Spatial domain recognition method and device of multi-graph convolutional network fused with gene ontology

Abstract

This invention discloses a spatial domain recognition method and apparatus using a multi-graph convolutional network that integrates gene ontology, relating to the fields of bioinformatics and artificial intelligence. The method includes: acquiring and preprocessing raw spatial transcriptome data to obtain a unified node feature vector; simultaneously constructing multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph; inputting the unified node feature vector and the multiple independent heterogeneous graphs into a pre-trained spatial domain recognition network; firstly, using multiple graph convolutional encoders to perform message passing on the multiple independent heterogeneous graphs, outputting corresponding multi-path embedding features; then, using an attention fusion module to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, outputting a globally fused embedding feature; finally, performing clustering and refinement processing on the globally fused embedding feature to output the spatial domain recognition result. This method can effectively integrate multi-dimensional biological prior knowledge, improving the accuracy and robustness of spatial domain recognition.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and artificial intelligence, specifically relating to a spatial domain recognition method for multi-graph convolutional networks that integrate gene ontology. Background Technology

[0002] Spatial transcriptomics, capable of high-throughput sequencing of cellular gene expression profiles while preserving tissue spatial location information, has become an important tool for elucidating complex tissue cellular heterogeneity, cellular microenvironment, and disease development mechanisms. Commonly used spatial transcriptomics platforms include 10x Visium, Slide-seq, and STARmap, which can simultaneously acquire gene expression levels and their two-dimensional coordinates for thousands of spatial loci.

[0003] However, spatial transcriptome data is characterized by high sparsity, high noise, and technical gene deletions. Learning high-quality low-dimensional embeddings from the original gene expression matrix to accurately identify spatial regions with consistent transcriptome characteristics in tissue sections remains a core challenge in the field. Spatial region identification technology requires the effective integration of gene expression information and spatial location information at spatial sites. The identification results directly impact downstream applications such as cell communication analysis, cellular microenvironment analysis, and disease biomarker mining.

[0004] Traditional spatial domain identification methods based on graph neural networks are mainly represented by STAGATE (Dong & Zhang, 2022, Nature Communications) and GraphST (Longetal., 2023, Nature Communications). These methods construct each spatial site as a graph node, establish a spatial adjacency graph based on the physical distance between spatial sites, and use graph attention autoencoders or graph neural networks to learn low-dimensional embeddings that fuse spatial information. Finally, spatial region identification is achieved through clustering. All of these methods use spatial physical distance as the sole basis for graph construction, and node features are derived solely from the original gene expression profiles, without incorporating multidimensional biological prior knowledge such as metabolic pathway activity or gene function ontology. In recent years, multi-graph fusion methods, represented by Path-MGCN (Zhou et al., 2025, Briefings in Bioinformatics) and SiDMGF (Ma et al., 2026, Briefings in Bioinformatics), have further incorporated pathway activity information into graph construction. These methods use multi-view graph convolutional networks (GCN) and dual-graph fusion frameworks to identify spatial regions, which has enriched the utilization of biological information to a certain extent.

[0005] Despite the progress made by the aforementioned existing technical solutions, the following shortcomings still exist: First, existing methods lack information integration at the gene ontology (GO) level. Gene ontology quantifies the activity of each spot at the gene function level by performing statistical enrichment analysis on the functional annotations of active genes in a spot, thereby calculating the functional similarity between spots and constructing a functional similarity graph. GO focuses on characterizing the similarity between spots at the biological process level, while KEGG (Kyoto Encyclopedia of Genes and Genomes) focuses on characterizing the similarity between spots at the metabolic pathway activity level. Their different descriptive perspectives naturally complement each other. Existing methods do not incorporate GO functional similarity into the graph construction or embedding learning process, resulting in learned embeddings lacking semantic information at the gene function level. Path-MGCN and SiDMGF both treat GO and KEGG pathways as interchangeable single information sources, selecting only one to construct the pathway graph each time, without simultaneously modeling them as independent graph structures in parallel.

[0006] Secondly, existing methods do not consider expression reliability in their spatial contrastive loss design. Current methods impose spatial structure constraints equally on all spots, but in reality, some spots have extremely low gene expression levels due to technical dropout. These noisy spots can interfere with contrastive learning, and existing methods lack a mechanism to weight losses based on expression reliability differences. Both Path-MGCN and SiDMGF introduce spatial structure losses, but they treat the loss contribution of all spots indiscriminately, failing to consider the expression reliability differences among spots caused by technical dropout.

[0007] Furthermore, existing methods lack adaptive fusion mechanisms for multi-view biological maps. Spatial proximity maps, KEGG pathway maps, and GO function maps belong to three heterogeneous views: physical space, metabolic activity, and gene function. The contribution ratios of these three types of information to spatial domain partitioning vary significantly across different tissues and sequencing datasets. Existing multi-pathway methods employ fixed-weight fusion strategies, which cannot adaptively adjust the information ratios of the three views based on data characteristics. This can easily lead to problems such as effective biological information being masked by redundant noise and mutual interference between view information, thus limiting generalization performance.

[0008] In summary, existing spatial domain identification methods have significant shortcomings in terms of the comprehensiveness of biological information utilization, the rationality of spatial comparison constraints, and the adaptability of multi-source information fusion. There is an urgent need for a new spatial domain identification method that can integrate gene ontology information and adaptively integrate multi-graph information. Summary of the Invention

[0009] To address the aforementioned problems in the existing technology, this invention provides a spatial domain recognition method and apparatus that integrates gene ontology with multi-graph convolutional networks. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, this invention proposes a spatial domain recognition method for multi-graph convolutional networks that integrate gene ontology, comprising the following steps: We obtained raw spatial transcriptome data, including gene expression matrices and two-dimensional spatial coordinates for each spatial site, and preprocessed the raw data to obtain a unified node feature vector shared by all spatial sites. Using spatial sites as nodes and a unified node feature vector as node features, multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph, are constructed simultaneously. Multiple independent heterogeneous graphs are input into a pre-trained spatial domain recognition network. First, multiple graph convolutional encoders with the same structure and independent parameters are used to perform message passing on the multiple independent heterogeneous graphs, and output multi-path embedding features respectively. Then, the attention fusion module is used to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, and output a global fused embedding feature. Clustering is performed on the globally fused embedded features to obtain preliminary spatial domain division labels; the preliminary spatial domain division labels are refined using a neighbor majority voting strategy to obtain spatial domain recognition results.

[0010] Secondly, this invention proposes a spatial domain recognition device for multi-graph convolutional networks incorporating gene ontology, used to implement the spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology provided in the first aspect of this invention. The device includes: The preprocessing module is used to acquire the original spatial transcriptome data, which includes gene expression matrices and two-dimensional spatial coordinates of each spatial site, and to preprocess the original data to obtain a unified node feature vector shared by all spatial sites. The graph construction module is used to simultaneously construct multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph, using spatial sites as nodes and a unified node feature vector as node features. The inference module is equipped with a pre-trained spatial domain recognition network, which includes multiple graph convolutional encoders with identical structures and independent parameters, and an attention fusion module. The multiple graph convolutional encoders are used to perform message passing on multiple independent heterogeneous graphs, and output multi-path embedding features accordingly. The attention fusion module is used to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, and output a globally fused embedding feature. The clustering and identification module is used to perform clustering processing on the globally fused embedded features to obtain preliminary spatial domain division labels; the preliminary spatial domain division labels are refined using a neighbor majority voting strategy to obtain spatial domain identification results.

[0011] The beneficial effects of this invention are: 1. The spatial domain recognition method of multi-graph convolutional networks integrating gene ontology provided by this invention first preprocesses the original data to obtain a unified node feature vector; then, based on the unified node feature vector, it constructs multiple independent heterogeneous graphs including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph; next, it inputs the multiple independent heterogeneous graphs into a pre-trained spatial domain recognition network, and sequentially uses multiple graph convolutional encoders with identical structures and independent parameters to perform message passing on the multiple independent heterogeneous graphs, correspondingly outputting multi-path embedding features; then, it uses an attention fusion module to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, outputting a global fused embedding feature; finally, it performs clustering processing on the global fused embedding feature and refines the obtained preliminary spatial domain division labels to finally obtain the spatial domain recognition result. This method simultaneously constructs spatial proximity maps, KEGG metabolic pathway similarity maps, and GO functional similarity maps to characterize spatial site similarity relationships from different dimensions, enriching the multi-scale biological information carried by the embeddings. At the same time, it configures independent graph convolutional encoders to extract features separately. Compared with existing technologies that model single graphs and single pathways, it can accurately distinguish the spatial domain boundaries caused by subtle differences in cell function, significantly improving the accuracy of spatial domain recognition. On the other hand, it adaptively fuses the three-way embedding features through a learnable attention fusion mechanism, avoiding information redundancy and interference, and solving the defects of fixed fusion of multiple views and poor generalization.

[0012] 2. This invention incorporates a ZINB decoder in the spatial domain recognition network, which participates in computation only during model training and is disabled during inference. Combined with confidence-weighted spatial contrastive loss training logic, and based on the reliability expression of missing probability quantification at specific sites, it achieves automatic deweighting of low-quality noise sites. Ablation experiments show that after removing the ZINB confidence weighting, the ARI decreases from 0.631 to 0.587 (a 7.0% decrease) on the DLPFC 151510 dataset and from 0.574 to 0.519 (a 9.6% decrease) on the 151670 dataset, exhibiting significant performance degradation across multiple DLPFC slices.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0014] Figure 1A schematic flowchart of a spatial domain recognition method using a multi-graph convolutional network that integrates gene ontology, provided in an embodiment of the present invention; Figure 2 This is another flowchart illustrating the spatial domain recognition method of a multi-graph convolutional network that integrates gene ontology, as provided in an embodiment of the present invention. Figure 3 This is a structural block diagram of a spatial domain recognition device for a multi-graph convolutional network that integrates gene ontology, provided in an embodiment of the present invention. Detailed Implementation

[0015] 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 embodiments described below 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.

[0016] The first aspect of this invention provides a spatial domain recognition method for multi-graph convolutional networks that integrate gene ontology. (See also: [link to related information]) Figure 1 and Figure 2 , Figure 1 This is a schematic flowchart of a spatial domain recognition method using a multi-graph convolutional network that integrates gene ontology, provided in an embodiment of the present invention. Figure 2 This is another schematic flowchart illustrating a spatial domain recognition method using a multi-graph convolutional network incorporating gene ontology, provided in an embodiment of the present invention. The method mainly includes the following steps: Step 1: Obtain the original spatial transcriptome data, which includes gene expression matrices and two-dimensional spatial coordinates for each spatial site, and preprocess the original data to obtain a unified node feature vector shared by all spatial sites.

[0017] In this embodiment, the raw spatial transcriptome data can be obtained from the output files of commercial spatial transcriptome sequencing platforms such as 10x Visium, Slide-seq, and STARmap. The data contains two types of core information: the first is a gene expression matrix, where each row of the matrix corresponds to a tissue spatial sampling site spot, and each column corresponds to a gene. The values ​​in the matrix represent the sequencing capture expression count of the gene at the corresponding site. The second is two-dimensional spatial coordinate information, which stores the two-dimensional planar coordinates of each spot on the tissue slice, used for subsequent construction of spatial proximity relationships.

[0018] Optionally, in this embodiment, the original data is preprocessed, including: screening for highly variable genes, normalizing the total amount, performing logarithmic transformation, and PCA dimensionality reduction on the gene expression matrix in the original data in sequence.

[0019] Specifically, since the original gene expression matrix contains tens of thousands of genes, many of which have extremely low expression levels and lack discriminative power, there is feature redundancy. By calculating the degree of expression variation of each gene across all spatial sites, highly variable genes with the highest expression variation are screened, while genes with low variation and no biological distinguishing value are removed. This significantly reduces the dimensionality of subsequent calculations and reduces the computational overhead and noise interference caused by invalid features.

[0020] Furthermore, since the total amount of sequencing captured varies at different spatial sites, it can cause systematic biases in expression levels between sites. Therefore, this embodiment uses the sum of all gene expressions at a single spatial site as a benchmark, and normalizes the original expression count of each gene within that site to eliminate data bias caused by inconsistent sequencing depths at different spots and to unify the dimensions of expression levels at each site.

[0021] Furthermore, given the large dynamic range of the normalized expression values, which can easily lead to gradient instability when directly input into the network, this embodiment also performs a logarithmic transformation on the normalized expression values ​​to compress the dynamic range of the values, weaken the interference of extremely high-expression genes on feature extraction, and make the data distribution more suitable for the input requirements of the subsequent neural network.

[0022] Even after the above processing, the resulting gene features still have high dimensionality and exhibit multicollinearity. Therefore, this embodiment employs Principal Component Analysis (PCA) to linearly reduce the dimensionality of the logarithmically transformed gene features, retaining the top principal components that can explain the main variation information in the data, thus compressing the high-dimensional gene features into a fixed low-dimensional vector. This low-dimensional vector serves as a unified node feature vector, shared by all spatial sites, and is used as the unified node input feature for the subsequent three types of heterogeneous graphs.

[0023] After the above preprocessing, the original spatial transcriptome data outputs a unified node feature vector with uniform dimensions, eliminating sequencing bias and removing redundant low-information genes. Each spatial site corresponds to an independent node feature vector.

[0024] Step 2: Using spatial sites as nodes and unified node feature vectors as node features, simultaneously construct multiple independent heterogeneous graphs, including spatial proximity graphs, KEGG metabolic pathway similarity graphs, and GO functional similarity graphs.

[0025] Optionally, in this step, this embodiment constructs three different graph structures to characterize the similarity relationships between spots from multiple perspectives. The nodes of the three graphs are all N spots from the original spatial transcriptome data, and the node features all adopt the unified node feature vector obtained in step 1. The first type of graph is the spatial proximity graph As, which is constructed based on the physical coordinate information of the spots and reflects the proximity relationships in tissue space. It can be constructed using the following method: calculating the Euclidean distance between each spatial site based on the two-dimensional spatial coordinates of each spatial site, and establishing undirected adjacency edges for spatial sites with distances less than a preset radius, thus obtaining the spatial proximity graph As.

[0026] Specifically, after obtaining the two-dimensional physical coordinate information of each spot, the Euclidean distance between each pair of spots is calculated. When the Euclidean distance between two spots does not exceed the preset radius r, an undirected edge is established between them and added to the edge set of the spatial proximity graph As.

[0027] Understandably, the value of the preset radius r can be set according to the characteristics of the dataset. For example, for the human dorsolateral prefrontal cortex (DLPFC) dataset, the preset radius r can be set to 300 micrometers; for the human breast cancer (HBC) dataset, the preset radius r can be set to 550 micrometers.

[0028] It should be noted that after constructing the spatial proximity graph As, it is also necessary to construct a positive sample pair matrix and a negative sample pair matrix based on the spatial proximity graph As. The positive sample pair matrix records the set of spatial neighbors for each spot, and the negative sample pair matrix records the set of non-neighbors. These matrices are used for the subsequent calculation of spatial contrast loss.

[0029] The second type of graph is the KEGG metabolic pathway similarity graph Ap. This graph is constructed based on the activity scores of KEGG metabolic pathways and reflects the similarity between spots at the metabolic pathway level. It can be constructed as follows: use gene set variation analysis tools to score the gene set in the KEGG database to obtain the pathway activity score matrix for each spatial site; based on the cosine similarity between the row vectors of the pathway activity score matrix, use the k-nearest neighbor algorithm to establish pathway similarity adjacency edges to obtain the KEGG metabolic pathway similarity graph Ap.

[0030] Specifically, the GSVA (Gene Set Variation Analysis) package in R can be used to score the full set of metabolic pathway genes in the KEGG database, obtaining the activity score matrix P of each spot on each KEGG pathway. Then, the cosine similarity between the row vectors of the P matrix is ​​calculated, and the k=15 neighbors with the highest similarity for each spot are selected to construct the edge set of the KEGG metabolic pathway graph Ap.

[0031] The third type of graph is the GO functional graph Ag, which is constructed based on GO gene function annotation and reflects the similarity between spots at the gene function level. It can be constructed as follows: use gene set variation analysis tools to score the GO gene set to obtain the GO activity score matrix for each spatial site; based on the cosine similarity between the row vectors of the GO activity score matrix, use the k-nearest neighbor algorithm to establish gene function similarity adjacency edges to obtain the GO functional similarity graph Ag.

[0032] Specifically, the GSVA package in R is used to score the entire GO gene set, obtaining the GO activity score matrix G for each spot. Then, the cosine similarity between the row vectors of the G matrix is ​​calculated, and the k=15 most similar neighbors of each spot are selected to construct the edge set of the GO functional graph Ag.

[0033] At this point, the construction of three independent heterogeneous graphs has been completed.

[0034] This invention constructs spatial proximity maps, KEGG metabolic pathway similarity maps, and GO functional similarity maps covering the entire BP (Biological Process), MF (Molecular Function), and CC (Cellular Component) branches simultaneously. Figure 3 The class-independent heterogeneous graph incorporates three complementary types of information—tissue physical location, metabolic pathway activity, and global gene function—in parallel modeling, overcoming the limitations of existing technologies that rely on only a single spatial mapping or selectively use only KEGG / GO pathway information. Combined with a post-configured three-way independent graph convolutional encoder, it extracts site embeddings from different biological perspectives, enabling the final globally fused embedding to simultaneously carry multi-layered biological features of spatial structure, metabolism, and gene function. This allows for a more accurate characterization of the real biological similarities between different spatial sites, significantly improving the accuracy of distinguishing heterogeneous spatial domains in tissues.

[0035] Step 3: Input multiple independent heterogeneous graphs into the pre-trained spatial domain recognition network. First, use multiple graph convolutional encoders with the same structure and independent parameters to perform message passing on the multiple independent heterogeneous graphs respectively, and output multi-path embedding features accordingly. Then, use the attention fusion module to adaptively integrate the multi-path embedding features based on the learnable attention fusion mechanism, and output the global fused embedding features.

[0036] Optionally, this embodiment constructs a ZINB (Zero-Inflated Negative Binomial) decoder that includes multiple graph convolutional network (GCN) encoders, an attention fusion module, and a decoder that participates in computation only during the model training phase and is not enabled during the inference phase.

[0037] Optionally, in this embodiment, each graph convolutional encoder includes two graph convolutional layers, each followed by a ReLU activation function; then, multiple sets of graph convolutional encoders with identical structures and independent parameters are used to perform message passing on multiple independent heterogeneous graphs, correspondingly outputting multi-path embedded features, including: The unified node feature vector is combined with multiple independent heterogeneous graphs and then input into the corresponding graph convolutional encoder. Two graph convolutional layers are used to extract features in sequence, and the ReLU activation function is used to complete the non-linear feature mapping, outputting multi-path embedding features with a unified 64-dimensional dimension.

[0038] Specifically, the three encoders take the same node features as input and use their respective graph structures As, Ap, and Ag to perform message passing and feature aggregation, outputting three 64-dimensional embeddings Hs, Hp, and Hg. Hs corresponds to the encoding result of the spatial proximity graph As, Hp corresponds to the encoding result of the KEGG metabolic pathway graph Ap, and Hg corresponds to the encoding result of the GO functional graph Ag.

[0039] Furthermore, the attention fusion module is used to adaptively integrate the three embedded features based on a learnable attention fusion mechanism, outputting a globally fused embedded feature, including: After stacking the multi-path embedding features along the dimensions, an attention score for each spatial location is calculated using a Linear-Tanh-Linear (linear-hyperbolic tangent-linear) network structure, and the attention weight is obtained by Softmax normalization. The multi-path embedding features are then weighted and summed using the attention weights, and then mapped via an MLP (Multi-Layer Perceptron) to generate a global fusion embedding feature that integrates multi-dimensional biological information such as spatial location, metabolic pathways, and gene function.

[0040] Specifically, the three embeddings Hs, Hp, and Hg are stacked along a new dimension to obtain a tensor of dimension N×3×64. Then, a Linear-Tanh-Linear structure is used to calculate the three-image attention weights for each spot. The first Linear layer maps the 3×64-dimensional input to 3 dimensions, the Tanh layer introduces a non-linear transformation, and the second Linear layer maps the 3-dimensional vector to the attention score. The attention scores are then normalized using a Softmax function to obtain the three-image attention weights. The three embeddings are weighted and summed according to the attention weights to obtain an intermediate representation of dimension N×64. Finally, the final fused embedding feature Z, with dimension N×64, is obtained through MLP mapping.

[0041] Unlike existing technologies that use fixed weights for fusion, this invention utilizes adaptive learning of attention weights to enable the model to dynamically adjust the contribution ratio of the three types of graph information based on data characteristics. When the information of a certain graph structure is more relevant to the spatial domain recognition task, its corresponding attention weight will automatically increase, weakening the interference of redundant information. There is no need for manual adjustment of fusion weights, and it can stably output high-quality fusion embeddings on multiple types and batches of spatial transcriptome datasets, greatly improving the model's cross-dataset generalization ability.

[0042] Furthermore, the ZINB decoder includes a fully connected layer, a BatchNorm layer, and a ReLU activation function; The fully connected layer is used to receive the globally fused embedded features output by the attention fusion module and complete the dimension mapping to obtain the dimension-mapped features; The BatchNorm layer contains learnable scaling and offset coefficients, which are used to perform standardization on the dimension-mapped features to stabilize the model training gradient, accelerate network convergence, and obtain standardized features. The ReLU activation function is used to perform a non-linear mapping on the standardized features to obtain the non-linearly mapped features. The ZINB decoder is also used to predict the probability of technical dropout for each gene at each spatial site based on the features after nonlinear mapping.

[0043] The spatial domain recognition network is trained through the following steps: 1) Input the globally fused embedded features into the ZINB decoder to predict the probability of technology missing for each gene at each spatial site.

[0044] Specifically, after the training data has been preprocessed, graph constructed, and processed by the graph convolutional encoder and attention fusion module, the resulting global fusion embedded features are input into the ZINB decoder. After passing through the fully connected layer, the BatchNorm layer, and the ReLU activation function, the dropout probability π of each spot for each gene is predicted, forming the distribution parameters of the ZINB decoder.

[0045] 2) Calculate the reliability confidence of each spatial site based on the probability of technology missing; construct a confidence-weighted spatial contrast loss based on the reliability confidence, and combine it with the reconstruction loss of the ZINB decoder to construct the total loss.

[0046] First, using the dropout probability π output by the ZINB decoder, the reliability confidence level of each spot is defined, and the calculation formula is as follows:

[0047] In the formula, Indicates the first i Reliability confidence level of each spatial location, Indicates the total number of genes. The first predicted by the ZINB decoder The spatial location of the first The probability of technical missing genes; Then, a confidence-weighted spatial contrast loss is constructed.

[0048] Specifically, L2 normalization is performed on the globally fused embedded features, and the Sigmoid cosine similarity between spatial sites is calculated using the following formula:

[0049] In the formula, Representing the The spatial location and the first Similarity of spatial locations, and They represent the first The spatial location and the first Embedding features of spatial sites; Represents the L2 norm; For the For each spatial location, the single-point spatial contrast loss is defined as:

[0050] In the formula, Indicates the first The contrast loss at the _th spatial location, this value is based on the _th ... The embedding similarity between each spatial location and its set of spatial neighbors and its set of non-neighbors is calculated. Indicates the first The set of neighbors of a spatial location. Indicates the first A set of non-neighbors of spatial locations This indicates the total number of spatial locations.

[0051] Using reliability confidence level as the weighting coefficient, the confidence level-weighted spatial comparison loss is:

[0052] In the formula, This represents the confidence-weighted spatial contrast loss.

[0053] The confidence-weighted spatial contrast loss designed in this embodiment makes high-confidence spots contribute more to the loss function, while the contribution of low-confidence spots to the loss function is automatically reduced, thereby achieving automatic deweighting of noisy spots.

[0054] Finally, combining the reconstruction loss of the ZINB decoder, the total loss is obtained as follows:

[0055] In the formula, Indicates the total loss. This represents the reconstruction loss of the ZINB decoder. , This is a hyperparameter.

[0056] 3) Based on the total loss, the learnable parameters of the graph convolutional encoder, attention fusion module and ZINB decoder are updated iteratively by backpropagation of the optimizer to train the spatial domain recognition network and obtain the pre-trained spatial domain recognition network.

[0057] Optionally, the training process in this embodiment can be implemented using the NAdam optimizer. After training reaches a preset number of rounds, such as 500, the training ends, and the trained spatial domain recognition network is obtained.

[0058] Subsequently, the graph convolutional encoder and attention fusion module in the trained spatial domain recognition network can be used for data processing to obtain globally fused embedded features.

[0059] Step 4: Perform clustering processing on the global fusion embedded features to obtain preliminary spatial domain division labels; refine the preliminary spatial domain division labels using the neighbor majority voting strategy to obtain spatial domain recognition results.

[0060] Specifically, this embodiment performs KMeans clustering on the globally fused embedded features and executes spatial neighbor majority voting refinement. The number of clusters K in KMeans clustering is set based on prior knowledge of the dataset; for the DLPFC dataset, K=7, and for the HBC dataset, K=20. The spatial neighbor majority voting refinement process includes refining the initial labels obtained from KMeans clustering. For each spot, the cluster labels of its preset number (set to 50) nearest neighbor spots are counted, and the label with the highest frequency is taken as the final label for that spot. This refinement process utilizes the label consistency of the spatial neighborhood to post-process the clustering results, improving the spatial continuity of the spatial domain labels.

[0061] Furthermore, in another embodiment of the present invention, it further includes: The UMAP (Uniform Manifold Approximation and Projection) algorithm was used to perform dimensionality reduction visualization on the refined preliminary spatial domain division labels. The visualization results were then compared with the actual labeled regions of the organization to evaluate the accuracy of spatial domain identification.

[0062] Specifically, the UMAP algorithm is first used to reduce the dimensionality of the globally fused embedded features to two dimensions to obtain the two-dimensional coordinates of each spatial site. Then, different colors are assigned to the sites according to the final spatial domain labels, and embedded feature distribution maps and tissue slice partition maps are drawn respectively. Finally, the actual tissue partition labels are retrieved and superimposed on the map for intuitive comparison, thereby verifying whether the spatial domain division results fit the actual tissue structure and completing the recognition accuracy assessment.

[0063] This invention, through UMAP visualization combined with real-world annotation comparison, can intuitively and quickly verify the discriminative power of the embedded features output by the multi-image fusion strategy of this invention, rapidly locate misclassified clustering regions, and intuitively demonstrate the partitioning advantages brought by the multi-biological prior fusion and noise weighting constraints of this invention, providing intuitive and visual support for model performance verification and parameter tuning.

[0064] The method proposed in this invention will be tested on the DLPFC and HBC datasets to illustrate the method flow and effects of this invention.

[0065] Experiment 1: Spatial Domain Identification of Spatial Transcriptome on the DLPFC Dataset Dataset: 10x Visium DLPFC dataset (151507 slices), with a total of 4226 spots, 33538 genes, and 7 cortical layers labeled.

[0066] Parameter settings: 3000 high-variable genes, preset radius r=300 for spatial proximity graph, k=15 nearest neighbors for KEGG / GO graph KNN algorithm, first hidden layer dimension of graph convolutional encoder is 128, second hidden layer dimension is 64, random inactivation rate is 0.1, learning rate of NAdam optimizer is 0.001, weight decay coefficient is 5×10⁻⁶. -4 The training was conducted for 500 rounds, with loss weights α=1.0 and γ=0.1, cluster number K=7, and refined neighbor number 50.

[0067] The implementation steps are as follows: S101. Read the original gene expression matrix and spot coordinates, screen for highly variable genes, and perform total normalization, logarithmic transformation, and PCA dimensionality reduction preprocessing. S102. Construct a spatial proximity map based on the spatial coordinates of the spot, and simultaneously generate a spatial positive sample pair matrix and a negative sample pair matrix; S103. Use the GSVA package in R language to score the gene set in the KEGG database, generate a pathway activity matrix, and construct a KEGG metabolic pathway similarity map using cosine similarity KNN. S104. Use the GSVA package in R language to score the entire GO gene set, generate the GO activity matrix, and construct the GO functional similarity graph using cosine similarity KNN. S105. Initialize the model spatial domain recognition network and optimize it using the NAdam optimizer; S106, Iterative training for 500 rounds: Forward propagation obtains the global fusion embedding feature Z and ZINB parameters, calculates the ZINB confidence-weighted spatial contrast loss, and combines the ZINB reconstruction loss to obtain the total loss; parameters are updated through backpropagation, and ARI is evaluated by KMeans clustering in each round and the optimal embedding is recorded. S107. Refine the optimal embedding Z by spatial neighbor majority voting and output the final spatial domain label. S108. Perform UMAP dimensionality reduction visualization on the refined embedding and compare it with the ground truth to display the spatial domain recognition results.

[0068] Experiment 2: Spatial Domain Identification of Spatial Transcriptome on HBC Dataset Dataset: Human Breast Cancer 10x Visium dataset, with 3798 spots, 36601 genes, and 20 fine-grained region labels (including tumor region, immune infiltration region, stromal region, etc.).

[0069] Parameter settings: 3000 high-variable genes, preset radius r=550 for spatial proximity graph, k=15 nearest neighbors for KEGG / GO graph KNN algorithm, first hidden layer dimension of graph convolutional encoder is 128, second hidden layer dimension is 64, random inactivation rate is 0.1, learning rate of NAdam optimizer is 0.001, weight decay coefficient is 5×10⁻⁶. -4 The training was conducted for 500 rounds, with loss weights α=1.0 and γ=1.0, cluster number K=20, and refined neighbor number 50.

[0070] Implementation steps: S201. Read the original gene expression matrix and spot coordinates, screen for highly variable genes, and perform total normalization, logarithmic transformation, and PCA dimensionality reduction preprocessing. S202. Construct a radius adjacency graph based on the spatial coordinates of the spot, and simultaneously generate a spatial positive sample pair matrix (graph_nei) and a negative sample pair matrix (graph_neg); S203. Use the GSVA package in R language to score the gene set in the KEGG database, generate a pathway activity matrix, and construct a pathway map using cosine similarity KNN (k=15). S204. Use the GSVA package in R language to score the entire GO gene set, generate the GO activity matrix, and construct the GO functional graph using cosine similarity KNN (k=15). S205. Initialize the model spatial domain recognition network and optimize it using the NAdam optimizer; S206, Iterative training for 500 rounds: Forward propagation obtains the global fusion embedding feature Z and ZINB parameters, calculates the ZINB confidence-weighted spatial contrast loss, and combines it with the ZINB reconstruction loss to obtain the total loss; parameters are updated through backpropagation, and ARI is evaluated by KMeans clustering in each round and the optimal embedding is recorded. S207. Refine the optimal embedding Z by spatial neighbor majority voting and output the final spatial domain label. S208. Perform UMAP dimensionality reduction visualization on the refined embedding and compare it with the ground truth to display the spatial domain recognition results.

[0071] The above datasets demonstrate that the proposed method is effective on 14 spatial transcriptome datasets (12 DLPFC slices, 1 HBC, and 1 MBA), showing a significant improvement in average ARI compared to the baseline model, especially on datasets with high dropout rates.

[0072] In summary, the spatial domain recognition method of multi-graph convolutional networks that integrate gene ontology provided by this invention simultaneously constructs spatial proximity graphs, KEGG metabolic pathway similarity graphs, and GO functional similarity graphs to characterize spatial site similarity relationships from different dimensions, enriching the multi-scale biological information carried by the embeddings. Furthermore, it configures independent graph convolutional encoders for feature extraction. Compared to existing technologies that use single-graph or single-pathway modeling, this method can accurately distinguish spatial domain boundaries caused by subtle differences in cellular function, significantly improving the accuracy of spatial domain recognition. On the other hand, it adaptively fuses the three-way embedding features through a learnable attention fusion mechanism, avoiding information redundancy and interference, and solving the defects of fixed fusion and poor generalization of multi-view methods.

[0073] Furthermore, this method incorporates a ZINB decoder in the spatial domain recognition network that participates in computation only during model training and is disabled during inference. Combined with confidence-weighted spatial contrastive loss training logic, and based on the site-specific missing probability quantification of reliability, it achieves automatic deweighting of low-quality noise sites. Ablation experiments show that removing the ZINB confidence weighting reduces the ARI from 0.631 to 0.587 (a 7.0% decrease) on the DLPFC 151510 dataset and from 0.574 to 0.519 (a 9.6% decrease) on the 151670 dataset, exhibiting significant performance degradation across multiple DLPFC slices.

[0074] Based on the same inventive concept, a second aspect of this invention also provides a spatial domain recognition device that integrates a multi-graph convolutional network based on gene ontology. See also... Figure 3 , Figure 3 This is a structural block diagram of a spatial domain recognition device for a multi-graph convolutional network that integrates gene ontology, provided in an embodiment of the present invention. The device includes: The preprocessing module is used to acquire the original spatial transcriptome data, which includes gene expression matrices and two-dimensional spatial coordinates of each spatial site, and to preprocess the original data to obtain a unified node feature vector shared by all spatial sites. The graph construction module is used to simultaneously construct multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph, using spatial sites as nodes and a unified node feature vector as node features. The inference module is equipped with a pre-trained spatial domain recognition network, which includes multiple graph convolutional encoders with identical structures and independent parameters, and an attention fusion module. The multiple graph convolutional encoders are used to perform message passing on multiple independent heterogeneous graphs, and output multi-path embedding features accordingly. The attention fusion module is used to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, and output a globally fused embedding feature. The clustering and identification module is used to perform clustering processing on the globally fused embedded features to obtain preliminary spatial domain division labels; the preliminary spatial domain division labels are refined using a neighbor majority voting strategy to obtain spatial domain identification results.

[0075] It should be noted that, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. Therefore, all embodiments of the spatial domain recognition method of multi-graph convolutional networks that integrate gene ontology are applicable to this device and can achieve the same or similar beneficial effects.

[0076] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A spatial domain recognition method for multi-graph convolutional networks integrating gene ontology, characterized in that, Includes the following steps: Obtain raw spatial transcriptome data, including gene expression matrices and two-dimensional spatial coordinates for each spatial site, and preprocess the raw data to obtain a unified node feature vector shared by all spatial sites. Using the spatial sites as nodes and the unified node feature vector as node features, multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph, are constructed simultaneously. The multiple independent heterogeneous graphs are input into a pre-trained spatial domain recognition network. First, multiple graph convolutional encoders with the same structure and independent parameters are used to perform message passing on the multiple independent heterogeneous graphs respectively, and output multi-path embedded features accordingly. The attention fusion module is then used to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, and output the globally fused embedding features. Clustering is performed on the globally fused embedded features to obtain preliminary spatial domain partitioning labels; The preliminary spatial domain division labels are refined using a neighbor majority voting strategy to obtain spatial domain identification results.

2. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, Preprocessing the raw data includes: The gene expression matrix in the original data is subjected to high-variability gene screening, total gene normalization, logarithmic transformation, and PCA dimensionality reduction in sequence.

3. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, The spatial proximity map, the KEGG metabolic pathway similarity map, and the GO functional similarity map were constructed using the following methods: The Euclidean distance between each spatial point is calculated based on the two-dimensional spatial coordinates of each spatial point, and undirected adjacent edges are established for spatial points whose Euclidean distance is less than the preset radius to obtain a spatial proximity graph. Gene set variation analysis tools were used to score the gene set in the KEGG database to obtain a pathway activity score matrix for each spatial site; based on the cosine similarity between the row vectors of the pathway activity score matrix, the k-nearest neighbor algorithm was used to establish pathway similarity adjacency edges to obtain a KEGG metabolic pathway similarity graph. The GO gene set was scored using a gene set variation analysis tool to obtain a GO activity score matrix for each spatial locus; Based on the cosine similarity between the row vectors of the GO activity score matrix, the k-nearest neighbor algorithm is used to establish gene functional similarity adjacency edges to obtain the GO functional similarity graph.

4. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, Each graph convolutional encoder includes two graph convolutional layers, each followed by a ReLU activation function; multiple sets of graph convolutional encoders with identical structures and independent parameters are used to perform message passing on multiple independent heterogeneous graphs, correspondingly outputting multi-path embedded features, including: Multiple independent heterogeneous graphs are input together with the unified node feature vector into the corresponding graph convolutional encoder. Feature extraction is performed sequentially using two graph convolutional layers, and nonlinear feature mapping is completed through the ReLU activation function, outputting multi-path embedded features with unified dimensions.

5. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, The multi-path embedding features are adaptively integrated using an attention fusion module based on a learnable attention fusion mechanism, outputting a globally fused embedding feature, including: After stacking the multipath embedding features along the dimensions, the attention score of the multipath embedding features is calculated for each spatial site using a Linear-Tanh-Linear structure, and the attention weight is obtained by Softmax normalization. The multipath embedding features are then weighted and summed using the attention weights, and then mapped by MLP to generate a global fusion embedding feature that integrates spatial location, metabolic pathway and gene function.

6. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, The spatial domain recognition network includes multiple graph convolutional encoders, the attention fusion module, and a ZINB decoder that only participates in computation during the model training phase and is not enabled during the inference phase. The ZINB decoder includes a fully connected layer, a BatchNorm layer, and a ReLU activation function; The fully connected layer is used to receive the globally fused embedding features output by the attention fusion module and complete the dimension mapping to obtain the dimension-mapped features. The BatchNorm layer contains learnable scaling and offset coefficients, which are used to perform standardization on the features after the dimension mapping to stabilize the model training gradient, accelerate network convergence, and obtain standardized features. The ReLU activation function is used to perform a non-linear mapping on the standardized features to obtain the non-linearly mapped features. The ZINB decoder is also used to predict the technical missing probability of each gene at each spatial site based on the features after nonlinear mapping.

7. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 6, characterized in that, The spatial domain recognition network is trained through the following steps: The global fusion embedding features are input into the ZINB decoder to predict the probability of technology missing for each gene at each spatial site. The reliability confidence level of each spatial site is calculated based on the probability of the missing technology described above; Based on the aforementioned reliability confidence level, a confidence-weighted spatial contrast loss is constructed, and combined with the reconstruction loss of the ZINB decoder, a total loss is constructed. Based on the total loss, the learnable parameters of the graph convolutional encoder, the attention fusion module, and the ZINB decoder are iteratively updated using backpropagation of the optimizer to iteratively train the spatial domain recognition network and obtain a pre-trained spatial domain recognition network.

8. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 7, characterized in that, The formula for calculating the reliability confidence level of each spatial location is as follows: In the formula, Indicates the first i Reliability confidence level of each spatial location, Indicates the total number of genes. The first predicted by the ZINB decoder The spatial location of the first The probability of technical missing genes; The expression for the confidence-weighted spatial contrast loss is: In the formula, This represents the confidence-weighted spatial contrast loss. This represents the total number of spatial locations. Indicates based on the first The contrastive loss is calculated by the embedding similarity between each spatial location and its set of spatial neighbors and non-neighbors; The expression for the total loss is: In the formula, Indicates the total loss. This represents the reconstruction loss of the ZINB decoder. , This is a hyperparameter.

9. The spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology according to claim 1, characterized in that, Also includes: The UMAP algorithm was used to perform dimensionality reduction visualization on the refined preliminary spatial domain division labels, and the visualization results were compared with the actual labeled areas of the organization to evaluate the accuracy of spatial domain identification.

10. A spatial domain recognition device for multi-graph convolutional networks incorporating gene ontology, used to implement the spatial domain recognition method for multi-graph convolutional networks incorporating gene ontology as described in any one of claims 1-9, characterized in that, The device includes: The preprocessing module is used to acquire the original spatial transcriptome data, which includes gene expression matrices and two-dimensional spatial coordinates of each spatial site, and to preprocess the original data to obtain a unified node feature vector shared by all spatial sites. The graph construction module is used to simultaneously construct multiple independent heterogeneous graphs, including a spatial proximity graph, a KEGG metabolic pathway similarity graph, and a GO functional similarity graph, using the spatial site as a node and the unified node feature vector as a node feature. The inference module is configured with a pre-trained spatial domain recognition network, which includes multiple graph convolutional encoders with identical structures and independent parameters, and an attention fusion module. The multiple graph convolutional encoders are used to perform message passing on the multiple independent heterogeneous graphs, and output multi-path embedding features accordingly. The attention fusion module is used to adaptively integrate the multi-path embedding features based on a learnable attention fusion mechanism, and output a globally fused embedding feature. The clustering and identification module is used to perform clustering processing on the global fusion embedded features to obtain preliminary spatial domain division labels; and to refine the preliminary spatial domain division labels using a neighbor majority voting strategy to obtain spatial domain identification results.

Images