Spatial domain resolution methods, apparatus, equipment and storage media

Abstract

This application relates to a spatial domain resolution method, apparatus, device, and storage medium. The method includes: constructing a physical spatial map feature matrix, a correlation map feature matrix, and a spatial feature matrix based on the spatial transcription data of the target sequencing object; sequentially performing gated weighted fusion on each matrix to obtain a target latent representation matrix; performing linear spatial reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix; performing gated weighted fusion on the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain a target optimized expression matrix; and based on the target optimized expression matrix, performing spatial clustering on each spatial transcription sequencing point of the target sequencing object to obtain the spatial domain resolution result. The technical solution of this application can accurately identify dispersed and small discontinuous spatial domains while preventing excessive feature smoothing or excessive noise retention.

Description

Technical Field

[0001] This application relates to the field of spatial transcriptomics technology, and in particular to a spatial domain resolution method, apparatus, device, and storage medium. Background Technology

[0002] In recent years, spatial transcriptomics technology has made groundbreaking progress. After acquiring spatial transcriptomics data, the primary and most critical downstream computational analysis task is spatial domain resolution, or spatial clustering. A spatial domain refers to a cell population or microenvironment region that has a specific distribution pattern in physical space and exhibits consistency in gene expression. Accurate spatial domain delineation is fundamental for subsequent cell communication inference, spatial developmental trajectory construction, and differential gene analysis.

[0003] The core idea of ​​existing spatial domain parsing algorithms is based on the fundamental assumption of "spatial continuity," meaning that physically adjacent sites tend to have similar cell types or microenvironment states. Mainstream algorithmic frameworks include graph neural network architectures such as STAGATE and GraphST, and probabilistic graph model architectures such as BayesSpace and BASS. These algorithms heavily rely on physical distance, failing to overcome physical limitations and resulting in feature loss and inaccurate identification of discontinuous domains. Furthermore, they lack an adaptive fusion mechanism between local spatial and global semantics, leading to rigid feature extraction. This makes it difficult to balance noise smoothing and specific signal preservation during clustering, hindering the accurate reconstruction of complex tissue heterogeneity. Summary of the Invention

[0004] This application provides a spatial domain resolution method, apparatus, device, and storage medium, with the aim of achieving connection of distant similar sites across physical distances and adaptive feature fusion in spatial domain resolution, accurately identifying dispersed and small discontinuous spatial domains such as tumor immune infiltration areas and microvascular networks, while preventing excessive feature smoothing or excessive noise retention.

[0005] In a first aspect, embodiments of this application provide a spatial domain resolution method, including:

[0006] Based on the spatial transcription data of the target sequencing object, construct the physical spatial map feature matrix, the correlation map feature matrix, and the spatial feature matrix of the target sequencing object;

[0007] The physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix are sequentially subjected to gated weighted fusion to obtain the target latent representation matrix.

[0008] The target latent representation matrix is ​​subjected to linear space reconstruction and nonlinear neural decoding to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, respectively.

[0009] The linear reconstruction matrix and the nonlinear reconstruction matrix are gated and weighted to obtain the target optimized expression matrix;

[0010] Based on the target optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain spatial domain resolution results.

[0011] Optionally, constructing the physical spatial map feature matrix, correlation map feature matrix, and spatial feature matrix of the target sequencing object based on its spatial transcription data includes:

[0012] Obtain the highly variable gene expression matrix, sequencing point spatial location matrix, and sequencing point spatial adjacency matrix of the target sequencing object;

[0013] The highly variable gene expression matrix is ​​projected into a latent space feature matrix;

[0014] Graph convolution is performed on the spatial adjacency matrix of the sequencing points and the latent space feature matrix to construct the physical space graph feature matrix;

[0015] The correlation adjacency matrix of sequencing points is obtained based on the latent space feature matrix;

[0016] Perform graph convolution operation on the sequencing point correlation adjacency matrix and the latent space feature matrix to construct the correlation graph feature matrix;

[0017] The spatial feature matrix is ​​constructed based on the spatial location matrix of the sequencing points.

[0018] Optionally, obtaining the sequencing point correlation adjacency matrix based on the latent space feature matrix includes:

[0019] Based on the latent space feature matrix, the correlation coefficient matrix of each of the spatial transcription sequencing points is calculated;

[0020] Based on a preset correlation coefficient threshold, the correlation coefficient matrix is ​​subjected to target correlation edge filtering to obtain the sequencing point correlation adjacency matrix.

[0021] Optionally, the step of sequentially performing gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix to obtain the target latent representation matrix includes:

[0022] A pre-trained first gating network is used to perform the gated weighted fusion of the physical space graph feature matrix and the correlation graph feature matrix to obtain a dual-graph fusion feature matrix.

[0023] A pre-trained second gating network is used to perform the gated weighted fusion of the dual-graph fusion feature matrix and the spatial feature matrix to obtain the target latent representation matrix.

[0024] Optionally, performing linear space reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix includes:

[0025] Project the target latent representation matrix into the original spatial feature matrix;

[0026] Graph convolution is performed on the spatial adjacency matrix of the sequencing points and the original spatial feature matrix to construct the linear reconstruction matrix;

[0027] The target latent representation matrix is ​​mapped to the nonlinear reconstruction matrix using a sinusoidal activation function.

[0028] Optionally, the step of gated weighted fusion of the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix includes:

[0029] The linear and nonlinear weight parameters of the linear and nonlinear reconstruction matrices in the gated weighted fusion are obtained using the following formulas:

[0030] ,

[0031] in, and These represent the linear weight parameter and the nonlinear weight parameter, respectively. and Let the linear reconstruction matrix and the nonlinear reconstruction matrix represent respectively. This represents a matrix concatenation operation. This refers to a pre-trained multilayer perceptron. This represents the Sigmoid activation function. This represents a matrix partitioning operation;

[0032] The target optimized expression matrix is ​​obtained by performing a weighted fusion operation on the linear reconstruction matrix and the nonlinear reconstruction matrix based on the linear weight parameters and the nonlinear weight parameters using the following formula:

[0033] ,

[0034] in, This represents the target optimization expression matrix. This refers to a pre-trained fully connected layer. It is a matrix dot product operation, which means multiplying corresponding elements of a matrix.

[0035] Optionally, before performing gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix to obtain the target latent representation matrix, the method further includes:

[0036] Joint parameter training is performed on the gated network parameters based on the mean square error loss function:

[0037] ,

[0038] in, This represents the original gene expression matrix in the spatial transcription data. and This indicates the preset balance coefficient.

[0039] Secondly, embodiments of this application provide a spatial domain resolution apparatus, comprising:

[0040] The feature construction module is used to construct the physical spatial map feature matrix, correlation map feature matrix, and spatial feature matrix of the target sequencing object based on the spatial transcription data of the target sequencing object;

[0041] The feature fusion module is used to sequentially perform gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix and the spatial feature matrix to obtain the target latent representation matrix;

[0042] The decoding and reconstruction module is used to perform linear space reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, respectively.

[0043] The reconstruction and fusion module is used to perform gated weighted fusion of the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix;

[0044] The spatial clustering module is used to perform spatial clustering of each spatial transcription sequencing point of the target sequencing object based on the target optimized expression matrix, and obtain spatial domain resolution results.

[0045] Thirdly, embodiments of this application provide a spatial domain resolution device, including:

[0046] One or more processors;

[0047] Memory, used to store one or more programs;

[0048] When the one or more programs are executed by the one or more processors, the one or more processors implement the spatial domain resolution method provided in any embodiment of this application.

[0049] Fourthly, embodiments of this application provide a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the spatial domain resolution method provided in any embodiment of this application.

[0050] This application provides a spatial domain resolution method, apparatus, device, and storage medium. Based on the spatial transcription data of the target sequencing object, it constructs a physical spatial graph feature matrix, a correlation graph feature matrix, and a spatial feature matrix. Gated weighted fusion is then performed between these matrices sequentially to obtain a target latent representation matrix. Furthermore, linear spatial reconstruction and nonlinear neural decoding are performed on the target latent representation matrix to obtain linear and nonlinear reconstruction matrices, which are then gated weighted and fused to obtain a target optimized expression matrix. Based on this optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain the spatial domain resolution result. This solves the problems of existing technologies where feature recognition highly depends on spatial continuity and feature fusion methods are rigid. It achieves connectivity of distant similar sites across physical distances and adaptive feature fusion in spatial domain resolution, accurately identifying dispersed and small discontinuous spatial domains such as tumor immune infiltration areas and microvascular networks, while preventing excessive feature smoothing or excessive noise retention. Attached Figure Description

[0051] Figure 1 A flowchart of a spatial domain resolution method provided in Embodiment 1 of this application;

[0052] Figure 2 A flowchart of a spatial domain resolution method provided in Embodiment 1 of this application;

[0053] Figure 3 This is a schematic diagram of the structure of a spatial domain resolution device provided in Embodiment 2 of this application;

[0054] Figure 4 This is a schematic diagram of the structure of a spatial domain resolution device provided in Embodiment 3 of this application. Detailed Implementation

[0055] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present application, not the entire structure.

[0056] Example 1

[0057] Figure 1This is a flowchart of a spatial domain resolution method provided in Embodiment 1 of this application. This embodiment is applicable to the case of spatial domain resolution of sequencing objects of spatial transcriptome. The method can be executed by a spatial domain resolution device, which can be implemented by hardware and / or software and is generally integrated into an electronic device, such as a computer device. The method specifically includes:

[0058] Step 110: Based on the spatial transcription data of the target sequencing object, construct the physical spatial map feature matrix, correlation map feature matrix, and spatial feature matrix of the target sequencing object.

[0059] The target sequencing object can refer to a biological tissue slice sample sequenced using spatial transcriptomics technology. The spatial transcriptome data can refer to the data obtained from spatial transcriptome sequencing that contains gene expression information and spatial location information of the sequencing points. It can typically include the spatial transcriptome expression matrix and spatial location matrix of each sequencing point of the target sequencing object.

[0060] Specifically, spatial transcription data can include a given target sequencing object containing sequencing sites and The original spatial transcriptome expression matrix of each gene, wherein, and They can be positive integers, This represents the total number of sequencing sites. This represents the total number of genes originally detected; optionally, further screening can be performed on the original spatial transcriptome expression matrix to identify the top-ranked genes. A number of highly variable genes were obtained, and logarithmic normalization and scaling were performed to obtain... sequencing sites and Highly variable gene expression matrix of 1 highly variable gene ,in, For a high number of variable genes, a setting of 3000 is preferred. Simultaneously, the two-dimensional spatial coordinates of each sequencing point are obtained. Spatial location matrix of sequencing sites .

[0061] Furthermore, the physical spatial map feature matrix can be a matrix characterizing gene expression features associated with local spatial topology between sequencing points, and can be obtained by fusing the features and associations of spatial transcriptome expression information and spatial location information of each sequencing point in spatial transcription data. The correlation map feature matrix can be a feature matrix characterizing the global expression similarity of sequencing points, and can be obtained by fusing the features and associations of spatial transcriptome expression information between sequencing points in spatial transcription data. The spatial feature matrix can be a matrix characterizing the detailed features of the spatial location of sequencing points, and can be obtained by fusing the features and associations of spatial location information between sequencing points in spatial transcription data.

[0062] In one optional implementation, based on the spatial transcription data of the target sequencing object, a physical spatial graph feature matrix, a correlation graph feature matrix, and a spatial feature matrix are constructed, including: obtaining the highly variable gene expression matrix, the sequencing point spatial location matrix, and the sequencing point spatial adjacency matrix of the target sequencing object; projecting the highly variable gene expression matrix into a latent spatial feature matrix; performing graph convolution operations on the sequencing point spatial adjacency matrix and the latent spatial feature matrix to construct the physical spatial graph feature matrix; obtaining the sequencing point correlation adjacency matrix based on the latent spatial feature matrix; performing graph convolution operations on the sequencing point correlation adjacency matrix and the latent spatial feature matrix to construct the correlation graph feature matrix; and constructing the spatial feature matrix based on the sequencing point spatial location matrix.

[0063] The sequencing point spatial adjacency matrix can be a matrix representing the spatial proximity relationship between sequencing points, or it can be obtained from the sequencing point spatial location matrix. Specifically, it can be based on the two-dimensional spatial coordinates in the sequencing point spatial location matrix. The K nearest neighbors of each site are calculated using Euclidean distance to construct the initial physical adjacency matrix. The sequencing point spatial adjacency matrix was further obtained by symmetric normalization. ,in, It is a degree matrix.

[0064] Furthermore, the latent space feature matrix can be a low-dimensional latent space feature matrix obtained by linearly projecting a high-dimensional, highly variable gene expression matrix. Specifically, optionally, the input highly variable gene expression matrix can be... By projecting the linear layer onto the low-dimensional latent space, the latent space feature matrix is ​​obtained. ,in, This is the pre-learned linear layer projection weight matrix. Subsequently, the spatial adjacency matrix of sequencing points is... and latent space feature matrix Perform graph convolution operations to construct the physical space graph feature matrix. Specifically, it can be expressed as Through the above-mentioned initial feature projection and physical graph convolution process, the sequencing points that are physically adjacent can transfer and smooth gene expression features, thereby achieving the fusion of spatial topology and gene expression features.

[0065] Furthermore, based on the latent space feature matrix, the gene expression feature associations between sequencing sites can be obtained, thus obtaining the sequencing site correlation adjacency matrix, which can be separated from the physical spatial position relationship between sequencing sites and only reflects their gene expression similarity.

[0066] In one optional implementation, obtaining the sequencing point relevance adjacency matrix based on the latent space feature matrix includes: calculating the correlation coefficient matrix of each spatial transcription sequencing point based on the latent space feature matrix; and performing target relevance edge screening on the correlation coefficient matrix according to a preset correlation coefficient threshold to obtain the sequencing point relevance adjacency matrix.

[0067] The correlation coefficient matrix can be a matrix representing the degree of gene expression similarity between any two sequencing sites. The preset correlation coefficient threshold can be a critical value used to screen for strongly correlated connections between sequencing sites. Target correlation edge filtering can be a threshold filtering operation on the elements of the correlation coefficient matrix, which can be used to filter out negative and weakly correlated noise, retaining only strongly positively correlated edges; it can also include normalizing the filtering results. Therefore, the resulting sequencing site correlation adjacency matrix can represent sequencing site associations determined solely by gene expression similarity.

[0068] Specifically, the correlation coefficient matrix for each spatial transcription sequencing point can be calculated based on the latent space feature matrix. This can include dynamically calculating the Pearson correlation coefficient matrix based on the current latent space feature matrix. Specifically, it can be calculated using the following formula:

[0069]

[0070]

[0071]

[0072] in, The latent space feature matrix represents the centralized feature matrix. This represents the vector obtained by averaging the latent space feature matrix across each latent feature dimension. This represents the result of normalizing the centered latent space feature matrix. Based on a preset correlation coefficient threshold... For the Pearson correlation coefficient matrix The target-related edge filtering process is performed to obtain the sequencing point relevance adjacency matrix. Then it can be expressed as:

[0073]

[0074] Among them, the preset correlation coefficient threshold The optimal value can be set to 0.75.

[0075] Furthermore, the correlation adjacency matrix of sequencing sites... and latent space feature matrix Graph convolution operations can be performed to construct a feature matrix of the correlation graph. Specifically, it can be expressed as Through the above correlation map construction process, gene expression similarity can be used as a weighting basis to globally weight and fuse gene expression features without being constrained by the spatial location of sequencing points, enabling physically discontinuous but similarly expressed sequencing points to transfer feature information.

[0076] Furthermore, a spatial feature matrix can be constructed based on the spatial location matrix of sequencing points. Optionally, to further refine the spatial boundaries, more accurately capture the nonlinear structure of space, and avoid excessive smoothing of the graph structure, an implicit neural representation module composed of multiple SineLayer layers can be introduced to map the original two-dimensional spatial coordinates to a high-dimensional feature space. Unlike the traditional ReLU (Rectified Linear Unit) activation function, the sinusoidal activation function can excellently preserve high-frequency details in the spatial coordinates. Specifically, the spatial coordinate matrix can be... Input by A spatial coordinate encoder composed of layers of sinusoidal networks obtains the spatial feature matrix. Specifically, it can be expressed as:

[0077]

[0078] in, Represents the sine function. , , and This represents the pre-learned spatial coordinate encoding weight matrix. The resulting spatial feature matrix, purely based on coordinate frequency information, is the encoded version. It can operate completely independently of the neighbor aggregation mechanism.

[0079] Step 120: Perform gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix in sequence to obtain the target latent representation matrix.

[0080] Gated weighted fusion can be an operation that assigns adaptive weights to multiple features and performs dynamic weighted fusion. The target latent representation matrix can be a feature matrix that integrates local spatial information of sequencing points, global expression similarity information, and spatial location details.

[0081] Specifically, through gated weighted fusion, the local spatial information and global expression similarity information represented by the physical space map feature matrix and the correlation map feature matrix can be adaptively weighted and dynamically weighted. Furthermore, the above-mentioned fused features and the spatial location details represented by the spatial feature matrix can be cross-modal adaptively weighted and fused to finally obtain the target latent representation matrix, which can then integrate three complementary types of information: spatial topology, gene expression similarity, and high-frequency location details.

[0082] In one optional implementation, the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix are sequentially subjected to gated weighted fusion to obtain the target latent representation matrix. This includes: using a pre-trained first gating network to perform gated weighted fusion of the physical space graph feature matrix and the correlation graph feature matrix to obtain a dual-graph fusion feature matrix; and using a pre-trained second gating network to perform gated weighted fusion of the dual-graph fusion feature matrix and the spatial feature matrix to obtain the target latent representation matrix.

[0083] The first pre-trained gating network can be a gating network containing pre-learned weight parameters. This first gating network can configure the weights of the physical spatial graph feature matrix and the correlation graph feature matrix in the fusion based on the weight parameters, so as to optimally allocate the contribution ratio of local spatial information and global related information of each sequencing point, so that the fused dual-graph fusion feature matrix can achieve a dynamic balance between spatial continuity constraints and cross-spatial semantic similarity.

[0084] Specifically, a pre-trained first gating network is used to perform gated weighted fusion of the physical space graph feature matrix and the correlation graph feature matrix to obtain a dual-graph fusion feature matrix. This may include: calculating the first gating weight vector. ,in, The first gating weight matrix is ​​learned in advance. The Sigmoid activation function is used; the resulting first gate weight vector Segmented into physical space graph weights Correlation plot weights The two graph features are weighted element-wise and then fused to obtain the dual-graph fused feature matrix. Specifically, it can be expressed as ,in, The first fusion weight matrix is ​​learned in advance. It is a matrix dot product operation, which means multiplying corresponding elements of a matrix.

[0085] Furthermore, the pre-trained second gating network can also contain pre-learned weight parameters. This second gating network configures the weights of the dual-graph fusion feature matrix and the spatial feature matrix in the fusion based on the weight parameters, so as to optimally allocate the contribution ratio of graph aggregation features and high-frequency details of spatial coordinates of each sequencing point. This allows the weight of spatial smoothing to be reduced in discontinuous regions with strong gene expression signal features, so that the target latent representation matrix obtained by fusion can achieve a dynamic balance between noise smoothing and local specificity preservation.

[0086] Specifically, a pre-trained second gating network is used to perform gated weighted fusion of the dual-image fusion feature matrix and the spatial feature matrix to obtain the target latent representation matrix. This may include: calculating the second gating weight vector. ,in, This represents the pre-learned second gating weight matrix. These are the pre-learned bias coefficients. This represents the concatenated dual-image fusion feature matrix and spatial feature matrix; the resulting second gating weight vector... Segmentation into graph fusion feature weights Spatial feature weights The target latent representation matrix is ​​obtained by weighted fusion calculation. ,in, This represents the pre-learned second fusion weight matrix. It is a matrix dot product operation, which means multiplying corresponding elements of a matrix.

[0087] Step 130: Perform linear space reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain the linear reconstruction matrix and the nonlinear reconstruction matrix.

[0088] Linear spatial reconstruction can be based on graph convolution and dimensionality restoration of latent features using a physical adjacency graph, used to enhance the continuity of the gene expression spatial domain. Nonlinear neural decoding can be a nonlinear feature mapping operation used to capture complex nonlinear patterns in gene expression. Therefore, by performing linear spatial reconstruction on the target latent representation matrix, the resulting linear reconstruction matrix can represent gene expression reconstruction components that characterize continuous spatial structure and smooth noise. Conversely, by performing nonlinear neural decoding on the target latent representation matrix, the resulting nonlinear reconstruction matrix can represent gene expression reconstruction components that retain high-frequency details and specific signals corresponding to small discontinuities.

[0089] In one optional implementation, the target latent representation matrix is ​​subjected to linear spatial reconstruction and nonlinear neural decoding to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, including: projecting the target latent representation matrix into the original spatial feature matrix; performing graph convolution operation on the sequencing point spatial adjacency matrix and the original spatial feature matrix to construct a linear reconstruction matrix; and using a sinusoidal activation function to map the target latent representation matrix into a nonlinear reconstruction matrix.

[0090] The original spatial feature matrix can be the feature matrix projected from the low-dimensional target latent representation matrix to the original high-dimensional gene-dimensional feature matrix. This is relevant to the spatial adjacency matrix of sequencing points. and the original spatial feature matrix Graph convolution operations can construct linear reconstruction matrices. Specifically, it can be expressed as ,in, This represents the pre-learned reconstruction weight matrix. The linear spatial reconstruction path described above utilizes the physical spatial graph for neighborhood smoothing, and can explicitly introduce spatial nearest neighbor constraints during the reconstruction process to enhance the continuity of the spatial domain and reduce noise.

[0091] Furthermore, a sinusoidal activation function is used to transform the target latent representation matrix. Mapped to a nonlinear reconstruction matrix Specifically, it can be expressed as ,in, Represents the sinusoidal activation function. This represents a fully connected dense layer. The above implementation uses a neural network structure based on a fully connected dense layer and incorporating a sinusoidal nonlinear mapping to implicitly fit the complex nonlinear relationship between gene expression characteristics and spatial distribution. By utilizing the high-frequency representation capability of the sinusoidal activation layer, it accurately preserves the expression details of small, discontinuous spatial domains, thus achieving the preservation of fine features of discontinuous spatial domains.

[0092] Step 140: Perform gated weighted fusion of the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix.

[0093] The target optimized expression matrix can be the optimal gene expression matrix that balances noise smoothing and specificity preservation, obtained through adaptive fusion of linear and nonlinear dual paths. Specifically, by gating and weighting the linear and nonlinear reconstruction matrices, and using pre-learned weights, the resulting target optimized expression matrix can achieve an adaptive balance between continuous spatial smoothing features and discontinuous high-frequency detail features.

[0094] In an optional implementation, the linear reconstruction matrix and the nonlinear reconstruction matrix are gated and weighted to obtain the target optimization expression matrix, including:

[0095] The linear reconstruction matrix is ​​obtained using the following formulas. and nonlinear reconstruction matrix Linear weight parameters in gated weighted fusion and nonlinear weight parameters :

[0096]

[0097] in, This represents a matrix concatenation operation. This refers to a pre-trained multilayer perceptron. This represents a matrix partitioning operation; the following formula is used, based on linear weight parameters. and nonlinear weight parameters For linear reconstruction matrix and nonlinear reconstruction matrix Perform weighted fusion operations to obtain the target optimized expression matrix. :

[0098]

[0099] in, This refers to a pre-trained fully connected layer. This is a matrix dot product operation, which represents the multiplication of corresponding elements in a matrix. The above implementation will linearly reconstruct the matrix. and nonlinear reconstruction matrix The parts are assembled and fed into a multilayer sensor. Perform multi-layer nonlinear feature transformation to mine the correlation between the features expressed by the two matrices and learn the fusion weight parameters; through Operation will The learned weights are divided into segments corresponding to the linear reconstruction matrix. and nonlinear reconstruction matrix The weighted branches are mapped to the 0-1 interval using the Sigmoid activation function to generate linear weight parameters. and nonlinear weight parameters This allows for the weighted fusion of the two matrices, ultimately yielding the target optimization representation matrix. .

[0100] Furthermore, in an optional implementation, a training method for each learnable parameter in this embodiment is also provided. Specifically, before sequentially performing gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix to obtain the target latent representation matrix, the method may further include: performing joint parameter training on the gated network parameters obtained by the gated weighted fusion based on the mean squared error loss function.

[0101]

[0102] in, This represents the original gene expression matrix in the spatial transcription data. and This represents the preset balance coefficient. Specifically, the gated network parameters can be learnable parameters in gated weighted fusion; further optionally, other learnable parameters can be jointly trained with the gated network parameters based on the aforementioned loss function. Specifically, an end-to-end training method can be adopted, using the aforementioned loss function. To optimize the objective, all learnable parameters are iteratively updated during training, gradually reducing the original gene expression matrix. With reconstruction matrix , The error between; when the loss function When the convergence reaches its minimum, all learnable parameters are optimal and can be used to obtain the final reconstruction result. This is to ensure that it most accurately reflects the organization's true spatial domain structure, which can then be used for subsequent spatial clustering.

[0103] Step 150: Based on the target optimized expression matrix, perform spatial clustering on each spatial transcription sequencing point of the target sequencing object to obtain the spatial domain resolution results.

[0104] The spatial domain resolution result can be the identification result of the tissue spatial domain formed after assigning spatial domain category labels to each sequencing point. Specifically, based on the target optimized expression matrix, the global expression similarity and spatial structure association of each spatial transcription sequencing point can be taken into account, and spatial clustering can be performed. A corresponding spatial domain category label is assigned to each sequencing point to obtain the spatial domain resolution result, thereby dividing the tissue slice into spatially continuous, gene-expression-consistent, and anatomically consistent spatial domains, achieving accurate spatial domain resolution. Specifically, when the number of clusters is known, the mclust algorithm is preferred for spatial clustering; when the number of clusters is uncertain, the Louvain community retrieval algorithm is preferred for spatial clustering.

[0105] For example, Figure 2 This is a flowchart illustrating a spatial domain resolution method provided in Embodiment 1 of this application. Figure 2 As shown, this spatial domain parsing method performs dimensionality reduction projection on the original gene expression features in the encoding stage, and achieves adaptive extraction and fusion of local spatial information, global semantic information and high-frequency coordinate information through dual-graph convolution, spatial coordinate encoding and two-level gating fusion to obtain low-dimensional feature representation. Furthermore, in the decoding stage, the low-dimensional features are reconstructed into a high-dimensional gene expression matrix through dual-path decoding and gating fusion mechanism. In the reconstruction process, both the spatial smoothness of continuous tissue regions and the preservation of details of discontinuous small domains are taken into account.

[0106] The technical solution of this embodiment constructs a physical spatial map feature matrix, a correlation map feature matrix, and a spatial feature matrix based on the spatial transcription data of the target sequencing object. Then, gated weighted fusion is performed between these matrices to obtain the target latent representation matrix. The target latent representation matrix is ​​then subjected to linear spatial reconstruction and nonlinear neural decoding to obtain linear reconstruction matrices and nonlinear reconstruction matrices, which are then gated weighted fused to obtain the target optimized expression matrix. Based on this optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain the spatial domain resolution result. This solves the problems of existing technologies where feature recognition highly depends on spatial continuity and feature fusion methods are rigid. It achieves connectivity of distant similar sites across physical distances and adaptive feature fusion in spatial domain resolution, accurately identifying dispersed and small discontinuous spatial domains such as tumor immune infiltration areas and microvascular networks, while preventing excessive feature smoothing or excessive noise retention.

[0107] Furthermore, to illustrate the beneficial effects of the spatial domain resolution method in this embodiment, this embodiment also provides comparative experimental results with existing technologies. Specifically, existing technologies include: BayesSpace, BASS, BANKSY, STAGATE, GraphST, SEDR, SCAN-IT, CCST, STMIGCL, stDCL, StDSGCL, and SAGE spatial clustering methods. The clustering metrics used to evaluate the performance of each spatial clustering method in the experiments include: ARI (Adjusted Rand Index) and NMI (Normalized Mutual Information), both ranging from [0-1], with larger values ​​indicating better performance. Two spatial transcriptome sequencing datasets with relatively small, non-contiguous spatial domains were collected, and two experimental results analyses were performed. Specifically, Experiment 1 collected data from the 10X Visium spatial transcriptome sequencing platform Mouse Liver, with data samples from mouse liver tissue; the original complete dataset contained a total of 8 samples. Table 1 shows detailed information on the eight datasets of Mouse Liver; Tables 2 and 3 show the results of the eight datasets of Mouse Liver on the evaluation metrics ARI and NMI, respectively.

[0108] Table 1

[0109]

[0110] Table 2

[0111]

[0112] Table 3

[0113]

[0114] The results above show that this application demonstrates improved performance compared to the aforementioned prior art, with average improvements of 18.45% and 11.20% in ARI and NMI, respectively. Furthermore, Experiment 2 collected data from the 10XVisium spatial transcriptome sequencing platform, Human Liver, with data samples derived from human liver tissue. The original complete dataset contained a total of 5 samples. Table 4 presents detailed information on the 5 datasets from Human Liver; Tables 5 and 6 respectively show the results of the 5 datasets from Human Liver on the evaluation metrics.

[0115] Table 4

[0116]

[0117] Table 5

[0118]

[0119] Table 6

[0120]

[0121] As can be seen from the above results, compared with the existing technologies, the performance of this application is improved, with an average improvement of 10.56% in ARI and 8.86% in NMI. Therefore, based on the above comparison results, the spatial domain resolution method proposed in this application achieves better performance and significantly improves the recognition of small, discontinuous spatial domains.

[0122] Example 2

[0123] Figure 3 This is a schematic diagram of the structure of a spatial domain resolution device provided in Embodiment 2 of this application, as shown below. Figure 3 As shown, the spatial domain resolution device includes: a feature construction module 310, a feature fusion module 320, a decoding and reconstruction module 330, a reconstruction and fusion module 340, and a spatial clustering module 350, wherein...

[0124] The feature construction module 310 is used to construct the physical spatial map feature matrix, the correlation map feature matrix, and the spatial feature matrix of the target sequencing object based on the spatial transcription data of the target sequencing object.

[0125] The feature fusion module 320 is used to sequentially perform gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix and the spatial feature matrix to obtain the target latent representation matrix;

[0126] The decoding and reconstruction module 330 is used to perform linear space reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix.

[0127] The reconstruction and fusion module 340 is used to perform gated weighted fusion of the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix;

[0128] The spatial clustering module 350 is used to perform spatial clustering on each spatial transcription sequencing point of the target sequencing object based on the target optimized expression matrix, and obtain spatial domain resolution results.

[0129] The technical solution of this embodiment constructs a physical spatial map feature matrix, a correlation map feature matrix, and a spatial feature matrix based on the spatial transcription data of the target sequencing object. Then, gated weighted fusion is performed between these matrices to obtain the target latent representation matrix. The target latent representation matrix is ​​then subjected to linear spatial reconstruction and nonlinear neural decoding to obtain linear reconstruction matrices and nonlinear reconstruction matrices, which are then gated weighted fused to obtain the target optimized expression matrix. Based on this optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain the spatial domain resolution result. This solves the problems of existing technologies where feature recognition highly depends on spatial continuity and feature fusion methods are rigid. It achieves connectivity of distant similar sites across physical distances and adaptive feature fusion in spatial domain resolution, accurately identifying dispersed and small discontinuous spatial domains such as tumor immune infiltration areas and microvascular networks, while preventing excessive feature smoothing or excessive noise retention.

[0130] Optionally, the feature construction module 310 specifically includes: a preprocessing unit for obtaining the highly variable gene expression matrix, sequencing point spatial location matrix, and sequencing point spatial adjacency matrix of the target sequencing object; a feature dimensionality reduction unit for projecting the highly variable gene expression matrix into a latent space feature matrix; a spatial graph unit for performing graph convolution operations on the sequencing point spatial adjacency matrix and the latent space feature matrix to construct the physical space graph feature matrix; a related feature acquisition unit for obtaining the sequencing point related adjacency matrix based on the latent space feature matrix; a related graph unit for performing graph convolution operations on the sequencing point related adjacency matrix and the latent space feature matrix to construct the related graph feature matrix; and a spatial feature unit for constructing the spatial feature matrix based on the sequencing point spatial location matrix.

[0131] Optionally, the relevant feature acquisition unit is specifically configured to: calculate the correlation coefficient matrix of each of the spatial transcription sequencing points based on the latent space feature matrix; and perform target relevant edge screening processing on the correlation coefficient matrix according to a preset correlation coefficient threshold to obtain the correlation adjacency matrix of the sequencing points.

[0132] Optionally, the feature fusion module 320 is specifically configured as follows: using a pre-trained first gating network, the physical space graph feature matrix and the correlation graph feature matrix are subjected to the gating weighted fusion to obtain a dual-graph fusion feature matrix; using a pre-trained second gating network, the dual-graph fusion feature matrix and the spatial feature matrix are subjected to the gating weighted fusion to obtain the target latent representation matrix.

[0133] Optionally, the decoding and reconstruction module 330 is specifically configured to: project the target latent representation matrix into the original spatial feature matrix; perform graph convolution operation on the sequencing point spatial adjacency matrix and the original spatial feature matrix to construct the linear reconstruction matrix; and use a sinusoidal activation function to map the target latent representation matrix into the nonlinear reconstruction matrix.

[0134] Optionally, the reconstruction and fusion module 340 is specifically configured to obtain the linear weight parameters and nonlinear weight parameters of the linear reconstruction matrix and the nonlinear reconstruction matrix in the gated weighted fusion using the following formulas:

[0135] ,

[0136] in, and These represent the linear weight parameter and the nonlinear weight parameter, respectively. and Let the linear reconstruction matrix and the nonlinear reconstruction matrix represent respectively. This represents a matrix concatenation operation. This refers to a pre-trained multilayer perceptron. This represents the Sigmoid activation function. The matrix partitioning operation is represented by the following formula, which, based on the linear weight parameters and the nonlinear weight parameters, performs a weighted fusion operation on the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix:

[0137] ,

[0138] in, This represents the target optimization expression matrix. This refers to a pre-trained fully connected layer. It is a matrix dot product operation, which means multiplying corresponding elements of a matrix.

[0139] Optionally, the spatial domain resolution device further includes: a joint training module, used to perform joint parameter training on the gated weighted fusion gated network parameters based on the mean square error loss function.

[0140] ,

[0141] in, This represents the original gene expression matrix in the spatial transcription data. and This indicates the preset balance coefficient.

[0142] The spatial domain resolution apparatus provided in this application can execute the spatial domain resolution method provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the execution method.

[0143] Example 3

[0144] Figure 4 This is a schematic diagram of the structure of a spatial domain resolution device provided in Embodiment 3 of this application, as shown below. Figure 4 As shown, the spatial domain resolution device includes a processor 510, a memory 520, an input device 530, and an output device 540; the number of processors 510 in the spatial domain resolution device can be one or more. Figure 4 Taking a processor 510 as an example; the processor 510, memory 520, input device 530, and output device 540 in the spatial domain resolution device can be connected via a bus or other means. Figure 4 Taking the example of a connection between China and Israel via a bus.

[0145] The memory 520, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the spatial domain resolution method in this embodiment (e.g., the feature construction module 310, feature fusion module 320, decoding and reconstruction module 330, reconstruction and fusion module 340, and spatial clustering module 350 in the spatial domain resolution device). The processor 510 executes various functional applications and data processing of the spatial domain resolution device by running the software programs, instructions, and modules stored in the memory 520, thereby implementing the aforementioned spatial domain resolution method.

[0146] The memory 520 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 520 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 520 may further include memory remotely located relative to the processor 510, which can be connected to a spatial domain resolution device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0147] Input device 530 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the spatial domain resolution device. Output device 540 may include display devices such as a display screen.

[0148] Example 4

[0149] Embodiment 4 of this application also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform a spatial domain resolution method, including:

[0150] Based on the spatial transcription data of the target sequencing object, construct the physical spatial map feature matrix, the correlation map feature matrix, and the spatial feature matrix of the target sequencing object;

[0151] The physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix are sequentially subjected to gated weighted fusion to obtain the target latent representation matrix.

[0152] The target latent representation matrix is ​​subjected to linear space reconstruction and nonlinear neural decoding to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, respectively.

[0153] The linear reconstruction matrix and the nonlinear reconstruction matrix are gated and weighted to obtain the target optimized expression matrix;

[0154] Based on the target optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain spatial domain resolution results.

[0155] Of course, the computer-executable instructions provided in the embodiments of this application are not limited to the method operations described above, but can also perform related operations in the spatial domain resolution method provided in any embodiment of this application.

[0156] Based on the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0157] It is worth noting that in the embodiments of the above-mentioned spatial domain resolution device, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of this application.

[0158] Although this application has been described in detail above with general descriptions, specific embodiments, and experiments, some modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of this application fall within the scope of protection claimed in this application.

Claims

1. A spatial domain resolution method, characterized in that, include: Based on the spatial transcription data of the target sequencing object, construct the physical spatial map feature matrix, the correlation map feature matrix, and the spatial feature matrix of the target sequencing object; The physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix are sequentially subjected to gated weighted fusion to obtain the target latent representation matrix. The target latent representation matrix is ​​subjected to linear space reconstruction and nonlinear neural decoding to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, respectively. The linear reconstruction matrix and the nonlinear reconstruction matrix are gated and weighted to obtain the target optimized expression matrix; Based on the target optimized expression matrix, spatial clustering is performed on each spatial transcription sequencing point of the target sequencing object to obtain spatial domain resolution results.

2. The method according to claim 1, characterized in that, The construction of the physical spatial map feature matrix, correlation map feature matrix, and spatial feature matrix of the target sequencing object based on the spatial transcription data of the target sequencing object includes: Obtain the highly variable gene expression matrix, sequencing point spatial location matrix, and sequencing point spatial adjacency matrix of the target sequencing object; The highly variable gene expression matrix is ​​projected into a latent space feature matrix; Graph convolution is performed on the spatial adjacency matrix of the sequencing points and the latent space feature matrix to construct the physical space graph feature matrix; The correlation adjacency matrix of sequencing points is obtained based on the latent space feature matrix; Perform graph convolution operation on the sequencing point correlation adjacency matrix and the latent space feature matrix to construct the correlation graph feature matrix; The spatial feature matrix is ​​constructed based on the spatial location matrix of the sequencing points.

3. The method according to claim 2, characterized in that, The step of obtaining the sequencing point correlation adjacency matrix based on the latent space feature matrix includes: Based on the latent space feature matrix, the correlation coefficient matrix of each of the spatial transcription sequencing points is calculated; Based on a preset correlation coefficient threshold, the correlation coefficient matrix is ​​subjected to target correlation edge filtering to obtain the sequencing point correlation adjacency matrix.

4. The method according to claim 1, characterized in that, The step of sequentially performing gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix to obtain the target latent representation matrix includes: A pre-trained first gating network is used to perform the gated weighted fusion of the physical space graph feature matrix and the correlation graph feature matrix to obtain a dual-graph fusion feature matrix. A pre-trained second gating network is used to perform the gated weighted fusion of the dual-graph fusion feature matrix and the spatial feature matrix to obtain the target latent representation matrix.

5. The method according to claim 2, characterized in that, The process of performing linear spatial reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain linear reconstruction matrices and nonlinear reconstruction matrices includes: Project the target latent representation matrix into the original spatial feature matrix; The linear reconstruction matrix is ​​constructed by performing graph convolution operation on the spatial adjacency matrix of the sequencing points and the original spatial feature matrix. The target latent representation matrix is ​​mapped to the nonlinear reconstruction matrix using a sinusoidal activation function.

6. The method according to claim 1, characterized in that, The step of gating and weighting the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix includes: The linear and nonlinear weight parameters of the linear and nonlinear reconstruction matrices in the gated weighted fusion are obtained using the following formulas: ， in, and These represent the linear weight parameter and the nonlinear weight parameter, respectively. and Let the linear reconstruction matrix and the nonlinear reconstruction matrix represent respectively. This represents a matrix concatenation operation. This refers to a pre-trained multilayer perceptron. This represents the Sigmoid activation function. This represents a matrix partitioning operation; The target optimized expression matrix is ​​obtained by performing a weighted fusion operation on the linear reconstruction matrix and the nonlinear reconstruction matrix based on the linear weight parameters and the nonlinear weight parameters using the following formula: ， in, This represents the target optimization expression matrix. This refers to a pre-trained fully connected layer. It is a matrix dot product operation, which means multiplying corresponding elements of a matrix.

7. The method according to claim 6, characterized in that, Before performing gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix, and the spatial feature matrix to obtain the target latent representation matrix, the method further includes: Joint parameter training is performed on the gated network parameters based on the mean square error loss function: ， in, This represents the original gene expression matrix in the spatial transcription data. and This indicates the preset balance coefficient.

8. A spatial domain resolution device, characterized in that, include: The feature construction module is used to construct the physical spatial map feature matrix, correlation map feature matrix, and spatial feature matrix of the target sequencing object based on the spatial transcription data of the target sequencing object; The feature fusion module is used to sequentially perform gated weighted fusion on the physical space graph feature matrix, the correlation graph feature matrix and the spatial feature matrix to obtain the target latent representation matrix; The decoding and reconstruction module is used to perform linear space reconstruction and nonlinear neural decoding on the target latent representation matrix to obtain a linear reconstruction matrix and a nonlinear reconstruction matrix, respectively. The reconstruction and fusion module is used to perform gated weighted fusion of the linear reconstruction matrix and the nonlinear reconstruction matrix to obtain the target optimized expression matrix; The spatial clustering module is used to perform spatial clustering of each spatial transcription sequencing point of the target sequencing object based on the target optimized expression matrix, and obtain spatial domain resolution results.

9. A spatial domain resolution device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the spatial domain resolution method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the spatial domain resolution method as described in any one of claims 1-7.

Images