Method and apparatus for generating expression profile of tissue genes

Abstract

The application discloses a tissue gene expression atlas generation method and device, relates to the technical field of image processing, and mainly aims to solve the problem of poor effectiveness of existing tissue gene expression atlas generation. The method comprises the following steps: acquiring multi-modal graph structure data of a tissue gene; determining a weighted graph representation of the multi-modal graph structure data; generating an atlas based on a neural network model that has completed model training, to obtain an expression atlas of the tissue gene, wherein an attention mechanism is introduced into the neural network model, and attention weights of each layer in the neural network model are used to update topological structure information of each layer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a method and device for generating an expression atlas of tissue genes. BACKGROUND

[0002] The study of complex tissues is not only limited to the analysis of molecular characteristics of a large number of cells, but also needs to understand how the spatial environment affects the state or function of cells. The transcriptional state of cells can be regulated by a gene regulatory network (GRN), which is a collection of regulatory relationships between transcription factors (TFs) and their downstream target genes. In recent years, breakthroughs in spatially resolved transcriptomics (SRT) technology have made it possible to analyze transcriptomic characteristics while preserving positional information, thereby providing unprecedented opportunities to capture transcriptional states with spatial context.

[0003] Currently, existing SRT data spatial domain detection calculation methods can be roughly divided into two categories: probabilistic models and deep learning models. The probabilistic methods include Hidden Markov Random Field (HMRF), BayesSpace, and SpatialPCA, which integrate spatial information using probabilistic graphical models. The deep learning methods include SpaGCN, STAGATE, DeepST, and GraphST, which use graph convolutional networks and their derivatives to capture dependencies driven by spatial data. However, these two types of models handle spatial domain modeling problems differently, focusing on statistical probability or neural network learning ability to understand the spatial context embedded in the data. They rely on predefined graph structures and cannot fully capture the complex and nonlinear relationships between gene expression and spatial context, nor can they fully represent actual biological interactions, greatly reducing the integration effect of gene expression data and spatial information. Therefore, there is an urgent need for a method for generating an expression atlas of tissue genes to solve the above problems. SUMMARY

[0004] Therefore, the present application provides a method and device for generating an expression atlas of tissue genes, which mainly aims to solve the problem of poor effectiveness of existing expression atlas generation of tissue genes.

[0005] According to one aspect of the present application, a method for generating an expression atlas of tissue genes is provided, comprising:

[0006] Obtaining multi-modal graph structure data of tissue genes;

[0007] Determining a weighted graph representation of the multi-modal graph structure data;

[0008] perform graph generation on the weighted graph representation based on a neural network model that has completed model training, to obtain an expression graph of the tissue gene, wherein an attention mechanism is introduced into the neural network model, and attention weights of each layer in the neural network model are used to update topological structure information of each layer.

[0009] Further, before the neural network model that has completed model training performs graph generation on the weighted graph representation to obtain the expression graph of the tissue gene, the method further comprises:

[0010] obtaining a gene expression training sample, wherein the gene expression training sample comprises a graph representation sample of a label graph expression and an edge expression;

[0011] constructing a neural network model that introduces an attention mechanism, wherein a bias term is included in attention weights of each layer in the neural network model, and the bias term is determined based on position encoding and an edge sample;

[0012] performing model training on the neural network model based on the gene expression training sample, to obtain a neural network model that has completed model training.

[0013] Further, the method further comprises:

[0014] obtaining a gene singular matrix, and decomposing an adjacency matrix of the gene singular matrix based on a singular value decomposition algorithm to generate position encoding, wherein the adjacency matrix is determined based on a gene point position and a neighbor gene point position.

[0015] Further, before the model training on the neural network model based on the gene expression training sample, the method further comprises:

[0016] obtaining the adjacency matrix;

[0017] performing enhancement processing on the expression sample based on the adjacency matrix to obtain an enhanced expression sample, so as to perform model training based on the enhanced expression sample.

[0018] Further, after the expression graph of the tissue gene is obtained, the method further comprises:

[0019] determining an assignment probability for clustering based on a preset distribution that matches different clustering tasks, and determining a target assignment distribution of the assignment probability based on a normalization algorithm;

[0020] determining a clustering loss function based on the assignment probability and the target assignment distribution, and constraining a clustering model based on the clustering loss function;

[0021] cluster the expression atlas based on the constrained clustering model to obtain an atlas clustering result of the clustering task.

[0022] Further, the determining the weighted graph representation of the multi-modal graph structure data comprises:

[0023] parsing gene points and neighbor gene points in the multi-modal graph structure data, and obtaining an adjacency matrix based on spatial distances between the gene points and the neighbor gene points;

[0024] determining the weighted graph representation based on gene expression data in the multi-modal graph structure data and the adjacency matrix.

[0025] Further, the method further comprises:

[0026] constructing a proximity network based on transcription representation data in the expression atlas, and performing spatial recognition processing based on the proximity network to identify a spatial domain of the tissue gene;

[0027] and based on the topological structure data in the expression atlas, performing denoising processing, and based on the denoised topological structure data, performing specificity recognition to determine a specific differential gene.

[0028] According to another aspect of the present application, an expression atlas generation device of a tissue gene is provided, comprising:

[0029] an acquisition module configured to acquire multi-modal graph structure data of a tissue gene;

[0030] a determination module configured to determine a weighted graph representation of the multi-modal graph structure data;

[0031] a generation module configured to generate an atlas based on the weighted graph representation by using a neural network model that has completed model training, to obtain an expression atlas of the tissue gene, wherein an attention mechanism is introduced into the neural network model, and each layer attention weight in the neural network model is used to update topological structure information of each layer.

[0032] Further, the device further comprises a construction module, a training module,

[0033] The acquisition module is further configured to acquire a gene expression training sample, wherein the gene expression training sample comprises a labeled graph expression and an edge expression graph representation sample.

[0034] The construction module is configured to construct a neural network model that introduces an attention mechanism, wherein each layer attention weight in the neural network model comprises a bias term, and the bias term is determined based on position encoding and an edge sample.

[0035] The training module is configured to perform model training on the neural network model based on the gene expression training sample, to obtain a neural network model after model training.

[0036] Further, the generating module is further configured to obtain a gene singular matrix, and perform singular value decomposition on an adjacency matrix of the gene singular matrix based on a singular value decomposition algorithm, to generate position encoding, wherein the adjacency matrix is determined based on gene point positions and neighbor gene point positions.

[0037] Further,

[0038] The obtaining module is further configured to obtain the adjacency matrix, and perform enhancement processing on the expression sample based on the adjacency matrix, to obtain an enhanced expression sample, and perform model training based on the enhanced expression sample.

[0039] Further, the apparatus further comprises:

[0040] The determining module is further configured to determine an assignment probability for clustering based on a preset distribution matched with different clustering tasks, and determine a target assignment distribution of the assignment probability based on a normalization algorithm, determine a clustering loss function based on the assignment probability and the target assignment distribution, and constrain a clustering model based on the clustering loss function, and perform clustering processing on the expression atlas based on the constrained clustering model, to obtain an atlas clustering result of the clustering task.

[0041] Further, the determining module is further configured to analyze gene point positions and neighbor gene point positions in the multi-modal graph structure data, and obtain an adjacency matrix based on spatial distances between the gene point positions and the neighbor gene point positions, and determine the weighted graph representation based on gene expression data in the multi-modal graph structure data and the adjacency matrix.

[0042] Further, the apparatus further comprises:

[0043] The processing module is configured to construct a proximity network based on transcription representation data in the expression atlas, perform spatial recognition processing based on the proximity network, to recognize a spatial domain of the tissue gene, perform denoising processing based on topological structure data in the expression atlas, and perform specificity recognition based on denoised topological structure data, to determine a specific difference gene.

[0044] According to still another aspect of the present application, a storage medium is provided, in which at least one executable instruction is stored, which causes a processor to perform operations corresponding to the above-mentioned expression atlas generation method of a tissue gene.

[0045] According to another aspect of the present application, a terminal is provided, comprising a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete communication with each other through the communication bus;

[0046] The memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform the corresponding operation of the above-mentioned tissue gene expression profile generation method.

[0047] By the above technical solution, the technical solution provided by the embodiments of the present application has at least the following advantages:

[0048] The present application provides a tissue gene expression profile generation method and device. Compared with the prior art, the embodiments of the present application obtain multi-modal graph structure data of a tissue gene; determine a weighted graph representation of the multi-modal graph structure data; perform graph generation on the weighted graph representation based on a neural network model that has completed model training, to obtain an expression profile of the tissue gene, the attention mechanism is introduced into the neural network model, the attention weight of each layer in the neural network model is used to update the topological structure information of each layer, the reinforced self-attention mechanism is realized, the topological structure information of the iterative evolution graph representation and the transcription signal representation are realized, and the global self-attention is used to replace the graph convolution, so that the global information and the spatial local information can be integrated, thereby improving the detection capability of the spatial functional domain and improving the integration effect of the gene expression data and the spatial information.

[0049] The above description is only a summary of the technical solutions of the present application. In order to enable the technical means of the present application to be more clearly understood, the following detailed description of the embodiments of the present application can be implemented according to the content of the description, and in order to enable the above and other purposes, characteristics and advantages of the present application to be more obvious and easy to understand, the following detailed description of the embodiments of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0050] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included only to illustrate preferred embodiments and are not intended to limit the present application. Moreover, the same reference numerals are used throughout the accompanying drawings to designate the same parts. In the drawings:

[0051] Figure 1 A flow chart of a tissue gene expression profile generation method provided by the embodiments of the present application is shown;

[0052] Figure 2 A schematic diagram of a neural network model processing flow provided by the embodiments of the present application is shown;

[0053] Figure 3 A block diagram of a tissue gene expression profile generation device provided by the embodiments of the present application is shown;

[0054] Figure 4 A structure diagram of a terminal is shown. DETAILED DESCRIPTION

[0055] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0056] The present embodiment provides a method for generating an expression atlas of a tissue gene, as shown in Figure 1 The method comprises:

[0057] 101. Obtain multi-modal graph structure data of a tissue gene.

[0058] In the present embodiment, the current execution subject as the processing end for generating an expression atlas can be a terminal device or a cloud server, etc., so as to obtain multi-modal graph structure data of a tissue gene. The tissue gene can be a gene image of a tissue cell obtained by cutting different species in different dimensions in a large-scale spatiotemporal omics data set, or a gene image of a tissue cell collected in real time, which is not limited in the present embodiment. The multi-modal graph structure data includes vector data, image data, and spatial coordinate data, etc., which are different modal graph data representing gene mechanisms, which are not limited in the present embodiment.

[0059] 102. Determine a weighted graph representation of the multi-modal graph structure data.

[0060] In the present embodiment, after obtaining the multi-modal graph structure data, in order to perform atlas generation thereon, the current execution end determines a weighted graph representation of the multi-modal graph structure data, i.e., the weighted graph representation includes expression data of a tissue gene on a gene spot or a cell in the multi-modal graph structure data and an adjacency matrix derived from different modalities in space, which can be represented as G(X1, A), where X1∈R M×N represents the expression of M component genes on N spots / cells, and A∈R N×N is an adjacency matrix derived from spatial multi-modal data, i.e., the relationship between each spot and its spatial neighbor can be captured to capture the inherent gene spatial local structure in SRT data, which is not limited in the present embodiment.

[0061] 103. Perform atlas generation on the weighted graph representation based on a neural network model trained by a completed model, to obtain an expression atlas of the tissue gene.

[0062] In the embodiments of the present application, after obtaining the weighted graph representation, the weighted graph representation is input into the neural network model as a model input for graph generation. At this time, an attention mechanism is introduced into the neural network model, and the attention weight of each layer in the neural network model is used to update the topological structure information of each layer. The topological structure information is used to represent the structural relationship in the gene graph, which is not limited in the embodiments of the present application. In addition, the attention weight in the self-attention mechanism is used to iteratively update the topological structure information of the graph representation between layers, which can strengthen the self-attention module and realize efficient learning and updating of the expression graph of the entire model.

[0063] It should be noted that the neural network model in the embodiments of the present application is used as a spatial perception graph transformer to transform the weighted graph representation. After the conversion of the i+1 layer, the expression graph of the tissue gene is obtained, which is represented as G(H L ,E L ), H L is a gene transcription signal representation, and E L is gene topological structure information, which is not limited in the embodiments of the present application.

[0064] In another embodiment of the present application, in order to further limit and illustrate, before obtaining the expression graph of the tissue gene by generating a graph based on the neural network model that has completed model training on the weighted graph representation, the method further comprises:

[0065] obtaining a gene expression training sample;

[0066] constructing a neural network model introducing an attention mechanism;

[0067] model training the neural network model based on the gene expression training sample to obtain a neural network model that has completed model training.

[0068] In order to realize the atlas generation by using the neural network with attention mechanism, so as to accurately construct the gene representation with clearer structure, the current execution end obtains the gene expression training sample in advance, at this time, the gene expression training sample includes the graph expression of the label graph and the edge expression, which can be obtained from the SRT database, the SRT database includes STOmics, SOAR, SpatialDB, CROST, 10x Genomics website (which can be obtained through 10x Genomics) and Census (which can be obtained through Cellxgene) and the like, for example, containing 96,700,729 cells / positions from 7,367 tissue sections, covering 365 tissue cells, including lung, skin, brain, liver, kidney, spinal cord and embryo, and covering normal tissues and various diseases, such as pancreatic ductal adenocarcinoma, amyotrophic lateral sclerosis, non-small cell lung cancer and hepatocellular carcinoma, the embodiments of the present application are not limited. The embodiments of the present application are not limited. Wherein, the graph sample includes the weighted graph sample as the model input, which is represented as G(X0,A).

[0069] It should be noted that, as Figure 2 indicated, after the current execution end constructs the neural network model with the introduced attention mechanism, the model training is performed on the neural network model based on the gene expression training sample, in the training process, the bias term is contained in the attention weight of each layer of the neural network model, wherein, during the training of each layer, the output of the previous layer is the graph expression H i and the edge expression E i , which is used as the input of the current layer for the training of the current layer. Wherein, the H i and E i are updated by the structure reinforced self-attention module, and the H i is represented as:

[0070]

[0071] E i is represented as:

[0072] Wherein, L is the number of layers of the neural network, e represents the update of the point pair topology structure, O is the output of each layer of the neural network, h represents the update of the transcriptome signal, Γ is the number of multiple heads of multi-head attention, V is the value matrix of the self-attention module, S is the attention weight of the self-attention module, and concact(*) is the row concatenation. At this time, the low-dimensional X1 of the gene expression and the relationship matrix A between each point and its spatial neighbor are represented as H 0 and E 0 respectively. Further, the application feedforward sublayer (FFN) and layer normalization (LN) in the neural network model obtain the final graph representation as:

[0073]

[0074] In addition, in the embodiments of the present application, in order to integrate global information and spatial local information, an attention mechanism architecture is introduced into the neural network model, that is, the attention process in the i-th layer and the t-th attention head is represented as:

[0075]

[0076] wherein, are the query, key and value matrices of the t-th head in the i-th layer. The dimensions of the query and key are d i The operator Θ is an element-wise multiplication. is a bias term added to the attention weight, which is determined based on the position encoding and the edge sample, so that the spatial information affects the global aggregation process. The sigmoid function σ(·) is used, so that the spatial information gates the value before aggregation, controlling the flow of information between sites. At the same time, the clip(·) operation is used to constrain the scaled dot product, thereby improving numerical stability. In addition, P S,E and P S,G are learned projection matrices.

[0077] In another embodiment of the present application, in order to further limit and illustrate, the step further comprises:

[0078] The gene singular matrix is obtained, and the adjacency matrix of the gene singular matrix is decomposed based on a singular value decomposition algorithm to generate the position encoding.

[0079] In order to realize the position encoding P as the input of the attention mechanism, thereby updating the attention weight in each attention, the current execution end obtains the gene singular matrix, denoted as U, V, U, V ∈ R M×r , respectively containing the first r left and right singular matrices, and then based on a singular value decomposition algorithm (Singular Value Decomposition) to decompose the adjacency matrix A of the gene singular matrix to generate the position encoding, which is represented by the formula:

[0080]

[0081] wherein, Σ ∈ R γ×γ is a diagonal matrix, || represents column concatenation, W PE ∈ R 2γ×N is a learned projection matrix, and P ∈ R M×N is the position encoding.

[0082] In another embodiment of the present application, in order to further limit and illustrate, before the step of training the neural network model based on the gene expression training samples, the method further comprises:

[0083] obtaining the adjacency matrix;

[0084] performing enhancement processing on the expression samples through the adjacency matrix to obtain enhanced expression samples, so as to perform model training based on the enhanced expression samples.

[0085] In order to enhance the gene expression of each gene point, thereby improving the accuracy of expression map generation, the current execution end generates enhanced spatial expression. Specifically, first, the adjacency matrix A is obtained, and the enhanced expression formula X1=X0+αX0A is used, wherein α is an adjustable parameter for controlling the influence of spatial neighborhood similarity on spatial domain identification. The present application does not make specific limitation based on the demand.

[0086] In another embodiment of the present application, in order to further limit and illustrate, after the step of obtaining the expression map of the tissue gene, the method further comprises:

[0087] determining an assignment probability for clustering based on a preset distribution matching different clustering tasks, and determining a target assignment distribution of the assignment probability based on a normalization algorithm;

[0088] determining a clustering loss function based on the assignment probability and the target assignment distribution, and constraining a clustering model based on the clustering loss function;

[0089] performing clustering processing on the expression map based on the constrained clustering model to obtain a graph clustering result of the clustering task.

[0090] In order to make the obtained expression map have stronger clustering performance, thereby improving the applicability of different clustering tasks, after the current execution end obtains the expression map, first, an assignment probability for clustering is determined based on a preset distribution matching different clustering tasks, and a target assignment distribution of the assignment probability is determined based on a normalization algorithm. Wherein, the preset distribution is preferably a student t distribution, so as to measure the probability of a point i being assigned to a clustering task j through the student t distribution, and then based on the combined expression map H L The calculation formula of the assignment probability is:

[0091]

[0092] Wherein, μ is the clustering center determined by the K-means clustering algorithm, is the i-th column of H L , and Q=[q ijrepresents the assignment probability of all tissue genes, and p is the degree of freedom of the student t distribution. Further, the current execution end determines the clustering loss function based on the assignment probability and the target assignment distribution, and at this time, the target assignment distribution is represented as where j ＝ ∑ i q ij ; wherein each q ij is normalized to generate the target assignment distribution P, and j' is the center point of each cluster. In addition, in the embodiment of the present application, the clustering loss function is determined based on minimizing the Kullback-Leibler (KL) divergence between Q and P distributions, and increasing the confidence of the assignment, which is represented as At this time, the loss function enhances the accuracy and stability of clustering by minimizing the difference between the distributions Q and P, which helps to improve the performance of the clustering task, therefore, the clustering model is constrained based on the clustering loss function, so as to perform clustering processing on the expression atlas based on the constrained clustering model, and obtain the atlas clustering result of the clustering task.

[0093] In another embodiment of the present application, in order to further limit and illustrate, the step of determining the weighted graph representation of the multi-modal graph structure data comprises:

[0094] Parsing the gene point position and the neighbor gene point position in the multi-modal graph structure data, and obtaining an adjacency matrix based on the spatial distance between the gene point position and the neighbor gene point position;

[0095] Determining the weighted graph representation based on the gene expression data in the multi-modal graph structure data and the adjacency matrix.

[0096] In order to better learn the gene structure, the weighted graph representation is G(X0,A), at this time, the current execution end can calculate the Euclidean distance between each pair of gene point positions spot according to the spatial coordinates and the corresponding histological image information, and select k nearest spatial neighbors for each point, further, based on the principal component analysis kernel PCA for embedding, calculate the cosine distance between the gene point position and the neighbor gene point position, and convert its index to a similarity matrix A, the specific formula is represented as: where matrix U∈R 15×Nis a low-dimensional matrix composed of 15 principal components PCs, and the matrix D represents the cosine distance of these adjacent spots as the spatial distance. In addition, the 15 principal components PCs are the first 15 principal components selected according to the corresponding eigenvalues after the PCA dimension reduction of X1. At this time, the principal component analysis (PCA) is a commonly used data dimension reduction technique, which is used to convert a high-dimensional data set into a low-dimensional space. After principal component analysis, a plurality of principal components PCs can be obtained.

[0097] In another embodiment of the present application, in order to further limit and illustrate, the step further comprises:

[0098] Based on the transcriptional representation data in the expression atlas, a proximity network is constructed, and spatial recognition processing is performed based on the proximity network to identify the spatial domain of the tissue gene;

[0099] Based on the topological structure data in the expression atlas, denoising processing is performed, and specificity recognition is performed based on the denoised topological structure data to determine the specific differential gene.

[0100] In order to improve the effectiveness and accuracy of generating the expression atlas, the current execution end can also perform denoising on the expression atlas. Specifically, first, based on the transcriptional representation data H L The proximity network is constructed, that is, each spot spot and its nearest 15 spots in the expression representation form an adjacency matrix, and spatial recognition processing is performed based on the proximity network. Leiden algorithm can be used for calculation to identify the spatial domain of the tissue gene. At the same time, based on the topological structure data E L Denoising processing is performed to obtain a denoised expression atlas In order to enhance the spatial expression pattern and domain specificity, specificity recognition is performed based on the denoised topological structure data to determine the specific differential gene. The embodiments of the present application are not limited.

[0101] The embodiment of the present application provides a kind of tissue gene expression atlas generation method, compared with prior art, the embodiment of the present application obtains the multi-modal graph structure data of tissue gene;Determine the weighted graph representation of the multi-modal graph structure data;Based on the neural network model of model training completed, the atlas generation of the weighted graph representation is carried out, and the expression atlas of the tissue gene is obtained, attention mechanism is introduced in the neural network model, and the topology structure information of each layer in the neural network model is used to update each layer, realize the topology structure information of reinforced self-attention mechanism, realize the purpose of iterative evolution graph representation and transcription signal representation, and by using global self-attention instead of graph convolution, global information and spatial local information can be integrated, so as to improve the detection ability of spatial functional domain, improve the integration effect of gene expression data and spatial information.

[0102] Further, as to the implementation of the method shown in the above Figure 1 The embodiment of the present application provides a kind of tissue gene expression atlas generation device, as shown in Figure 3 The device comprises:

[0103] The acquisition module 21 is used for acquiring the multi-modal graph structure data of tissue gene;

[0104] The determination module 22 is used for determining the weighted graph representation of the multi-modal graph structure data;

[0105] The generation module 23 is used for carrying out atlas generation based on the neural network model of model training completed to the weighted graph representation, and the expression atlas of the tissue gene is obtained, attention mechanism is introduced in the neural network model, and the topology structure information of each layer in the neural network model is used to update each layer.

[0106] Further, the device further comprises: construction module, training module,

[0107] The acquisition module is also used for acquiring gene expression training sample, and the gene expression training sample includes labeled graph expression, edge expression graph representation sample;

[0108] The construction module is used for constructing the neural network model of introducing attention mechanism, and the bias term is contained in the attention weight of each layer in the neural network model, and the bias term is determined based on position coding and edge sample;

[0109] The training module is used for carrying out model training based on the gene expression training sample to the neural network model, and the neural network model of completing model training is obtained.

[0110] Further, the generating module is further configured to obtain a gene singular matrix, decompose an adjacency matrix of the gene singular matrix based on a singular value decomposition algorithm, and generate position encoding, wherein the adjacency matrix is determined based on gene points and neighbor gene points.

[0111] Further,

[0112] The obtaining module is further configured to obtain the adjacency matrix, perform enhancement processing on the expression sample based on the adjacency matrix to obtain an enhanced expression sample, and perform model training based on the enhanced expression sample.

[0113] Further, the apparatus further comprises:

[0114] The determining module is further configured to determine an assignment probability for clustering based on a preset distribution matched with different clustering tasks, determine a target assignment distribution of the assignment probability based on a normalization algorithm, determine a clustering loss function based on the assignment probability and the target assignment distribution, constrain a clustering model based on the clustering loss function, and perform clustering processing on the expression atlas based on the constrained clustering model to obtain an atlas clustering result of the clustering task.

[0115] Further, the determining module is further configured to analyze gene points and neighbor gene points in the multi-modal graph structure data, obtain an adjacency matrix based on a spatial distance between the gene points and the neighbor gene points, and determine the weighted graph representation based on gene expression data in the multi-modal graph structure data and the adjacency matrix.

[0116] Further, the apparatus further comprises:

[0117] The processing module is configured to construct a proximity network based on transcriptional representation data in the expression atlas, perform spatial recognition processing based on the proximity network to recognize a spatial domain of the tissue gene, perform denoising processing based on topological structure data in the expression atlas, and perform specificity recognition based on denoised topological structure data to determine a specific differential gene.

[0118] The embodiment of the present application provides a kind of tissue gene expression atlas generation device, compared with prior art, the embodiment of the present application obtains the multi-modal graph structure data of tissue gene;Determine the weighted graph representation of the multi-modal graph structure data;Based on the neural network model of model training completed, the atlas generation is carried out to the weighted graph representation, obtains the expression atlas of the tissue gene, attention mechanism is introduced in the neural network model, and the topology structure information of each layer in the neural network model is used to update each layer, realize the self-attention mechanism of combination reinforcement, realize the purpose of iterative evolution graph representation topology structure information and transcription signal representation, and by using global self-attention instead of graph convolution, global information and spatial local information can be integrated, to improve the detection capability of spatial functional domain, improve the integration effect of gene expression data and spatial information.

[0119] According to an embodiment of the present application, a storage medium is provided, and the storage medium stores at least one executable instruction. The computer executable instruction can execute the tissue gene expression atlas generation method in any method embodiment described above.

[0120] Figure 4 A structural schematic diagram of a terminal according to an embodiment of the present application is shown, and the specific implementation of the terminal is not limited in the specific embodiments of the present application.

[0121] As shown in Figure 4 The terminal can include a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0122] The processor 302, the communications interface 304, and the memory 306 can communicate with each other through the communications bus 308.

[0123] The communications interface 304 is configured to communicate with network elements such as clients or other servers.

[0124] The processor 302 is configured to execute the program 310, and can execute the related steps in the tissue gene expression atlas generation method embodiments described above.

[0125] Specifically, the program 310 can include program code, and the program code includes computer operation instructions.

[0126] The processor 302 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to perform the embodiments of the present application. The terminal comprises one or more processors, which can be the same type of processors, such as one or more CPUs; or can be different types of processors, such as one or more CPUs and one or more ASICs.

[0127] The memory 306 is used to store programs 310. The memory 306 can include a high-speed RAM memory, and can also include a non-volatile memory, such as at least one disk memory.

[0128] The programs 310 can be specifically used to make the processor 302 perform the following operations:

[0129] Obtain multi-modal graph structure data of tissue genes;

[0130] Determine a weighted graph representation of the multi-modal graph structure data;

[0131] Generate an atlas of the weighted graph representation based on a neural network model that has completed model training, to obtain an expression atlas of the tissue genes, the neural network model introducing an attention mechanism, and each layer attention weight in the neural network model being used to update topological structure information of each layer.

[0132] Obviously, those skilled in the art should understand that the above-mentioned modules or steps of the present application can be realized by general computing devices, which can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be realized by program codes executable by computing devices, so that they can be stored in storage devices and executed by computing devices, and in some cases, the steps shown or described can be executed in different order, or they can be manufactured into individual integrated circuit modules, or multiple modules or steps can be manufactured into a single integrated circuit module. Thus, the present application is not limited to any specific combination of hardware and software.

[0133] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating an expression profile of genes in a tissue, characterized by, The method comprises the following steps: obtaining multi-modal graph structure data of tissue genes; determining a weighted graph representation of the multi-modal graph structure data; generating an atlas based on a neural network model trained on the weighted graph representation, obtaining an expression atlas of the tissue genes, the neural network model introducing an attention mechanism, and the attention weight of each layer of the neural network model being used to update the topological structure information of each layer; before the step of generating an atlas based on the neural network model trained on the weighted graph representation, the method further comprises the following steps: obtaining a gene expression training sample, wherein the gene expression training sample comprises a labeled graph expression and an edge expression graph representation sample; constructing a neural network model introducing an attention mechanism, wherein the bias term in the attention weight of each layer of the neural network model is determined based on position encoding and edge samples; training the neural network model based on the gene expression training sample to obtain a neural network model trained on the gene expression training sample; wherein, at each layer training, the output of the previous layer is a graph representation and an edge representation to be used as input for the current layer training, and the graph representation and the edge representation are updated when the graph representation is represented as: ； The edge expression is represented as: ； a number of layers of a neural network model, denotes an update of a point pair topology, denotes an output of each layer in a neural network model, denotes an update of a transcriptome signal, a number of heads of multi-head attention, a value matrix of a self-attention module, an attention weight of a self-attention module, a concatenation along a row, a layer variable of a neural network model, a head variable of multi-head attention.

2. The method of claim 1, wherein, the method further comprises the following steps: obtaining a gene singular matrix and decomposing an adjacency matrix of the gene singular matrix based on a singular value decomposition algorithm to generate position encoding, wherein the adjacency matrix is determined based on gene points and neighbor gene points.

3. The method of claim 2, wherein, before the step of training the neural network model based on the gene expression training sample, the method further comprises the following steps: obtaining the adjacency matrix; enhancing the gene expression training sample based on the adjacency matrix to obtain an enhanced expression sample, and training the model based on the enhanced expression sample.

4. The method of claim 1, wherein, after obtaining the expression atlas of the tissue genes, the method further comprises the following steps: determining an assignment probability for clustering based on a preset distribution matching different clustering tasks, and determining a target assignment distribution of the assignment probability based on a normalization algorithm; determining a clustering loss function based on the assignment probability and the target assignment distribution, and constraining a clustering model based on the clustering loss function; performing clustering processing on the expression atlas based on the constrained clustering model to obtain an atlas clustering result of the clustering task.

5. The method of claim 1, wherein, The method further comprises the following steps: analyzing gene points and neighbor gene points in the multi-modal graph structure data, and obtaining an adjacency matrix based on the spatial distance between the gene points and the neighbor gene points; determining the weighted graph representation based on gene expression data in the multi-modal graph structure data and the adjacency matrix.

6. The method according to any one of claims 1 to 5, characterized in that, The method further comprises the following steps: constructing a proximity network based on transcriptional representation data in the expression atlas, and performing spatial recognition processing based on the proximity network to recognize the spatial domain of the tissue genes; performing denoising processing based on topological structure data in the expression atlas, and performing specificity recognition based on the denoised topological structure data to determine specific difference genes.

7. An apparatus for generating an expression profile of a gene of a tissue, characterized by, The method comprises the following steps: an obtaining module for obtaining multi-modal graph structure data of tissue genes; a determining module for determining a weighted graph representation of the multi-modal graph structure data; The generating module is configured to perform graph generation on the weighted graph representation based on a neural network model that has completed model training, to obtain an expression graph of the tissue gene, wherein the neural network model introduces an attention mechanism, and attention weights of each layer in the neural network model are used to update topological structure information of each layer. The device further comprises a constructing module and a training module, The obtaining module is further configured to obtain a gene expression training sample, wherein the gene expression training sample comprises a labeled graph expression and an edge expression graph representation sample. The constructing module is configured to construct a neural network model that introduces an attention mechanism, wherein bias terms are included in attention weights of each layer in the neural network model, and the bias terms are determined based on position encoding and edge samples. The training module is configured to perform model training on the neural network model based on the gene expression training sample, to obtain a neural network model that has completed model training. wherein, at each layer training, the output of the previous layer is a graph representation and an edge representation to be used as input for the current layer training, and the graph representation and the edge representation are updated when the graph representation is represented as: ； The edge expression is represented as: ； is a number of layers of a neural network model, is an update of a point pair topology, is an output of each layer in a neural network model, is an update of a transcriptome signal, is a number of heads of multi-head attention, is a value matrix of a self-attention module, is an attention weight of a self-attention module, is a concatenation along a row, is a layer variable of a neural network model, is a head variable of multi-head attention.

8. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to implement the steps of the method of claim 1.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1-8. The processor executes the computer program to implement the steps of the method of claim 1.

Images