopping mechanism was employed during training, continuously monitoring changes in the loss value and automatically saving the model weights corresponding to the minimum loss value. The number of attention heads in the multi-head attention layer was selected from the parameter space {1, 2, 4, 8}.

[0036] Evaluation Metrics: To evaluate model performance, DACN selected two representative metrics: ARI and NMI. ARI measures the similarity between predicted clustering results and true labels by quantifying the consistency of pairwise samples. NMI, based on information theory, evaluates the mutual information dependency between predicted results and true labels. Higher values ​​for these metrics indicate better model performance.

[0037] Environment Setup: The DACN method was implemented using Python 3.9 and R 4.3, employing the PyTorch framework and integrating the Scanpy and Anndata packages for spatial transcriptome data analysis. The rpy2 package facilitated interaction between Python and R, and the Mclust package within the R environment was used to perform cluster analysis. Model training was completed on a workstation platform configured with an NVIDIA RTX 1060 graphics processor (8GB VRAM).

[0038] Table 1

[0039] 2. Results Analysis The model of this invention was compared with six advanced benchmark models.

[0040] The spatial clustering benchmark model is as follows: SpaGCN is the first model to apply graph neural networks to spatial clustering, capturing features between spatially adjacent nodes and performing cluster analysis through graph convolutional networks.

[0041] SEDR is a spatial clustering model that combines autoencoders with graph convolutional networks.

[0042] STAGATE is a spatial clustering model that uses an adaptive attention mechanism to dynamically allocate the weights of neighboring nodes during training.

[0043] SpaceFlow is a spatial clustering model that introduces contrastive learning into spatial transcriptome data analysis. It constructs negative samples by randomly perturbing the spatial graph structure and trains the model to distinguish them from the real graph structure.

[0044] GraphST is a spatial clustering model that combines self-supervised contrastive learning with graph convolutional networks.

[0045] DeepST is a spatial clustering model based on the integration of multimodal information.

[0046] The experimental results are analyzed in detail below: Performance Evaluation (RQ1, RQ4): This invention systematically evaluated the clustering performance of DACN, as shown in the attached figure. Figures 2-4 As shown. To ensure comprehensive comparison, seven representative methods were selected on the DLPFC dataset: SEDR, SpaGCN, GraphST, SpaceFlow, DeepST, Sigra, and STAGATE. For 12 brain slices, two unsupervised metrics, ARI and NMI, were calculated to measure the consistency between the predicted domain and the manually labeled data. Figure 2 The box plot shows that DACN maintained the highest and most stable score across all samples, significantly outperforming the other methods (t-test, p < 0.05), demonstrating its robustness and effectiveness in spatial domain recognition.

[0047] Taking slice 151674 as an example ( Figure 2 (b) SpaGCN and GraphST have blurred boundaries and obvious layer misalignment; STAGATE and SpaceFlow have clear structures but low ARI; SEDR and DeepST are generally consistent, but invert Layer 2 / 3. The layer distribution obtained by DACN is almost completely consistent with the true anatomical values, with smooth boundaries, no inverted adjacent layers, and correctly classifies the gray "unannotated area" in the lower right corner into the Layer 2 / 3 neighborhood. All other methods show long-distance mismatch.

[0048] Differential gene testing ( Figure 2c) Further verification: The WM region clustered by DACN shows high expression of MBP, and the Layer 1 region shows high expression of MALAT1, completely overlapping with the ground space pattern. UMAP visualization ( Figure 2 d) shows that the latent space of DACN is compact within clusters and has the greatest separation between clusters, indicating that its low-dimensional embedding effectively preserves the intrinsic structure of the organization.

[0049] On high-resolution BaristaSeq data (spot diameter 5–10 μm, 3 slices), DACN also outperforms in ARI / NMI across the board. Figure 3 a); Visualization of slice 2 ( Figure 3 In b), the layered structures of SpaGCN, DeepST, and GraphST are broken, while the DACN domain has a coherent boundary and the smallest misalignment area. Batch effect test ( Figure 3 The results (cd) show that, after Harmony correction, DACN's iLISI is closest to the true batch number, and the batch mixing is the most uniform, significantly better than the control method.

[0050] In subcellular osmFISH data, DACN continues to achieve the highest ARI / NMI ( Figure 4 a); Spatial Patterns ( Figure 4 b) The boundaries are the smoothest. The expression distribution of the glutamatergic marker gene Syt6 / Rorb and the oligodendrocyte marker Plp1 / Anln completely overlaps with the DACN domain. Figure 4 c), further confirming its biological interpretability.

[0051] 3. DACN Model Analysis A) Ablation experiment.

[0052] 1) Impact of the AAE module (RQ2): To verify the importance of the AAE module in improving DACN performance, this invention compares it with DACN on the DLPFC, BaristaSeq, and osmFISH datasets.

[0053] • DACNw / oAAE. Removes the AAE module.

[0054] Experimental results are as follows Figure 5As shown. To evaluate the contribution of the core AAE module, this invention removed the AAE component (denoted as DACNw / oAAE) through ablation experiments and systematically evaluated its performance impact. To ensure fairness and robustness of the comparison, all experiments were performed on three spatial transcriptome datasets with different spatial resolutions. DACN consistently outperformed its ablated version on all datasets. On the low-resolution DLPFC dataset, DACN achieved stable improvements in both ARI and NMI metrics; while on the high-resolution BaristaSeq and osmFISH datasets, the introduction of the AAE module brought more significant performance gains. These results fully demonstrate that integrating the AAE module significantly enhances DACN's ability to capture complex spatial transcriptome patterns.

[0055] B) Hyperparameter Analysis (RQ3) To comprehensively evaluate the impact of the number of attention heads (n), this invention compares the performance of DLCN under different n values ​​(1, 2, 4, 8). To ensure the robustness and representativeness of the comparison, all experiments were conducted on three spatial transcriptome datasets with different spatial resolutions. Figure 6 The results show that DLCN employing a 4-head attention mechanism exhibits the most stable performance across all three datasets: it demonstrates a significant advantage on the DLPFC dataset, while its performance improvement is relatively limited on the BaristaSeq and osmFISH datasets. This difference may stem from the low gene detection throughput of the BaristaSeq and osmFISH datasets, which limits the effectiveness of the multi-head attention mechanism under low-throughput conditions.

[0056] Example 2 The present invention also provides a single-cell spatial transcriptome data analysis device, comprising: The first processing module is used to acquire single-cell spatial transcriptome data; the spatial transcriptome data includes gene expression maps and spatial site coordinates. The second processing module is used to construct a graph structure that takes into account both gene expression similarity and spatial continuity based on single-cell spatial transcriptome data, and to use an adversarial autoencoder to perform data characterization on the gene expression map. The third processing module is used to learn robust latent representations by combining graph neural networks, construct training loss by reconstructing gene expression maps, and obtain clustering prediction results of spatial sites by combining mcluster clustering.

[0057] As one embodiment of the present invention, the data representation process of the adversarial autoencoder employs an improved hybrid encoder and a generator for generative adversarial training. The improved hybrid encoder integrates a multi-head attention mechanism and a residual connection strategy.

[0058] Example 3 The present invention also provides a single-cell spatial transcriptome data analysis system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a single-cell spatial transcriptome data analysis method when executed by the processor.

[0059] Example 4 The present invention also provides a storage medium storing a computer program, which executes a single-cell spatial transcriptome data analysis method when running.

[0060] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for analyzing single-cell spatial transcriptome data, characterized in that, include: Acquire single-cell spatial transcriptome data; The spatial transcriptome data includes gene expression maps and spatial site coordinates; Based on single-cell spatial transcriptome data, a graph structure that balances gene expression similarity and spatial continuity is constructed, and an adversarial autoencoder is used to characterize the gene expression map. We combine graph neural networks to learn robust latent representations and construct training loss through gene expression map reconstruction. At the same time, we combine mcluster clustering to obtain clustering prediction results of spatial sites.

2. The method for analyzing single-cell spatial transcriptome data as described in claim 1, characterized in that, The data representation process of the adversarial autoencoder employs a modified hybrid encoder and a generator for generative adversarial training. The modified hybrid encoder integrates a multi-head attention mechanism and a residual connection strategy.

3. The single-cell spatial transcriptome data analysis method as described in claim 1, characterized in that, The loss calculation for adversarial autoencoders is defined as follows: ； ； in, and These represent the potential embeddings of real data and simulated data, respectively. The representative discriminator is configured to assign predicted labels to both real and synthetic latent embeddings; the discriminator loss... and generator loss It is calculated using the binary cross-entropy function between the true label and the predicted label.

4. The single-cell spatial transcriptome data analysis method as described in claim 1, characterized in that, The improved hybrid encoder is defined as follows: ； ； ； in, The feature matrix of spatial sites is obtained from the preprocessed gene expression matrix and has a dimension of 200. This represents the number of attention heads in the multi-head attention layer; the feature matrix of each spatial location is divided into... Each segment is a separate segment, and each segment is responsible for feature extraction by a corresponding attention head. for Each attention head processes a subset of features, with a dimension of 200 / Three projection matrices , , The spatial representation of the input is converted into query, key, and value vectors, respectively. Used for and Scaling the dot product; Indicates the first Low-dimensional embeddings of spatial sites extracted by attention heads; This is the concatenation function, used to aggregate the low-dimensional embeddings of spatial locations from all attention head outputs; parameters As a weighting factor, it controls the fusion ratio of input and output features in the residual connection; This indicates a fully connected layer, used to implement linear transformations; This represents the potential embedding representation of a spatial location.

5. A single-cell spatial transcriptome data analysis device, characterized in that, include: The first processing module is used to acquire single-cell spatial transcriptome data; The spatial transcriptome data includes gene expression maps and spatial site coordinates; The second processing module is used to construct a graph structure that takes into account both gene expression similarity and spatial continuity based on single-cell spatial transcriptome data, and to use an adversarial autoencoder to perform data characterization on the gene expression map. The third processing module is used to learn robust latent representations by combining graph neural networks, construct training loss by reconstructing gene expression maps, and obtain clustering prediction results of spatial sites by combining mcluster clustering.

6. The single-cell spatial transcriptome data analysis device as described in claim 5, characterized in that, The data representation process of the adversarial autoencoder employs a modified hybrid encoder and a generator for generative adversarial training. The modified hybrid encoder integrates a multi-head attention mechanism and a residual connection strategy.

7. A single-cell spatial transcriptome data analysis system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the single-cell spatial transcriptome data analysis method as described in any one of claims 1-4 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the single-cell spatial transcriptome data analysis method as described in any one of claims 1-4.

Images