WO2025123187A1 - Feature extraction method and apparatus for cell grouping, and cell grouping method and apparatus - Google Patents

Abstract

The present application relates to a feature extraction method and apparatus for cell grouping, and a cell grouping method and apparatus. The feature extraction method for cell grouping comprises: performing disturbance on a cell feature matrix of a sample cell, so as to obtain a first feature matrix and a second feature matrix; performing disturbance on an adjacency matrix of the sample cell, so as to obtain a first adjacency matrix and a second adjacency matrix; by means of a feature encoder to be trained, performing encoding on a first image to obtain a first embedding matrix, and performing encoding on a second image to obtain a second embedding matrix; fusing the first embedding matrix with the second embedding matrix, so as to obtain a target embedding matrix; performing decoding on the target embedding matrix, so as to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; and on the basis of at least one of a first difference and a second difference, iteratively optimizing parameters of said feature encoder until an iteration stop condition is met, so as to obtain a trained feature encoder, wherein the feature encoder is used for extracting features required for cell grouping.

Description

Cell clustering feature extraction method, cell clustering method and device Technical Field

The present application relates to the field of computer technology and artificial intelligence technology, and in particular to a method for extracting cell clustering features, a cell clustering method and a cell clustering device.

Background Art

With the development of artificial intelligence technology, it has become possible to automatically cluster cells. The purpose of cell clustering is to divide cells into different clusters based on the similarity and difference of each cell's characteristics, ensuring that the cells in each cluster are as similar as possible and the cells in different clusters are as different as possible.

In traditional methods, principal component analysis (PCA) is generally used to represent cell characteristics in order to reduce the cell feature dimension while minimizing information loss. However, this method cannot extract the deep-level characteristics of cell gene expression, resulting in inaccurate results of cell clustering based on the represented information.

Summary of the invention

Based on this, it is necessary to provide a cell clustering feature extraction method and device, a cell clustering method and device, a computer device and a computer-readable storage medium that can accurately extract the features required for cell clustering in order to address the above technical problems.

In a first aspect, the present application provides a method for extracting cell clustering features. The method comprises:

Perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix;

Perturbing the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix;

The first graph is encoded by a feature encoder to be trained to obtain a first embedding matrix, and the second graph is encoded to obtain a second embedding matrix; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix;

Fusing the first embedding matrix and the second embedding matrix to obtain a target embedding matrix;

Decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix;

According to at least one of the first difference and the second difference, the parameters of the feature encoder to be trained are iteratively optimized until the iteration stop condition is met, thereby obtaining a trained feature encoder; the feature encoder is used to extract features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

In a second aspect, the present application also provides a cell clustering method. The method comprises:

Inputting a graph composed of a cell feature matrix and an adjacency matrix of cells to be grouped into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to a first graph and a second graph; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the adjacency matrix of the sample cells;

Clustering is performed on the embedded vectors corresponding to each of the cells to be grouped in the extracted embedded matrix to obtain a cell grouping result of the cells to be grouped.

In a third aspect, the present application also provides a cell clustering feature extraction device. The device comprises:

A first perturbation module is used to perturb the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix;

A second perturbation module is used to perturb the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix;

An encoding module, configured to encode a first graph through a feature encoder to be trained to obtain a first embedding matrix, and encode a second graph to obtain a second embedding matrix; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix;

A fusion module, configured to fuse the first embedding matrix and the second embedding matrix to obtain a target embedding matrix;

A decoding module, used for decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix;

A parameter optimization module is used to iteratively optimize the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference until the iteration stop condition is met, thereby obtaining a trained feature encoder; the feature encoder is used to extract features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

In a fourth aspect, the present application also provides a cell clustering device. The device comprises:

The embedding matrix determination module is used to input the graph composed of the cell feature matrix and the adjacency matrix of the cells to be grouped into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to the first graph and the second graph; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of the sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the cell feature matrix of the sample cells; The cell's adjacency matrix is perturbed;

The clustering module is used to cluster the embedding vectors corresponding to each of the cells to be grouped in the extracted embedding matrix to obtain the cell grouping results of the cells to be grouped.

In a fifth aspect, the present application further provides a computer device. The computer device includes a memory and one or more processors, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the one or more processors execute the steps in the cell clustering feature extraction method or the cell clustering method described in each embodiment of the present application.

In a sixth aspect, the present application further provides a computer-readable storage medium, wherein a computer program is stored thereon, and when the computer program is executed by a processor, one or more processors execute the steps in the cell clustering feature extraction method or the cell clustering method described in each embodiment of the present application.

The above-mentioned cell clustering feature extraction method and device, cell clustering method and device, computer equipment and computer-readable storage medium perturb the cell feature matrix of the sample cells to obtain the first feature matrix and the second feature matrix, and perturb the adjacency matrix of the sample cells to obtain the first adjacency matrix and the second adjacency matrix, which can obtain richer features under different views of the graph, and then encode the first and second images of different views with rich information respectively through the feature encoder to be trained, and fuse the encoding results of the first and second images to obtain the target embedding matrix, so that the target embedding matrix can have richer and deeper features, and then decode the target embedding matrix to obtain the reconstructed cell feature matrix and the reconstructed adjacency matrix to determine the reconstruction loss, so that the features required for cell clustering can be accurately extracted through the feature encoder that can extract deep features.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are merely embodiments of the present application, and ordinary technicians in this field can obtain other drawings based on the disclosed drawings without paying any creative work.

FIG1 is a schematic diagram of a process for extracting cell clustering features in one embodiment;

FIG2 is a schematic diagram of the overall process of a method for extracting cell clustering features in one embodiment;

FIG3 is a schematic diagram of a process for cell clustering in one embodiment;

FIG4 is a structural block diagram of a cell clustering feature extraction device in one embodiment;

FIG5 is a structural block diagram of a cell clustering feature extraction device in another embodiment;

FIG6 is a block diagram of a cell clustering device in one embodiment;

FIG7 is a structural block diagram of a cell clustering device in another embodiment;

FIG8 is a diagram showing the internal structure of a computer device in one embodiment;

FIG. 9 is a diagram showing the internal structure of a computer device in another embodiment.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

In an exemplary embodiment, as shown in FIG1 , a method for extracting cell clustering features is provided, and this embodiment is illustrated by applying the method to a computer device. The computer device may be a terminal or a server. The terminal may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server may be implemented as an independent server or a server cluster consisting of multiple servers. In this embodiment, the method comprises the following steps:

Step 102 , perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix.

The cell feature matrix is a matrix used to characterize the features of each sample cell.

In an exemplary embodiment, the features of the sample cells represented by the cell feature matrix may include gene features and adjacency relationship features of the sample cells.

In an exemplary embodiment, the cell feature matrix can be obtained based on the gene expression matrix and adjacency matrix of the sample cells. The gene expression matrix of the sample cells is a matrix used to characterize the genes of each sample cell. The adjacency matrix of the sample cells is a matrix used to characterize the adjacency relationship between each sample cell.

In an exemplary embodiment, the adjacency matrix of the sample cells can be determined based on the spatial position matrix of the sample cells. The adjacency matrix of the sample cells is used to characterize the distance similarity between each sample cell. The computer device can determine the distance similarity between each sample cell based on the spatial position matrix of the sample cells, and obtain the adjacency matrix of the sample cells based on the distance similarity. Among them, the spatial position matrix of the sample cells is a matrix used to characterize the spatial position of each sample cell.

In an exemplary embodiment, the rows in the gene expression matrix of the sample cells represent the sample cells, and the columns represent the genes. The rows in the spatial position matrix of the sample cells represent the sample cells, and the columns represent the coordinate values of the sample cells. In other embodiments, it can also be the other way around, the columns in the gene expression matrix of the sample cells represent the sample cells, and the rows represent the genes. The columns in the spatial position matrix of the sample cells represent the sample cells, and the rows represent the coordinate values of the sample cells.

In an exemplary embodiment, each row in the adjacency matrix of the sample cell represents a sample cell node and each For example, the i-th row represents the distance similarity between sample cell node i and each sample cell node, and the i-th row and j-th column represent the distance similarity between sample cell node i and sample cell node j.

In an exemplary embodiment, the computer device may perform different perturbations on the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix.

Step 104 , perturb the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix.

In an exemplary embodiment, the computer device may perform different perturbations on the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix.

Step 106, encoding the first graph through the feature encoder to be trained to obtain a first embedding matrix, and encoding the second graph to obtain a second embedding matrix; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix.

In an exemplary embodiment, the first Among them, _X1 represents the first characteristic matrix, and A _m represents the first adjacency matrix. Among them, _X2 represents the second characteristic matrix, and _Ad represents the second adjacency matrix.

In an exemplary embodiment, the computer device may input the first image and the second image into corresponding feature encoders to be trained for encoding, and obtain a first embedding matrix and a second embedding matrix, respectively. The parameters of the feature encoders to be trained corresponding to the first image and the second image are shared.

Step 108: The first embedding matrix and the second embedding matrix are fused to obtain a target embedding matrix.

In an exemplary embodiment, the computer device may linearly add the first embedding matrix and the second embedding matrix to obtain a target embedding matrix.

Step 110, decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix.

In an exemplary embodiment, the computer device may input the target embedding matrix into a decoder for decoding to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix.

In an exemplary embodiment, the computer device may input the target embedding matrix into a decoder for decoding to obtain a reconstructed cell feature matrix, and then determine a reconstructed adjacency matrix based on the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix. In an exemplary embodiment, the computer device may perform an inner product of the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix to obtain a reconstructed adjacency matrix.

In an exemplary embodiment, the decoder may be a graph convolutional neural network based decoder.

Step 112, iteratively optimizing the feature encoder to be trained according to at least one of the first difference and the second difference Parameters are adjusted until the iteration stop condition is met to obtain a trained feature encoder; the feature encoder is used to extract the features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

In an exemplary embodiment, the computer device may determine a reconstruction loss value based on at least one of the first difference and the second difference, and iteratively optimize the parameters of the feature encoder to be trained based on the reconstruction loss value until an iteration stop condition is met, thereby obtaining a trained feature encoder.

In an exemplary embodiment, the computer device may determine a feature matrix reconstruction loss value based on the first difference, determine an adjacency matrix reconstruction loss value based on the second difference, and determine a reconstruction loss value based on the sum of the feature matrix reconstruction loss value and the adjacency matrix reconstruction loss value.

In an exemplary embodiment, the computer device may determine the feature matrix reconstruction loss value based on the 2 norm of the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cell. In an exemplary embodiment, the calculation formula of the feature matrix reconstruction loss value L _REC-F may be:

Where N is the number of sample cells and X is the cell feature matrix of the sample cells. is the reconstructed cell feature matrix.

In an exemplary embodiment, the computer device may determine the adjacency matrix reconstruction loss value according to the 2-norm of the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cell. In an exemplary embodiment, the calculation formula of the adjacency matrix reconstruction loss value L _REC-A may be:

Where N is the number of sample cells and A is the adjacency matrix of sample cells. is the reconstructed adjacency matrix.

In an exemplary embodiment, the reconstruction loss value L _REC may be calculated using the following formula:
L _REC = L _REC-F + L _REC-A

In an exemplary embodiment, the iteration stopping condition may include that the reconstruction loss value is less than or equal to a first preset loss threshold. In some other embodiments, the iteration stopping condition may include that the number of iterations is greater than or equal to a preset number threshold.

The above-mentioned cell clustering feature extraction method perturbs the cell feature matrix of the sample cells to obtain the first feature matrix and the second feature matrix, and perturbs the adjacency matrix of the sample cells to obtain the first adjacency matrix and the second adjacency matrix, which can obtain richer features under different views of the graph, and then encodes the first graph and the second graph of different views with rich information respectively through the twin feature encoder to be trained, and fuses the encoding results of the first graph and the second graph The target embedding matrix is obtained by combining the target embedding matrix so that the target embedding matrix can have richer and deeper features. The target embedding matrix is then decoded to obtain the reconstructed cell feature matrix and the reconstructed adjacency matrix to determine the reconstruction loss. The feature encoder can be accurately trained to obtain the features required for cell clustering, so that the feature encoder that can extract deep features can accurately extract the features required for cell clustering.

In an exemplary embodiment, perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix includes: acquiring a first noise matrix and a second noise matrix; and perturbing the cell feature matrix of the sample cells according to the first noise matrix and the second noise matrix, respectively, to obtain a first feature matrix and a second feature matrix.

In an exemplary embodiment, the computer device may obtain the first noise matrix and the second noise matrix by randomly sampling from a Gaussian distribution.

In an exemplary embodiment, the computer device may multiply the first noise matrix with the cell feature matrix of the sample cells to obtain a first feature matrix, and multiply the second noise matrix with the cell feature matrix of the sample cells to obtain a second feature matrix.

In the above embodiment, the cell feature matrix of the sample cells is perturbed according to the first noise matrix and the second noise matrix respectively to obtain the first feature matrix and the second feature matrix, so that a richer feature matrix can be obtained.

In an exemplary embodiment, perturbing the adjacency matrix of sample cells to obtain a first adjacency matrix and a second adjacency matrix includes: determining the distance similarity between each sample cell node according to the adjacency matrix of the sample cells; deleting the edges connecting the sample cell nodes whose distance similarity meets a preset condition from the adjacency matrix of the sample cells to obtain the first adjacency matrix; determining the importance of each sample cell node according to the adjacency matrix of the sample cells, and adjusting the distance similarity between each sample cell node according to the importance to obtain the second adjacency matrix.

Among them, one sample cell node corresponds to one sample cell.

In an exemplary embodiment, the edges connected between sample cell nodes of a preset condition may be edges connected between sample cells whose distance similarity between sample cells is less than or equal to a preset similarity threshold, or edges connected between sample cells corresponding to a preset number of distance similarities in ascending order, or edges connected between sample cells corresponding to a preset percentage of distance similarities in ascending order. For example, edges connected between sample cells corresponding to the top 10% of distance similarities in ascending order may be deleted.

In an exemplary embodiment, the computer device can determine the importance of each sample cell node according to the adjacency matrix of the sample cells through the personalized PageRank (PPR) algorithm, and adjust the distance similarity between each sample cell node according to the importance to obtain a second adjacency matrix.

In the above embodiment, the edges between the sample cell nodes whose distance similarity meets the preset conditions are deleted from the adjacency matrix of the sample cells to obtain a first adjacency matrix, the importance of each sample cell node is determined according to the adjacency matrix of the sample cells, and the distance similarity between each sample cell node is adjusted according to the importance to obtain a second adjacency matrix, so that By introducing the personalized PageRank (PPR) algorithm to generate the second adjacency matrix, the remote information capture capability of the shallow network structure model is improved, which further improves the clustering capability.

In an exemplary embodiment, before perturbing the cell feature matrix of the sample cells to obtain the first feature matrix and the second feature matrix, the method also includes: acquiring a spatial position matrix and a gene expression matrix of the sample cells; the spatial position matrix is used to characterize the spatial position of each sample cell; the gene expression matrix is used to characterize the genes of each sample cell; based on the spatial position matrix, determining the adjacency matrix of the sample cells; and combining the adjacency matrix and the gene expression matrix to obtain the cell feature matrix of the sample cells.

In an exemplary embodiment, the computer device may concatenate the adjacency matrix and the gene expression matrix to obtain a cell feature matrix of the sample cells.

For example: Assume that the gene expression matrix is an N*M matrix, and the adjacency matrix is an N*N matrix. Wherein, N represents the number of sample cells, and M represents the number of genes. The splicing combination method is: gene expression matrix | adjacency matrix, that is, the adjacency matrix is placed on the right side of the gene expression matrix for splicing, or the adjacency matrix can be placed on the left side of the gene expression matrix for splicing. In other embodiments, assuming that the gene expression matrix is an M*N matrix, and the adjacency matrix is an N*N matrix, the adjacency matrix can be placed below or above the gene expression matrix for splicing.

In the above embodiment, the adjacency matrix of the sample cells is determined according to the spatial position matrix, and the adjacency matrix and the gene expression matrix are combined to obtain the cell feature matrix of the sample cells, thereby obtaining various feature information of the sample cells.

In an exemplary embodiment, encoding a first image by a feature encoder to be trained to obtain a first embedding matrix, and encoding a second image to obtain a second embedding matrix includes: taking the first image and the second image as images to be encoded, respectively, inputting the feature matrix and the adjacency matrix in the image to be encoded into the feature encoder to be trained corresponding to the image to be encoded, so as to encode the image to be encoded layer by layer; sharing parameters between the feature encoders to be trained corresponding to the first image and the second image; in the process of encoding the image to be encoded layer by layer, taking the first encoding layer of the feature encoder to be trained as the current layer, and taking the feature matrix input to the first encoding layer as the embedding matrix input to the current layer, and according to the embedding matrix and the adjacency matrix input to the current layer, The product of the normalized matrices determines the fusion matrix of the current layer; the fusion matrix of the current layer and the embedding matrix input to the current layer are weightedly fused according to the weights in the current layer to obtain the embedding matrix output by the current layer; the embedding matrix output by the current layer is used as the input of the next layer, and the next layer is used as the new current layer; the step of determining the fusion matrix of the current layer according to the product of the normalized matrix of the embedding matrix input to the current layer and the adjacency matrix and subsequent steps are returned to execute; the embedding matrix output by the last layer in the feature encoder to be trained is used as the embedding matrix corresponding to the image to be encoded; wherein, when the image to be encoded is the first image, the corresponding embedding matrix is the first embedding matrix; when the image to be encoded is the second image, the corresponding embedding matrix is the second embedding matrix.

In an exemplary embodiment, the computer device may weight the product of the embedding matrix of the current layer and the normalized matrix of the adjacency matrix through the weight matrix of the current layer to obtain a first result, and weight the product of the embedding matrix of the current layer and the normalized matrix of the adjacency matrix through the weight matrix of the current layer to obtain a first result. The embedding matrix of the layer is weighted to obtain the second result, and then the nonlinear activation function value of the first result and the nonlinear activation function value of the second result are added to obtain the embedding matrix output by the current layer.

In an exemplary embodiment, the formula for the feature encoder layer l is as follows:

in, A _m and A _d are the first adjacency matrix and the second adjacency matrix, respectively. _{D m} and D _d are the degree matrices of the first adjacency matrix A _m and the second adjacency matrix A _d , respectively. I is the identity matrix. and are the normalized matrices of the first adjacency matrix A _m and the second adjacency matrix A _d, respectively. and is the weight matrix of the feature encoder layer l. ^{b (l)} is the bias vector of the feature encoder layer l. σ is the nonlinear activation function. For example, the nonlinear activation function can be ReLU or Tanh. is the encoding result of the lth layer of the feature encoder corresponding to the first image. It is the encoding result of the l-1th layer of the feature encoder corresponding to the first figure. is the encoding result of the lth layer of the feature encoder corresponding to the second figure. is the encoding result of the feature encoder layer l-1 corresponding to the second figure. When the number of encoder layers l = 1, is the first input image, The second image is input.

In the above embodiment, since there is no need to mine negative samples, positive samples only need to be encoded through the encoder of the twin network structure. Therefore, compared with the method of using contrastive learning, the space utilization is reduced, and it is achieved that while learning richer feature representations, the space utilization is reduced.

In an exemplary embodiment, decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix includes: inputting the adjacency matrix of the sample cells and the target embedding matrix into the decoder to decode the target embedding matrix layer by layer; in the process of decoding the target embedding matrix layer by layer, taking the first decoding layer of the decoder as the current layer, and taking the target embedding matrix input to the first decoding layer as the input of the current layer; weighting the normalized product of the input of the current layer and the adjacency matrix of the sample cells to obtain the decoding result output by the current layer, taking the decoding result output by the current layer as the input of the next layer, and taking the next layer as the new current layer, returning to execute the step of weighting the normalized product of the input of the current layer and the adjacency matrix of the sample cells to obtain the decoding result output by the current layer and the subsequent steps, taking the decoding result output by the last layer in the decoder as the reconstructed cell feature matrix; determining the reconstructed adjacency matrix according to the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix.

In an exemplary embodiment, during the decoding process, the computer device can weight the normalized product of the input of the current layer and the adjacency matrix of the sample cell with the weight matrix of the current layer, and then take the nonlinear activation function value of the weighted result to obtain the decoding result output by the current layer.

In an exemplary embodiment, the calculation formula of the decoding result of the k-th layer decoder is:

Where H ^(k) is the decoding result of the kth layer of the decoder. ^{H (k-1)} is the decoding result of the k-1th layer of the decoder. σ is a nonlinear activation function. A is the adjacency matrix of the sample cell. I is the identity matrix, and D is the degree matrix of A. is the normalized value of A. W ^(k) is the parameter matrix of the kth layer of the decoder.

In an exemplary embodiment, the computer device may perform an inner product on the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix to obtain a reconstructed adjacency matrix.

In the above embodiment, the decoder performs layer-by-layer decoding to obtain a reconstructed cell feature matrix and a reconstructed adjacency matrix, so that the reconstruction loss can be accurately calculated. According to the reconstruction loss value, the low-dimensional feature embedding can take into account both the gene expression information and the spatial position information, thereby improving the accuracy of the trained feature encoder.

In an exemplary embodiment, iteratively optimizing the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference includes: determining a reconstruction loss value according to at least one of the first difference and the second difference; determining a target loss value according to at least one of a clustering guidance loss value and a de-redundancy loss value, and the reconstruction loss value; iteratively optimizing the parameters of the feature encoder to be trained according to the target loss value; wherein the clustering guidance loss value is determined according to the difference between the soft allocation matrix and the target distribution matrix; the soft allocation matrix is determined according to the clustering result obtained by clustering based on the target embedding matrix; the target distribution matrix is obtained by normalizing the soft allocation matrix; and the de-redundancy loss value is determined according to the difference between the first embedding matrix and the second embedding matrix.

In an exemplary embodiment, the computer device may determine the target loss value according to the sum of at least one of the clustering guidance loss value and the de-redundancy loss value and the reconstruction loss value.

In an exemplary embodiment, the calculation formula of the target loss value L may be:
L＝L _REC +L _C +L _RR

Where L _REC is the reconstruction loss value, L _C is the clustering guidance loss value, and L _RR is the redundancy removal loss value.

In an exemplary embodiment, the iteration stopping condition may be that the target loss value is less than or equal to a second preset loss threshold. In other embodiments, the iteration stopping condition may include that the number of iterations is greater than or equal to a preset number threshold.

In the above embodiment, the target loss value is determined according to at least one of the clustering guidance loss value and the de-redundancy loss value, and the reconstruction loss value. The feature embedding related to the clustering task can be effectively learned according to the clustering guidance loss value. The redundant information in the embedding can be eliminated according to the de-redundancy loss value, and a distinguishable embedding is generated for each sample cell node. The low-dimensional feature embedding can take into account both gene expression information and spatial position information according to the reconstruction loss value, thereby further This further improves the accuracy of the trained feature encoder.

In an exemplary embodiment, the step of determining the clustering guidance loss value includes: clustering according to the target embedding matrix to obtain the reference cluster center; determining the soft assignment matrix between the vectors corresponding to each sample cell in the target embedding matrix and the reference cluster center; the soft assignment matrix is used to characterize the probability that the vector is assigned to each reference cluster center; normalizing the soft assignment matrix to generate a target distribution matrix; and determining the clustering guidance loss value based on the difference between the distribution of the soft assignment matrix and the target distribution matrix.

Among them, each row in the target embedding matrix is a vector corresponding to a sample cell.

In an exemplary embodiment, the computer device may cluster the vectors corresponding to each sample cell in the target embedding matrix to obtain reference cluster centers.

In an exemplary embodiment, the computer device may use Student's t distribution to calculate the soft assignment matrix between the vectors corresponding to each sample cell in the target embedding matrix and the reference cluster center.

In an exemplary embodiment, the element p _ij in the i-th row and j-th column of the target distribution matrix P can be calculated by the following formula:

Wherein, _qij is the element in the i-th row and j-th column in the soft allocation matrix Q. j' represents the index of a column in the soft allocation matrix.

In an exemplary embodiment, the computer device may calculate the clustering guidance loss value using KL divergence according to the soft assignment matrix and the target distribution matrix. The formula is as follows:

In the above embodiment, a soft assignment distribution and a target distribution are generated based on the clustering result of the target embedding matrix, and then the two distributions are aligned by using a clustering guidance loss value to guide network learning, thereby improving the accuracy of the feature encoder and being able to effectively learn feature embeddings related to the clustering task.

In an exemplary embodiment, the de-redundancy loss value includes a first de-redundancy loss value; the step of determining the de-redundancy loss value includes: determining a node similarity matrix based on the similarity between the first embedding matrix and the second embedding matrix; and determining the first de-redundancy loss value based on the difference between the node similarity matrix and the unit matrix.

In an exemplary embodiment, the similarity may be cosine similarity. The node similarity matrix _SN may be calculated by the following formula:

Where _H1 is the first embedding matrix. _H2 is the second embedding matrix. _H2T represents the transpose of _H2 , ^and || || represents the modulus operation of the matrix.

In an exemplary embodiment, the computer device may calculate the MSE of the node similarity matrix and the identity matrix to determine the first de-redundancy loss value. The first de-redundancy loss value L _RR-N may be calculated by the following formula:

Where N is the number of sample cells. S _N is the node similarity matrix. I is the identity matrix.

In the above embodiment, a node similarity matrix is determined based on the similarity between the first embedding matrix and the second embedding matrix, and a first de-redundancy loss value is determined based on the difference between the node similarity matrix and the unit matrix. This decorrelation operation can enable the network to reduce redundant information between sample cell nodes in the latent space, thereby making the learned embedding more discriminative, enhancing the robustness of the feature encoder, and solving the problem of representation collapse existing in traditional methods.

In an exemplary embodiment, the de-redundancy loss value also includes a second de-redundancy loss value; the step of determining the de-redundancy loss value also includes: mapping the first embedding matrix and the second embedding matrix to cluster-level embeddings respectively to obtain a first cluster-level embedding matrix and a second cluster-level embedding matrix; determining a cluster-level similarity matrix based on the similarity between the first cluster-level embedding matrix and the second cluster-level embedding matrix; determining a second de-redundancy loss value based on the difference between the cluster-level similarity matrix and the unit matrix.

Among them, cluster-level embedding is obtained by clustering the vectors of each sample cell node in the embedding matrix and averaging the vectors in the same cluster.

In an exemplary embodiment, a computer device may use a read function The first embedding matrix and the second embedding matrix are mapped to the cluster-level embedding to obtain a first cluster-level embedding matrix and a second cluster-level embedding matrix.

In an exemplary embodiment, the similarity may be cosine similarity. The cluster-level similarity matrix _SF may be calculated by the following formula:

Where _Z1 is the first cluster-level embedding matrix and _Z2 is the second cluster-level embedding matrix.

In an exemplary embodiment, the computer device may calculate the MSE of the cluster-level similarity matrix and the identity matrix to determine the second de-redundancy loss value. The second de-redundancy loss value L _RR-F may be calculated by the following formula:

Where D is the dimension of the target embedding matrix H. _{S F} is the cluster-level similarity matrix. I is the identity matrix.

In an exemplary embodiment, the computer device may determine the redundancy loss value according to the sum of the first redundancy loss value and the second redundancy loss value. The calculation formula of the redundancy loss value L _RR is:
_LRR ＝LRR _-N +LRR _-F

In the above embodiment, mapping the node embedding to the cluster-level embedding can reduce redundant information at the feature level, thereby making the learned embedding more discriminative.

In an exemplary embodiment, the method further includes: inputting a graph consisting of a cell feature matrix and an adjacency matrix of cells to be grouped into a trained feature encoder to obtain an extracted embedding matrix; clustering the embedding vectors corresponding to each cell to be grouped in the extracted embedding matrix to obtain a cell grouping result of the cells to be grouped.

The extracted embedding matrix refers to the embedding matrix outputted after the feature encoder encodes the input graph. The cell clustering result of the cells to be clustered refers to the result of dividing the cells to be clustered into multiple cell clusters.

In an exemplary embodiment, the cells to be grouped may be cells other than sample cells. In other embodiments, the cells to be grouped may be sample cells themselves.

In an exemplary embodiment, each row in the extracted embedding matrix is an embedding vector corresponding to a cell to be grouped. The computer device can cluster the embedding vectors corresponding to each cell to be grouped in the extracted embedding matrix, and divide the cells to be grouped corresponding to the vectors belonging to the same cluster in the clustering result into the same cell group.

In the above embodiment, the embedding matrix of the cells to be clustered is extracted by a feature encoder that is trained to extract deep features, and then clustering is performed based on the extracted embedding matrix that has learned the deep features, which can improve the accuracy of cell clustering.

In an exemplary embodiment, the method further includes: clustering the embedding vectors corresponding to each sample cell in the target embedding matrix to obtain a cell clustering result of the sample cells.

The cell clustering result of the sample cells refers to the result of dividing the sample cells into a plurality of cell clusters.

In the above embodiment, when a group of sample cells needs to be clustered, a feature encoder that can extract deep-level features can be first trained based on the cell feature matrix and adjacency matrix of the sample cells, so that the obtained target embedding matrix can reflect the deep-level features, thereby clustering can be directly performed according to the target embedding matrix to obtain accurate cell clustering results for the sample cells.

As shown in FIG2 , it is a schematic diagram of the overall process of the cell clustering feature extraction method in each embodiment of the present application. First, the cell feature matrix of the sample cells is perturbed using the first noise matrix N ₁ and the second noise matrix N ₂ to obtain the first feature matrix X ₁ and the second feature matrix X ₂ , and the adjacency matrix A is perturbed differently to obtain the first adjacency matrix A _m and the second adjacency matrix A _d . And the second picture Input to the parameter-sharing feature encoder to be trained, obtain the first embedding matrix _H1 and the second embedding matrix _H2 , and transform the first embedding matrix The matrix _H1 and the second embedding matrix _H2 are linearly added to obtain the target embedding matrix H. Then, the target embedding matrix H is input into the decoder to obtain the reconstructed cell feature matrix and the reconstructed adjacency matrix The reconstruction loss value L _REC is determined. The redundancy removal loss value L _RR is determined according to the similarity between the first embedding matrix H ₁ and the second embedding matrix H _2. The clustering guidance loss value _LC is determined according to the result of clustering the target embedding matrix H. The parameters of the feature encoder are optimized according to the reconstruction loss value L _REC , the redundancy removal loss value L _RR and the clustering guidance loss value _LC , and finally a trained feature encoder is obtained. The feature encoder can be used to extract the features required for cell clustering.

In an exemplary embodiment, as shown in FIG3 , a cell clustering method is provided, and this embodiment is illustrated by applying the method to a computer device. The computer device may be a terminal or a server. The terminal may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server may be implemented as an independent server or a server cluster consisting of multiple servers. In this embodiment, the method comprises the following steps:

Step 302, input the graph composed of the cell feature matrix and the adjacency matrix of the cells to be grouped into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to the first graph and the second graph; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of the sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the adjacency matrix of the sample cells.

Step 304 , clustering the embedding vectors corresponding to the cells to be grouped in the extracted embedding matrix to obtain the cell grouping results of the cells to be grouped.

The above-mentioned cell clustering method extracts the embedding matrix of the cells to be clustered through a feature encoder that has been trained to extract deep features, and then performs clustering based on the extracted embedding matrix that has learned deep features, thereby improving the accuracy of cell clustering.

In an exemplary embodiment, the training steps of the feature encoder include: perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix; perturbing the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix; encoding the first graph through the feature encoder to be trained to obtain a first embedding matrix, and encoding the second graph to obtain a second embedding matrix; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; fusing the first embedding matrix and the second embedding matrix to obtain a target embedding matrix; decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; iteratively optimizing the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference. until the iteration stop condition is met, and a trained feature encoder is obtained; the feature encoder is used to extract the features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

In the above embodiment, the cell feature matrix of the sample cells is perturbed to obtain the first feature matrix and the second feature matrix, and the adjacency matrix of the sample cells is perturbed to obtain the first adjacency matrix and the second adjacency matrix, so that richer features can be obtained under different views of the graph, and then the first graph and the second graph of different views with rich information are respectively encoded by the twin feature encoder to be trained, and the encoding results of the first graph and the second graph are fused to obtain the target embedding matrix, so that the target embedding matrix can have richer and deeper features, and then the target embedding matrix is decoded to obtain the reconstructed cell feature matrix and the reconstructed adjacency matrix to determine the reconstruction loss, so that the feature encoder can be accurately trained, so that the features required for cell clustering can be accurately extracted through the feature encoder that can extract deep features.

The model evaluation test of the feature encoder was conducted on the data of four stages of mouse embryonic development: E9.5 E1S1, E9.5 E2S2, E9.5 E2S3 and E9.5 E2S4, and human breast cancer (10x Visium) data. Widely used evaluation indicators were used: Adjusted Rand index (ARI), with a range of [-1, 1], the closer to 1, the better; Normalized Mutual Information (NMI), with a range of [0, 1], the closer to 1, the better; Fowlkes-Mallows index (FMI), with a range of [0, 1], the closer to 1, the better. These three indicators are used to measure the degree of fit between two distributions. The larger the value, the more consistent the clustering effect is with the actual situation.

Table 1. Test results of the model on the E9.5 E1S1 stage data of mouse embryo

It can be seen from Table 1 that the algorithm proposed in this application is far superior to the existing algorithm model in most indicators.

Table 2. Test results of the model on the E9.5 E2S2 stage data of mouse embryo

It can be seen from Table 2 that the algorithm proposed in this application is far superior to the existing algorithm model in all indicators.

Table 3. Test results of the model on the E9.5 E2S3 stage data of mouse embryo

It can be seen from Table 3 that the application proposed by this patent is far superior to the existing algorithm model in all indicators.

Table 4. Test results of the model on the E9.5 E2S4 stage data of mouse embryo

It can be seen from Table 4 that the algorithm proposed in this application is far superior to the existing algorithm model in all indicators.

Table 5. Test results of the model on human breast cancer (10x Visium) data

It can be seen from Table 5 that the algorithm proposed in this application is far superior to the existing algorithm model in all indicators.

It should be understood that, although the steps in the flowcharts of the above embodiments are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. The steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or at least part of the steps or stages in other steps.

Based on the same inventive concept, the embodiment of the present application also provides a cell clustering feature extraction device for implementing the cell clustering feature extraction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more cell clustering feature extraction device embodiments provided below can refer to the limitations of the cell clustering feature extraction method above, and will not be repeated here.

In an exemplary embodiment, as shown in FIG4 , a cell clustering feature extraction device 400 is provided, comprising: a first perturbation module 402, a second perturbation module 404, an encoding module 406, a fusion module 408, a decoding module 410 and a parameter optimization module 412, wherein:

The first perturbation module 402 is used to perturb the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix.

The second perturbation module 404 is used to perturb the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix.

The encoding module 406 is used to encode the first image through the feature encoder to be trained to obtain a first embedding matrix, and encode the second image to obtain a second embedding matrix; the first image includes a first feature matrix and a first adjacency matrix; the second image includes a second feature matrix and a second adjacency matrix.

A fusion module 408, configured to fuse the first embedding matrix and the second embedding matrix to obtain a target embedding matrix;

The decoding module 410 is used to decode the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix.

The parameter optimization module 412 is used to iteratively optimize the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference until the iteration stop condition is met to obtain a trained feature encoder; the feature encoder is used to extract the features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

In an exemplary embodiment, the first perturbation module 402 is further used to obtain a first noise matrix and a second noise matrix; and to perturb the cell feature matrix of the sample cells according to the first noise matrix and the second noise matrix, respectively, to obtain a first feature matrix and a second feature matrix.

In an exemplary embodiment, the second perturbation module 404 is further used to determine the distance similarity between each sample cell node according to the adjacency matrix of the sample cells; delete the edges connected between the sample cell nodes whose distance similarity meets the preset conditions from the adjacency matrix of the sample cells to obtain a first adjacency matrix; determine the importance of each sample cell node according to the adjacency matrix of the sample cells, and adjust the distance similarity between each sample cell node according to the importance to obtain a second adjacency matrix. Connect the matrix.

In an exemplary embodiment, the first perturbation module 402 is also used to obtain a spatial position matrix and a gene expression matrix of sample cells; the spatial position matrix is used to characterize the spatial position of each sample cell; the gene expression matrix is used to characterize the genes of each sample cell; based on the spatial position matrix, the adjacency matrix of the sample cells is determined; the adjacency matrix and the gene expression matrix are combined to obtain a cell feature matrix of the sample cells.

In an exemplary embodiment, the encoding module 406 is further used to use the first image and the second image as images to be encoded, respectively, and input the feature matrix and the adjacency matrix in the image to be encoded into the feature encoder to be trained corresponding to the image to be encoded, so as to encode the image to be encoded layer by layer; parameters are shared between the feature encoders to be trained corresponding to the first image and the second image; in the process of encoding the image to be encoded layer by layer, the first encoding layer of the feature encoder to be trained is used as the current layer, and the feature matrix input to the first encoding layer is used as the embedding matrix input to the current layer, and the fusion matrix of the current layer is determined according to the product of the embedding matrix input to the current layer and the normalized matrix of the adjacency matrix; The fusion matrix of the current layer and the embedding matrix input to the current layer are weightedly fused according to the weights in the current layer to obtain the embedding matrix output by the current layer, the embedding matrix output by the current layer is used as the input of the next layer, and the next layer is used as the new current layer, and the step of determining the fusion matrix of the current layer according to the product of the normalized matrix of the embedding matrix input to the current layer and the adjacency matrix and subsequent steps are returned to execute, and the embedding matrix output by the last layer in the feature encoder to be trained is used as the embedding matrix corresponding to the image to be encoded; wherein, when the image to be encoded is the first image, the corresponding embedding matrix is the first embedding matrix; when the image to be encoded is the second image, the corresponding embedding matrix is the second embedding matrix.

In an exemplary embodiment, the method is also used to input the adjacency matrix of sample cells and the target embedding matrix into the decoder to decode the target embedding matrix layer by layer; in the process of decoding the target embedding matrix layer by layer, the first decoding layer of the decoder is used as the current layer, and the target embedding matrix input to the first decoding layer is used as the input of the current layer; the normalized product of the input of the current layer and the adjacency matrix of the sample cells is weighted to obtain the decoding result output by the current layer, the decoding result output by the current layer is used as the input of the next layer, and the next layer is used as the new current layer, and the step of weighting the normalized product of the input of the current layer and the adjacency matrix of the sample cells to obtain the decoding result output by the current layer and the subsequent steps are returned, and the decoding result output by the last layer in the decoder is used as the reconstructed cell feature matrix; the reconstructed adjacency matrix is determined according to the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix.

In an exemplary embodiment, the parameter optimization module 412 is also used to determine a reconstruction loss value based on at least one of the first difference and the second difference; determine a target loss value based on at least one of the clustering guidance loss value and the de-redundancy loss value, and the reconstruction loss value; iteratively optimize the parameters of the feature encoder to be trained based on the target loss value; wherein the clustering guidance loss value is determined based on the difference between the soft allocation matrix and the target distribution matrix; the soft allocation matrix is determined based on the clustering result obtained by clustering based on the target embedding matrix; the target distribution matrix is obtained by normalizing the soft allocation matrix; and the de-redundancy loss value is determined based on the difference between the first embedding matrix and the second embedding matrix.

In an exemplary embodiment, the parameter optimization module 412 is also used to perform clustering according to the target embedding matrix to obtain reference cluster centers; determine the soft assignment matrix between the vectors corresponding to each sample cell in the target embedding matrix and the reference cluster centers; the soft assignment matrix is used to characterize the probability that the vectors are assigned to each reference cluster center; normalize the soft assignment matrix to generate a target distribution matrix; and determine the clustering guidance loss value based on the difference between the distribution of the soft assignment matrix and the target distribution matrix.

In an exemplary embodiment, the parameter optimization module 412 is further configured to determine a node similarity matrix according to the similarity between the first embedding matrix and the second embedding matrix; and determine a first de-redundancy loss value according to the difference between the node similarity matrix and the identity matrix.

In an exemplary embodiment, the parameter optimization module 412 is also used to map the first embedding matrix and the second embedding matrix into cluster-level embeddings, respectively, to obtain a first cluster-level embedding matrix and a second cluster-level embedding matrix; determine a cluster-level similarity matrix based on the similarity between the first cluster-level embedding matrix and the second cluster-level embedding matrix; and determine a second de-redundancy loss value based on the difference between the cluster-level similarity matrix and the unit matrix.

In an exemplary embodiment, as shown in FIG. 5 , the cell clustering feature extraction device 400 further includes: a clustering module 414 .

The encoding module 406 is also used to input the graph composed of the cell feature matrix and the adjacency matrix of the cells to be grouped into the trained feature encoder to obtain the extracted embedding matrix.

The clustering module 414 is also used to cluster the embedding vectors corresponding to each of the cells to be grouped in the extracted embedding matrix to obtain the cell grouping results of the cells to be grouped.

In an exemplary embodiment, the clustering module 414 is further used to cluster the embedding vectors corresponding to each sample cell in the target embedding matrix to obtain a cell clustering result of the sample cells.

In an exemplary embodiment, as shown in FIG6 , a cell clustering device 600 is provided, comprising: an embedding matrix determination module 602 and a clustering module 604, wherein:

The embedding matrix determination module 602 is used to input the graph composed of the cell feature matrix and the adjacency matrix of the cells to be clustered into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to the first graph and the second graph; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of the sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the adjacency matrix of the sample cells.

The clustering module 604 is used to cluster the embedding vectors corresponding to the cells to be clustered in the extracted embedding matrix to obtain the cell clustering results of the cells to be clustered.

In an exemplary embodiment, the cell clustering device 600 further includes: a model training module 606 .

The model training module 606 is also used to perturb the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix; perturb the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix; encode the first graph through the feature encoder to be trained to obtain a first embedding matrix, and encode the second graph to obtain a second embedding matrix; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; fuse the first embedding matrix and the second embedding matrix to obtain a target embedding matrix; decode the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; iteratively optimize the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference until the iteration stop condition is met to obtain a trained feature encoder; the feature encoder is used to extract the features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells.

The above-mentioned cell clustering feature extraction device and each module in the cell clustering device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be shown in FIG7 . The computer device includes a processor, a memory, and a network interface connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a cell clustering feature extraction method or a cell clustering method is implemented.

In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in FIG8 . The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be implemented through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a cell clustering feature extraction method or a cell clustering method is implemented. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse. Standard, etc.

Those skilled in the art will understand that the structure shown in FIG. 7 or FIG. 8 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, etc., but are not limited to this.

The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims (18)

A method for extracting cell clustering features, characterized in that the method comprises: Perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix; Perturbing the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix; The first graph is encoded by a feature encoder to be trained to obtain a first embedding matrix, and the second graph is encoded to obtain a second embedding matrix; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix; Fusing the first embedding matrix and the second embedding matrix to obtain a target embedding matrix; Decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; According to at least one of the first difference and the second difference, the parameters of the feature encoder to be trained are iteratively optimized until the iteration stop condition is met, thereby obtaining a trained feature encoder; the feature encoder is used to extract features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells. The method according to claim 1, characterized in that the perturbing the cell characteristic matrix of the sample cells to obtain the first characteristic matrix and the second characteristic matrix comprises: Obtain a first noise matrix and a second noise matrix; The cell feature matrix of the sample cells is perturbed according to the first noise matrix and the second noise matrix respectively to obtain a first feature matrix and a second feature matrix. The method according to claim 1 or 2, characterized in that the perturbing the adjacency matrix of the sample cells to obtain the first adjacency matrix and the second adjacency matrix comprises: Determine the distance similarity between each sample cell node according to the adjacency matrix of the sample cell; Deleting the edges connecting the sample cell nodes whose distance similarity meets the preset conditions from the adjacency matrix of the sample cells to obtain a first adjacency matrix; The importance of each of the sample cell nodes is determined according to the adjacency matrix of the sample cells, and the distance similarity between each of the sample cell nodes is adjusted according to the importance to obtain a second adjacency matrix. The method according to any one of claims 1 to 3, characterized in that before perturbing the cell feature matrix of the sample cells to obtain the first feature matrix and the second feature matrix, the method further comprises: Obtain the spatial position matrix and gene expression matrix of the sample cells; the spatial position matrix is used to characterize each The spatial position of the sample cells; the gene expression matrix is used to characterize the genes of each of the sample cells; Determining an adjacency matrix of the sample cells according to the spatial position matrix; The adjacency matrix and the gene expression matrix are combined to obtain a cell feature matrix of the sample cells. The method according to any one of claims 1 to 4, characterized in that encoding the first image by the feature encoder to be trained to obtain the first embedding matrix, and encoding the second image to obtain the second embedding matrix comprises: The first image and the second image are respectively used as images to be encoded, and the feature matrix and the adjacency matrix in the image to be encoded are input into the feature encoder to be trained corresponding to the image to be encoded, so as to encode the image to be encoded layer by layer; the parameters of the feature encoders to be trained corresponding to the first image and the second image are shared; In the process of encoding the to-be-encoded graph layer by layer, a first encoding layer of the to-be-trained feature encoder is used as a current layer, and a feature matrix input to the first encoding layer is used as an embedding matrix input to the current layer, and a fusion matrix of the current layer is determined according to a product of the embedding matrix input to the current layer and a normalized matrix of the adjacency matrix; The fusion matrix of the current layer and the embedding matrix input to the current layer are weightedly fused according to the weights in the current layer to obtain the embedding matrix output by the current layer, the embedding matrix output by the current layer is used as the input of the next layer, and the next layer is used as the new current layer, and the step of determining the fusion matrix of the current layer according to the product of the embedding matrix input to the current layer and the normalized matrix of the adjacency matrix and subsequent steps are returned to execute, and the embedding matrix output by the last layer in the feature encoder to be trained is used as the embedding matrix corresponding to the image to be encoded; Among them, when the image to be encoded is the first image, the corresponding embedding matrix is the first embedding matrix; when the image to be encoded is the second image, the corresponding embedding matrix is the second embedding matrix. The method according to any one of claims 1 to 5, characterized in that the decoding of the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix comprises: Inputting the adjacency matrix of the sample cells and the target embedding matrix into a decoder to decode the target embedding matrix layer by layer; In a process of decoding the target embedding matrix layer by layer, taking a first decoding layer of the decoder as a current layer, and taking the target embedding matrix input to the first decoding layer as an input of the current layer; The normalized product of the input of the current layer and the adjacency matrix of the sample cells is weighted to obtain a decoding result output by the current layer, the decoding result output by the current layer is used as the input of the next layer, and the next layer is used as the new current layer, and the step of weighting the normalized product of the input of the current layer and the adjacency matrix of the sample cells to obtain the decoding result output by the current layer and the subsequent steps are returned, and the decoding result output by the last layer in the decoder is used as the reconstructed cell feature matrix; Determine the reconstructed adjacency matrix according to the reconstructed cell feature matrix and the transpose of the reconstructed cell feature matrix Formation. The method according to any one of claims 1 to 6, characterized in that the iterative optimization of the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference comprises: determining a reconstruction loss value based on at least one of the first difference and the second difference; Determining a target loss value according to at least one of a clustering guidance loss value and a de-redundancy loss value, and the reconstruction loss value; Iteratively optimizing the parameters of the feature encoder to be trained according to the target loss value; The clustering guidance loss value is determined according to the difference between the soft assignment matrix and the target distribution matrix; the soft assignment matrix is determined according to the clustering result obtained by clustering based on the target embedding matrix; the target distribution matrix is obtained by normalizing the soft assignment matrix; The de-redundancy loss value is determined according to a difference between the first embedding matrix and the second embedding matrix. The method according to claim 7, characterized in that the step of determining the clustering guidance loss value comprises: Performing clustering according to the target embedding matrix to obtain reference cluster centers; Determine a soft assignment matrix between the vectors corresponding to each sample cell in the target embedding matrix and the reference cluster center; the soft assignment matrix is used to represent the probability that the vectors are assigned to each reference cluster center; Normalizing the soft allocation matrix to generate a target distribution matrix; The clustering guidance loss value is determined according to a difference between the distributions of the soft assignment matrix and the target distribution matrix. The method according to claim 7 or 8, characterized in that the de-redundancy loss value includes a first de-redundancy loss value; and the step of determining the de-redundancy loss value includes: Determining a node similarity matrix according to the similarity between the first embedding matrix and the second embedding matrix; The first de-redundancy loss value is determined according to a difference between the node similarity matrix and a unit matrix. The method according to claim 9, characterized in that the de-redundancy loss value further includes a second de-redundancy loss value; and the step of determining the de-redundancy loss value further includes: Mapping the first embedding matrix and the second embedding matrix into cluster-level embeddings respectively to obtain a first cluster-level embedding matrix and a second cluster-level embedding matrix; Determining a cluster-level similarity matrix according to the similarity between the first cluster-level embedding matrix and the second cluster-level embedding matrix; The second de-redundancy loss value is determined according to the difference between the cluster-level similarity matrix and the identity matrix. The method according to any one of claims 1 to 10, characterized in that the method further comprises: The graph composed of the cell feature matrix and the adjacency matrix of the cells to be grouped is input into the feature encoding method completed by the training. In the encoder, the extracted embedding matrix is obtained; Clustering is performed on the embedded vectors corresponding to each of the cells to be grouped in the extracted embedded matrix to obtain a cell grouping result of the cells to be grouped. The method according to any one of claims 1 to 10, characterized in that the method further comprises: Clustering is performed on the embedding vectors corresponding to each of the sample cells in the target embedding matrix to obtain a cell clustering result of the sample cells. A cell clustering method, characterized in that the method comprises: Inputting a graph composed of a cell feature matrix and an adjacency matrix of cells to be grouped into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to a first graph and a second graph; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the adjacency matrix of the sample cells; Clustering is performed on the embedded vectors corresponding to each of the cells to be grouped in the extracted embedded matrix to obtain a cell grouping result of the cells to be grouped. The method according to claim 13, characterized in that the step of training the feature encoder comprises: Perturbing the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix; Perturbing the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix; The first graph is encoded by a feature encoder to be trained to obtain a first embedding matrix, and the second graph is encoded to obtain a second embedding matrix; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix; Fusing the first embedding matrix and the second embedding matrix to obtain a target embedding matrix; Decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; According to at least one of the first difference and the second difference, the parameters of the feature encoder to be trained are iteratively optimized until the iteration stop condition is met, thereby obtaining a trained feature encoder; the feature encoder is used to extract features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells. A cell clustering feature extraction device, characterized in that the device comprises: A first perturbation module is used to perturb the cell feature matrix of the sample cells to obtain a first feature matrix and a second feature matrix; A second perturbation module is used to perturb the adjacency matrix of the sample cells to obtain a first adjacency matrix and a second adjacency matrix; An encoding module, configured to encode a first graph through a feature encoder to be trained to obtain a first embedding matrix, and encode a second graph to obtain a second embedding matrix; the first graph includes the first feature matrix and the first adjacency matrix; the second graph includes the second feature matrix and the second adjacency matrix; A fusion module, configured to fuse the first embedding matrix and the second embedding matrix to obtain a target embedding matrix; A decoding module, used for decoding the target embedding matrix to obtain at least one of a reconstructed cell feature matrix and a reconstructed adjacency matrix; A parameter optimization module is used to iteratively optimize the parameters of the feature encoder to be trained according to at least one of the first difference and the second difference until the iteration stop condition is met, thereby obtaining a trained feature encoder; the feature encoder is used to extract features required for cell clustering; the first difference is the difference between the reconstructed cell feature matrix and the cell feature matrix of the sample cells; the second difference is the difference between the reconstructed adjacency matrix and the adjacency matrix of the sample cells. A cell clustering device, characterized in that the device comprises: The embedding matrix determination module is used to input a graph composed of a cell feature matrix and an adjacency matrix of cells to be grouped into a pre-trained feature encoder to obtain an extracted embedding matrix; the pre-trained feature encoder is pre-trained according to a first graph and a second graph; the first graph includes a first feature matrix and a first adjacency matrix; the second graph includes a second feature matrix and a second adjacency matrix; the first feature matrix and the second feature matrix are obtained by perturbing the cell feature matrix of sample cells; the first adjacency matrix and the second adjacency matrix are obtained by perturbing the adjacency matrix of the sample cells; The clustering module is used to cluster the embedding vectors corresponding to each of the cells to be grouped in the extracted embedding matrix to obtain the cell grouping results of the cells to be grouped. A computer device comprises a memory and one or more processors, wherein the memory stores computer-readable instructions, and wherein the one or more processors implement the steps of the method described in any one of claims 1 to 14 when executing the computer-readable instructions. A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions are suitable for being loaded by one or more processors and executing the method according to any one of claims 1-14.