Unsupervised convolutional neural network model training method and unsupervised convolutional neural network model clustering method and device

Abstract

The invention discloses a training method of an unsupervised convolutional neural network model and a clustering method and device thereof, and the method comprises the steps: carrying out the preprocessing of original space transcriptome data, obtaining a space group slice image for training, and enabling the original space transcriptome data to comprise a plurality of bins; the space group slice images for training are input into an unsupervised convolutional neural network model, and the unsupervised convolutional neural network model learns the bin space position information and the gene expression similarity at the same time; and optimizing the unsupervised convolutional neural network model based on a cell space information continuity loss function and a gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the prior art, the method has the advantages that after preprocessing, the clustering result can be denoised, and the compactness and continuous integrity of the tissue anatomical structure can be kept by keeping the clustering result through learning of cell space position information and gene expression similarity.

Description

Training method of unsupervised convolutional neural network model and clustering method and device thereof Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a training method for an unsupervised convolutional neural network model and a clustering method and device thereof. Background Art

[0002] A related technology provides a new spatial transcriptomics method based on DNB nanospheres, called Stereo-seq, which combines high resolution and sensitivity, providing researchers with more spatial group technology options. Clustering of spatial transcriptomics single-cell data is the first step in analyzing the spatial group data generated by the above technology. Various downstream analyses are based on clustering results, such as cell type identification, identification of spatially differential genes (SVGs), and inference of cell developmental trajectories. Existing tools, such as Seurat, Scanpy, etc., only use scRNA-seq gene expression data that does not contain spatial information for clustering analysis. The above-mentioned single-cell data clustering method fails to take into account the spatial location information when clustering spatial transcriptome data, and the clustering results are usually scattered when mapped to spatial transcription slices, which cannot well portray the anatomical structure of the tissue.

[0003] Summary of the Invention

[0004] The present disclosure provides a training method for an unsupervised convolutional neural network model and a clustering method, device, electronic device, and storage medium thereof.

[0005] According to a first aspect of the present disclosure, a method for training an unsupervised convolutional neural network model is provided, comprising:

[0006] Preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins;

[0007] Inputting the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0008] The unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0009] In some embodiments, the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0010] Inputting the training space group slice image into the unsupervised convolutional neural network model comprises:

[0011] Inputting the training space group slice images into the first batch of normalization layers of the unsupervised convolutional neural network model to perform a first normalization process to obtain the training space group slice images after the first normalization process;

[0012] Inputting the first normalized training space group slice image into the first convolutional layer for feature extraction to obtain the first feature-extracted training space group slice image;

[0013] Inputting the feature-extracted training space group slice images into the second batch normalization layer to obtain second-normalized training space group slice images;

[0014] Inputting the second normalized training space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted training space group slice image;

[0015] The training space group slice images after the second feature extraction are output to the third batch of normalized layers to obtain the target cluster categories output by the unsupervised convolutional neural network model.

[0016] In some embodiments, the method further comprises:

[0017] Obtaining a target cluster category output by the unsupervised convolutional neural network model;

[0018] When it is determined that the number of the target cluster categories is greater than a preset convergence threshold, the outlier bins are merged based on the Euclidean distance until the number of the target cluster categories is less than or equal to the preset convergence threshold.

[0019] In some embodiments, fusing the outlier bins based on the Euclidean distance includes:

[0020] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0021] Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively;

[0022] Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories;

[0023] The outlier bins whose Euclidean distances meet the preset fusion conditions are merged.

[0024] In some embodiments, before fusing the outlier bins whose Euclidean distances meet a preset fusion condition, the method further includes:

[0025] The preset fusion condition is created, and the preset fusion condition includes: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0026] In some embodiments, preprocessing the original spatial transcriptome data to obtain a training spatial group slice image includes:

[0027] Filtering the original spatial transcriptome data according to a preset filtering rule to obtain filtered spatial transcriptome data;

[0028] performing standardization processing on the filtered spatial transcriptome data to obtain standardized spatial transcriptome data;

[0029] Performing dimensionality reduction processing on the standardized spatial transcriptome data to obtain spatial transcriptome data after dimensionality reduction processing;

[0030] The spatial transcriptome data after the dimensionality reduction process is converted into the training spatial group slice image.

[0031] In some embodiments, before inputting the training space group slice image into the unsupervised convolutional neural network model, the method further comprises:

[0032] Pre-training the pre-training space group slice images based on the pre-training process loss function to obtain a pre-trained unsupervised convolutional neural network model;

[0033] Inputting the training space group slice image into the unsupervised convolutional neural network model comprises:

[0034] The training space group slice images are input into the pre-trained unsupervised convolutional neural network model.

[0035] According to a second aspect of the present disclosure, a clustering method based on an unsupervised convolutional neural network model is provided, characterized by comprising:

[0036] Inputting the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0037] Obtain the target cluster category output by the unsupervised convolutional neural network model.

[0038] In some embodiments, after obtaining the target cluster category output by the unsupervised convolutional neural network model, the method further includes:

[0039] When it is determined that the number of the target cluster categories is greater than a preset convergence threshold, the outlier bins are merged based on the Euclidean distance until the number of the target cluster categories is less than or equal to the preset convergence threshold.

[0040] In some embodiments, fusing the outlier bins based on the Euclidean distance includes:

[0041] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0042] Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively;

[0043] Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories;

[0044] Merge the outlier bins whose Euclidean distances meet the preset fusion conditions;

[0045] The preset fusion conditions include: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0046] In some embodiments, the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0047] The step of inputting the to-be-processed spatial transcriptome data into the unsupervised convolutional neural network model comprises:

[0048] Inputting the to-be-processed spatial transcriptome data into the first batch of normalization layers of the unsupervised convolutional neural network model for a first normalization process to obtain a spatial group slice image after the first normalization process;

[0049] Inputting the first normalized space group slice image into the first convolutional layer for feature extraction to obtain a first feature-extracted space group slice image;

[0050] Inputting the feature-extracted space group slice image into the second batch of normalized layers to obtain a second normalized space group slice image;

[0051] Inputting the second normalized space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted space group slice image;

[0052] The spatial group slice images after the second feature extraction are output to the third batch of normalization layers to obtain the target clustering categories output by the unsupervised convolutional neural network model.

[0053] According to a third aspect of the present disclosure, there is provided a training apparatus for an unsupervised convolutional neural network model, comprising:

[0054] A preprocessing unit, configured to preprocess the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes a plurality of bins;

[0055] An input unit, configured to input the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0056] A training unit is used to optimize the unsupervised convolutional neural network model based on a cell spatial information continuity loss function and a gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0057] In some embodiments, the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0058] The input unit is further used for:

[0059] Inputting the training space group slice images into the first batch of normalization layers of the unsupervised convolutional neural network model to perform a first normalization process to obtain the training space group slice images after the first normalization process;

[0060] Inputting the first normalized training space group slice image into the first convolutional layer for feature extraction to obtain the first feature-extracted training space group slice image;

[0061] Inputting the feature-extracted training space group slice images into the second batch normalization layer to obtain second-normalized training space group slice images;

[0062] Inputting the second normalized training space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted training space group slice image;

[0063] The training space group slice images after the second feature extraction are output to the third batch of normalized layers to obtain the target cluster categories output by the unsupervised convolutional neural network model.

[0064] In some embodiments, the apparatus further comprises:

[0065] An acquisition unit, configured to acquire a target cluster category output by the unsupervised convolutional neural network model;

[0066] The fusion unit is configured to, when determining that the number of the target cluster categories is less than a preset convergence threshold, fuse the outlier bins based on the Euclidean distance until the number of the target cluster categories is greater than or equal to the preset convergence threshold.

[0067] In some embodiments, the fusion unit is further configured to:

[0068] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0069] Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively;

[0070] Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories;

[0071] The outlier bins whose Euclidean distances meet the preset fusion conditions are merged.

[0072] In some embodiments, the apparatus further comprises:

[0073] A creation unit is used to create the preset fusion condition before fusing the outlier bins whose Euclidean distances meet the preset fusion condition, wherein the preset fusion condition includes: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0074] In some embodiments, the pre-processing unit is further configured to:

[0075] Filtering the original spatial transcriptome data according to a preset filtering rule to obtain filtered spatial transcriptome data;

[0076] performing standardization processing on the filtered spatial transcriptome data to obtain standardized spatial transcriptome data;

[0077] Performing dimensionality reduction processing on the standardized spatial transcriptome data to obtain spatial transcriptome data after dimensionality reduction processing;

[0078] The spatial transcriptome data after the dimensionality reduction process is converted into the training spatial group slice image.

[0079] In some embodiments, the apparatus further comprises:

[0080] A pre-training unit, configured to pre-train the pre-training space group slice images based on a pre-training process loss function before the input unit inputs the training space group slice images into the unsupervised convolutional neural network model, to obtain a pre-trained unsupervised convolutional neural network model;

[0081] The input unit is also used to input the training space group slice image into the pre-trained unsupervised convolutional neural network model.

[0082] According to a fourth aspect of the present disclosure, a clustering device based on an unsupervised convolutional neural network model is provided, comprising:

[0083] An input unit, configured to input the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial location information and gene expression similarity;

[0084] An acquisition unit is used to obtain the target cluster category output by the unsupervised convolutional neural network model.

[0085] In some embodiments, the apparatus further comprises:

[0086] A fusion unit is used to, after obtaining the target cluster category output by the unsupervised convolutional neural network model, fuse the outlier bins based on the Euclidean distance when it is determined that the number of the target cluster categories is less than a preset convergence threshold, until the number of the target cluster categories is greater than or equal to the preset convergence threshold.

[0087] In some embodiments, the fusion unit is further configured to:

[0088] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0089] Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively;

[0090] Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories;

[0091] Merge the outlier bins whose Euclidean distances meet the preset fusion conditions;

[0092] The preset fusion conditions include: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0093] In some embodiments, before inputting the spatial transcriptome data to be processed into the unsupervised convolutional neural network model, the apparatus further comprises: a training unit;

[0094] The training unit is further configured to:

[0095] Preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins;

[0096] Inputting the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0097] The unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0098] In some embodiments, the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0099] The input unit is further configured to:

[0100] Inputting the to-be-processed spatial transcriptome data into the first batch of normalization layers of the unsupervised convolutional neural network model for a first normalization process to obtain a spatial group slice image after the first normalization process;

[0101] Inputting the first normalized space group slice image into the first convolutional layer for feature extraction to obtain a first feature-extracted space group slice image;

[0102] Inputting the feature-extracted space group slice image into the second batch of normalized layers to obtain a second normalized space group slice image;

[0103] Inputting the second normalized space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted space group slice image;

[0104] The spatial group slice images after the second feature extraction are output to the third batch of normalization layers to obtain the target clustering categories output by the unsupervised convolutional neural network model.

[0105] According to a fifth aspect of the present disclosure, there is provided an electronic device, including:

[0106] at least one processor; and

[0107] a memory communicatively connected to the at least one processor; wherein,

[0108] The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect or the second aspect.

[0109] According to a sixth aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the method described in the first aspect or the second aspect.

[0110] According to a seventh aspect of the present disclosure, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the method as described in the first or second aspect above.

[0111] The present disclosure provides a training method for an unsupervised convolutional neural network model and a clustering method and device thereof, which pre-processes the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins; the training spatial group slice image is input into the unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; the unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the related art, the embodiment of the present application can denoise the clustering results after pre-processing, and by learning the cell spatial position information and gene expression similarity, the clustering result can maintain the compactness and continuous integrity of the tissue anatomical structure.

[0112] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] The accompanying drawings are provided to facilitate a better understanding of the present invention and do not constitute a limitation of the present disclosure.

[0114] FIG1 is a flow chart of a method for training an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0115] FIG2 is a framework diagram of an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0116] FIG3 is a flow chart of another clustering method based on an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0117] FIG4 is a schematic diagram of the structure of a training device for an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0118] FIG5 is a schematic diagram of the structure of another unsupervised convolutional neural network model training device provided by an embodiment of the present disclosure;

[0119] FIG6 is a schematic structural diagram of another clustering device based on an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0120] FIG7 is a schematic structural diagram of another clustering device based on an unsupervised convolutional neural network model provided by an embodiment of the present disclosure;

[0121] FIG8 is a schematic block diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0122] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0123] The following describes the training method, device, electronic device and storage medium of the unsupervised convolutional neural network model of the embodiments of the present disclosure with reference to the accompanying drawings.

[0124] FIG1 is a flow chart of a method for training an unsupervised convolutional neural network model provided in an embodiment of the present disclosure.

[0125] As shown in Figure 1, the method includes the following steps:

[0126] Step 101 : pre-processing original spatial transcriptome data to obtain training spatial group slice images, wherein the original spatial transcriptome data includes multiple bins.

[0127] The preprocessing described in the embodiments of the present application includes but is not limited to quality control, standardization, and dimensionality reduction, so as to achieve the purpose of being able to input into an unsupervised convolutional neural network model.

[0128] It should be noted that bins and spot cells have corresponding sizes or corresponding ranges in the idle data. For example, a spot cell can correspond to bin 100 and can be 55 microns in size. Typically, a spot cell contains 5-10 cells. The bin size corresponding to a cell is approximately bin 14 (mouse brain) - bin 20 (liver cancer cell). Therefore, in practical applications, bins can be replaced with spot cells, and the specific embodiments of this application are not limited thereto.

[0129] Step 102: input the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity.

[0130] In an embodiment of the present invention, the preprocessed spatial transcriptome data is treated as a high-dimensional multi-channel sparse training space group slice image for processing, each bin / spot point in the preprocessed spatial transcriptome data is treated as each pixel point in the training space group slice image, and the multidimensional gene expression feature is regarded as a channel of the training space group slice image. In an embodiment of the present invention, the three-dimensional training space group slice image matrix corresponding to the model is represented as (m, n, d). Among them, m is the maximum value of the horizontal coordinate of the training space group slice image, n is the maximum value of the vertical coordinate of the training space group slice image, and d is the number of principal components after dimensionality reduction. Before the constructed three-dimensional training space group slice image matrix is ​​input into the unsupervised convolutional neural network model, (m, n, d) will be z-score standardized in the d dimension, that is, for each dimensional data X i :

[0131] In the above formula, X i is the original data of the matrix (m,n,d) on channel i, is the mean of the above data in channel i, is the variance of the above data in channel i, X′ i is the matrix after normalization.

[0132] Step 103: Optimize the unsupervised convolutional neural network model based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0133] During the training of the unsupervised convolutional neural network model, the spatial information continuity loss function L spatial and gene expression similarity loss function L similarityAt the same time, we learn the spatial location information and gene expression similarity of bin / spot points, so that the spots with continuous spatial information and similar gene expression are clustered into the same category. For the matrix after argmax in the d dimension, the above loss function is defined as follows: L=αL similarity +βL spatial

[0134] in, is the model output data, C is the number of current clusters, α and β are the gene expression similarity loss function L similarity And the spatial continuity loss function L spatial The weight of .

[0135] It should be noted that the output dimension of the unsupervised convolutional neural network model is (N, d), where N is the m×n size after reshaping the training space group slice image.

[0136] The present disclosure provides an unsupervised convolutional neural network model training method and a clustering method thereof, wherein the original spatial transcriptome data is preprocessed to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes a plurality of bins; the training spatial group slice image is input into the unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; the unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the related art, the embodiment of the present application can denoise the clustering results after preprocessing, and by learning the cell spatial position information and gene expression similarity, the clustering result can maintain the compactness and continuous integrity of the tissue anatomical structure.

[0137] To facilitate understanding of the architecture of the unsupervised convolutional neural network model, as shown in Figure 2, the unsupervised convolutional neural network model includes: a first batch normalization layer (Batch Normalization), a first convolutional layer (Conv2D), a second batch normalization layer (Batch Normalization), an activation layer (ReLU), a second convolutional layer (Conv2D), and a third batch normalization layer (Batch Normalization). As an optional method, the unsupervised convolutional neural network model also includes an activation layer (ReLU), a third convolutional layer (Conv1D), and a fourth batch normalization layer (Batch Normalization). The activation layer (ReLU), the third convolutional layer (Conv1D), and the fourth batch normalization layer (Batch Normalization) are added to improve the performance of the unsupervised convolutional neural network model.

[0138] Example 1: During training, set the parameters of the deep neural network model (unsupervised convolutional neural network model) as follows:

[0139] Number of 2D convolution kernels: f=15,

[0140] 2D convolution kernel size: k=3,

[0141] 2D convolution operation step size: s = 1,

[0142] 2D convolution kernel l2 regularization factor: r l2 =0.00001,

[0143] Activation function LeakyReLU slope: slope = 0.2,

[0144] The size of the output feature map of the 2D 1×1 convolution operation remains unchanged from the size of the input feature map.

[0145] Gene expression similarity loss function L similarity Weight: α = 0.4,

[0146] Spatial continuity loss function L spatial Weight: β = 0.7,

[0147] Number of pretraining steps: pretrain_epochs = 400.

[0148] Number of target labels min_label∈{7,7,7,7,5,5,5,5,7,7,7,7}

[0149] For 10XDLPFC, the sample numbers are sample_id∈{151507,151508,151509,151510,151669,151670, 151671,151672,151673,151674,151675,151676}.

[0150] The model pre-training steps were set to 400, and the total number of epochs was set to 2100. The Adam algorithm (with parameters β1 = 0.9, β2 = 0.999) was used for gradient descent optimization, and the initial learning rate was set to 0.002.

[0151] Based on the setting of the above parameters, the training space group slice image is input into the unsupervised convolutional neural network model, including: inputting the training space group slice image into the first batch of standardization layers of the unsupervised convolutional neural network model for first standardization processing to obtain the training space group slice image after the first standardization processing; inputting the training space group slice image after the first standardization processing into the first convolution layer for feature extraction to obtain the training space group slice image after the first feature extraction; inputting the training space group slice image after the feature extraction into the second batch of standardization layers to obtain the training space group slice image after the second standardization processing; inputting the training space group slice image after the second standardization processing into the activation layer for processing as the input of the second convolution layer, and obtaining the training space group slice image after the second feature extraction; outputting the training space group slice image after the second feature extraction into the third batch of standardization layers to obtain the target cluster category output by the unsupervised convolutional neural network model.

[0152] With respect to the added activation layer (ReLU), the third convolution layer (Conv1D), and the fourth batch normalization layer (Batch Normalization) in the unsupervised convolutional neural network model, the inputting of the training space group slice image into the unsupervised convolutional neural network model includes: inputting the training space group slice image into the first batch normalization layer of the unsupervised convolutional neural network model for first normalization processing to obtain the training space group slice image after the first normalization processing; inputting the training space group slice image after the first normalization processing into the first convolution layer for feature extraction to obtain the training space group slice image after the first feature extraction; inputting the training space group slice image after the feature extraction into the second batch normalization layer to obtain the training space group slice image after the second normalization processing; inputting the training space group slice image after the second normalization processing into the second batch normalization layer to obtain the training space group slice image after the second normalization processing; The training space group slice image is input into the activation layer for processing and then used as the input of the second convolutional layer, and the training space group slice image after the second feature extraction is obtained; the training space group slice image after the second feature extraction is output into the third batch normalization layer, and the training space group slice image after the third normalization is obtained; the training space group slice image after the third normalization is input into the activation layer for processing and then used as the input of the third convolutional layer, and the training space group slice image after the third feature extraction is obtained; the training space group slice image after the third feature extraction is output into the fourth batch normalization layer, and the target cluster category output by the unsupervised convolutional neural network model is obtained.

[0153] As an implementable method of the embodiment of the present application, please continue to refer to Figure 2. After the training spatial group slice image is input into the unsupervised convolutional neural network model, it is first input into a first batch normalization layer (Batch Normalization), which is connected to it by two repeated first convolution layers. Specifically, the first convolution layer is composed of a 3×3 two-dimensional convolution Conv2D, a second batch normalization layer (Batch Normalization), and finally connected to a ReLU activation layer. Finally, it will be connected to a 1×1 second convolution layer, and finally through a third batch normalization layer (Batch Normalization) to obtain the output of the unsupervised convolutional neural network model.

[0154] In actual applications, after training is completed, the cluster categories obtained may not reach the target cluster categories. In order to further improve the clustering accuracy, the embodiment of the present application fuses some scattered and independent bin / spot points. This fusion process is also called a post-processing strategy. The fusion strategy is as follows: obtain the target cluster category output by the unsupervised convolutional neural network model; when it is determined that the number of target cluster categories is greater than the preset convergence threshold, fuse the outlier bins based on the Euclidean distance until the number of target cluster categories is less than or equal to the preset convergence threshold.

[0155] When fusing outlier bins based on Euclidean distance, it can be implemented in the following ways, but not limited to: performing dimensionality reduction processing on the bins of each cluster category to the target dimension; calculating the first mean of the bins after dimensionality reduction processing on the target dimension; calculating the mean and variance of the Euclidean distance between different bins under different cluster categories; and fusing the outlier bins whose Euclidean distance meets the preset fusion conditions.

[0156] Calculate the matrix (N) after dimensionality reduction in each category i ,d) The first mean P in the target dimension d i , where N i is the number of bin / spot points belonging to each category. Where i∈{0,1,…,c}, c is the label number of the target cluster, the above set is the final target cluster category, calculate the above category P i , the second mean P j The Euclidean distance D between two i,j .

[0157] Before fusing the outlier bins whose Euclidean distances meet the preset fusion conditions, the preset fusion conditions are created. The preset fusion conditions need to meet the following two conditions (condition a and condition b):

[0158] a. The class to which the outlier belongs and other classes are spatially separable. i,j The set of outliers is denoted as D, and the set of outliers is defined as O: O={o|o∈||D-μ D ||>σ D}

[0159] In the above formula, μ D is the mean of the Euclidean distance set D, σ D The variance of the Euclidean distance set D.

[0160] b. Point D corresponding to the minimum Euclidean distance in the above set D spat_i,spot_j ∈O.

[0161] The above steps are repeated until the output of the unsupervised convolutional neural network model reaches the preset convergence threshold.

[0162] Optionally, the post-processed data is fed into an unsupervised convolutional neural network model for further processing until the number of labels specified is reached, at which point training is terminated. For example, the maximum number of post-processing steps is set to 100. For example, after 2000 pre-training steps, post-processing is performed, and then the unsupervised convolutional neural network model is fed into the model for 100 training steps.

[0163] In order to adapt to the input of the unsupervised convolutional neural network model, the original spatial transcriptome data needs to be preprocessed before inputting the unsupervised convolutional neural network model. In practical applications, the preprocessing can be limited to the following methods: filtering the original spatial transcriptome data according to preset filtering rules to obtain filtered spatial transcriptome data; standardizing the filtered spatial transcriptome data to obtain standardized spatial transcriptome data; after quality control, the original spatial transcriptome data described in the embodiment of the present application will be subjected to gene filtering for the spatial transcriptome data of 10X Visium (the gene must be expressed in at least 5 spot cells / bin); spot / bin and gene filtering will be performed simultaneously for Stereo-seq data to filter out the number of genes corresponding to 5% of the spot / bin points to screen each spot point.

[0164] The filtered spatial transcriptome data is standardized to obtain standardized spatial transcriptome data; the standardization processing described in the embodiment of the present application includes but is not limited to library size standardization processing, log conversion processing and other spatial transcriptome standardized data processing processes. The specific implementation method can refer to the detailed description of the relevant technology, and the embodiment of the present application will not be repeated here.

[0165] Example 1 uses 12 10X DLPFC cerebral cortex samples, each with approximately 4,000 bin / spots, each 55 μm in size. Data cleaning and preprocessing were performed on these samples according to the steps in the previous example, with the gene filtering condition requiring the gene to be expressed in at least five bin / spots (min_cells = 5).

[0166] The standardized spatial transcriptome data is subjected to dimensionality reduction processing to obtain spatial transcriptome data after dimensionality reduction processing, and the spatial transcriptome data after dimensionality reduction processing is converted into the training space group slice image. In the embodiment of the present application, the training spatial transcriptome data is treated as a high-dimensional multi-channel sparse training space group slice image, each bin / spot point in the training spatial transcriptome data is treated as each pixel point in the training space group slice image, and the multidimensional gene expression features are regarded as channels of the training space group slice image.

[0167] As a feasible way of the embodiment of the present application, when the dimensionality reduction processing is performed on the standardized spatial transcriptome data, the number of PCA dimensions can be set to 15 dimensions. It should be noted that the specific numerical values ​​described in the above embodiment are only examples in Example 1, and are not specific limitations of the embodiment of the present application. The numerical values ​​mentioned in the above embodiment can be configured according to different clustering requirements.

[0168] The above embodiment describes the training process of the unsupervised convolutional neural network model in detail, but in the training of unsupervised

[0169] Before the convolutional neural network model is constructed, it is also necessary to initialize the parameters and / or weights of the unsupervised convolutional neural network model. Therefore, before inputting the training space group slice image into the unsupervised convolutional neural network model, the method further includes: pre-training the pre-training space group slice image based on the pre-training process loss function to obtain the pre-trained unsupervised convolutional neural network model; and inputting the training space group slice image into the pre-trained unsupervised convolutional neural network model.

[0170] The unsupervised convolutional neural network model needs to learn the inherent patterns of the pre-trained spatial group slice image itself so that it can converge to a better direction later. The loss function of the pre-training process is the mean square error:

[0171] Where N is the size of m×n after reshaping, y is the pre-training spatial group slice image, and the dimension is (N, d); Output data for the model, the dimension is also (N, d).

[0172] According to Example 1, the number of pre-training steps of the unsupervised convolutional neural network model is set to 400, and the total number of iterations epochs = 2100. The gradient descent optimization algorithm uses the Adam algorithm (parameters β1 = 0.9, β2 = 0.999), and the initial learning rate is set to 0.002. If the unsupervised convolutional neural network model has not reached the set number of cluster categories after 2000 steps of training, the post-processing strategy is started in the training phase until the min_label set by the user is reached. It should be noted that the embodiment of the present application does not specifically limit the parameter settings of the pre-training phase.

[0173] The above example 1 is explained using 12 10X DLPFC cerebral cortex samples as an example, and the following example 2 is explained using Stereo-seq mouse brain samples as an example.

[0174] Example 2:

[0175] Example 2 used Stereo-seq mouse brain samples to test the robustness of the present invention on Stereo-seq spatial transcriptome data at four different resolutions. Bin 20 samples had 526,716 bins / spots; bin 50 samples had 84,724 bins / spots; bin 100 samples had 21,368 bins / spots; and bin 200 samples had 5,420 bins / spots. Stereo-seq sequencing captured 27,279 genes.

[0176] The above samples were cleaned and preprocessed, where min_cells = 5, and min_genes is the number of genes corresponding to 5% of bin / spot points filtered out.

[0177] The input data of the unsupervised convolutional neural network model is constructed, where for bin100, bin200 data d=50; for bin20 and bin50 data d=100.

[0178] Model training and clustering are performed for each input sample.

[0179] Set the unsupervised convolutional neural network model parameters as follows:

[0180] Number of 2D convolution kernels: f=15,

[0181] 2D convolution kernel size: k=3,

[0182] 2D convolution operation step size: s = 1,

[0183] 2D convolution kernel l2 regularization factor: r l2 =0.00001,

[0184] Activation function LeakyReLU slope: slope = 0.2,

[0185] The size of the output feature map of the 2D 1×1 convolution operation remains unchanged from the size of the input feature map.

[0186] Gene expression similarity loss function L similarity Weight: α = 0.4,

[0187] Spatial continuity loss function L spatial Weight: β = 0.7,

[0188] Pretraining steps: pretrain_epochs = 400. min_label = 20.

[0189] The number of pre-training steps for the unsupervised convolutional neural network model was set to 400, and the total number of epochs was set to 2100. The Adam algorithm (with parameters β1 = 0.9, β2 = 0.999) was used for gradient descent optimization, and the initial learning rate was set to 0.002.

[0190] If the unsupervised convolutional neural network model has not reached the set number of clusters after 2000 steps of training, the post-processing strategy will be started in the subsequent training until the min_label set by the user is reached, and the clustering result file after the unsupervised convolutional neural network model training is saved.

[0191] Table 1 shows that the LISI clustering performance indicators of the present invention at different resolutions of Chinese rat brain are significantly better than other similar algorithms; in addition, in terms of time and space complexity, the embodiment of the present application also has great advantages over other similar algorithms (see Table 2).

[0192] Table 1 LISI clustering performance index results at different resolutions on mouse brain data

[0193] Table 2 Time and space complexity results of different resolutions on the mouse brain data

[0194] In summary, the embodiment of the present application simultaneously learns the correlation of bin / spot spatial location information and the intrinsic pattern of gene expression through an unsupervised convolutional neural network model. By designing an end-to-end architecture, data processing to the unsupervised convolutional neural network model result output encapsulation forms a closed loop. There is no need to understand the complex implementation process. You only need to specify a few parameters such as the input spatial group data path and the final number of clusters to run this algorithm.

[0195] The present invention achieves high clustering accuracy. The unsupervised convolutional neural network model outperforms existing algorithms in terms of clustering evaluation indicators (ARI) (0.532) and NMI (0.644) for 12 samples from an existing public dataset. ARI and NMI comparison results for each algorithm on 12 samples from the 10x DLPFC dataset are shown in Tables 1 and 2. Furthermore, the unsupervised convolutional neural network model described in the present application embodiment has a simple architecture, low computational memory overhead (2.6 GB), and short training time (1 minute). The runtime and memory requirements of each algorithm on 12 samples from the 10x DLPFC dataset are shown in Table 1.

[0196] Compared to other spatial group clustering algorithms, the embodiments of this application can better denoise clustering results and maintain the compactness and continuous integrity of tissue anatomical structures, especially for tissues such as the mouse brain that require fine and complete structures. In addition, the method proposed in the embodiments of this application is highly scalable and robust. In addition to being suitable for 10X Visium data, it is also applicable to Stereo-seq data at different resolutions (bin20, bin50, bin100, bin200) and data from multiple spatiotemporal group platforms, such as MERFISH, seqFISH, Slide-seqV2, etc.

[0197] The above embodiment describes a training method based on an unsupervised convolutional neural network model. The following embodiment describes a clustering method based on an unsupervised convolutional neural network model, as shown in FIG3 , including:

[0198] Step 201: inputting the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial location information and gene expression similarity;

[0199] For the description of the unsupervised convolutional neural network model, please refer to the detailed description of the above embodiment, so it will not be repeated here.

[0200] Step 202: Obtain the target cluster category output by the unsupervised convolutional neural network model.

[0201] The present disclosure provides a training method for an unsupervised convolutional neural network model and a clustering method and device thereof, which pre-processes the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins; the training spatial group slice image is input into the unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; the unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the related art, the embodiment of the present application can denoise the clustering results after pre-processing, and by learning the cell spatial position information and gene expression similarity, the clustering result can maintain the compactness and continuous integrity of the tissue anatomical structure.

[0202] In some embodiments, after obtaining the target cluster categories output by the unsupervised convolutional neural network model, the method further includes: when it is determined that the number of the target cluster categories is less than a preset convergence threshold, fusing the outlier bins based on Euclidean distance until the number of the target cluster categories is greater than or equal to the preset convergence threshold.

[0203] When fusing outlier bins based on Euclidean distance, the following methods may be used but are not limited to: performing dimensionality reduction processing on the bins of each cluster category to the target dimension; calculating the first mean of the bins after dimensionality reduction processing on the target dimension; calculating the second mean between different bins under different cluster categories; calculating the Euclidean distance between the first mean and the second mean; fusing the outlier bins whose Euclidean distances meet the preset fusion conditions; wherein the preset fusion conditions include: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0204] For the fusion processing process, please refer to the detailed description of the above embodiment, which will not be described in detail in the embodiment of the present application.

[0205] The unsupervised convolutional neural network model described in the embodiment of the present application includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0206] When executing step 201 to input the spatial transcriptome data to be processed into the unsupervised convolutional neural network model, the following methods can be adopted but are not limited to: inputting the spatial transcriptome data to be processed into the first batch of normalization layers of the unsupervised convolutional neural network model for first normalization processing to obtain a space group slice image after first normalization processing; inputting the space group slice image after the first normalization processing into the first convolution layer for feature extraction to obtain a space group slice image after first feature extraction; inputting the space group slice image after feature extraction into the second batch of normalization layers to obtain a space group slice image after second normalization processing; inputting the space group slice image after the second normalization processing into the activation layer for processing as the input of the second convolution layer, and obtaining a space group slice image after second feature extraction; outputting the space group slice image after the second feature extraction into the third batch of normalization layers to obtain the target clustering category output by the unsupervised convolutional neural network model.

[0207] It should be noted that the above embodiment is described using a three-layer convolutional neural network as an example. In actual application, the three-layer convolutional neural network can be replaced with other common network architectures such as ResNet, U-net and other arbitrary network architectures, and the specific embodiments of this application are not limited.

[0208] It should be noted that further improvements, modifications, and alternatives to the embodiments of this application include but are not limited to:

[0209] Modification of the network structure based on the core architecture of the deep neural network model proposed in the embodiments of the present application, or substitution of PCA for gene expression feature extraction using an autoencoder network or a variational autoencoder network based on the design concept of the embodiments of the present application;

[0210] Integrate H&E image expression information to correct gene expression information or directly input the image as another dimension feature into the model for training.

[0211] The design of the model pseudo-label can replace argmax with other pseudo-label generation strategies such as soft-Kmeans, Kmeans, UMAP, etc.

[0212] When multiple samples are used as inputs of the unsupervised convolutional neural network model of the embodiment of the present application, the unsupervised convolutional neural network model has a batch effect correction effect for adjacent slices of different batches, which will further expand the scope of application of the embodiment of the present application.

[0213] Corresponding to the above-mentioned unsupervised convolutional neural network model training method, the present invention also provides an unsupervised convolutional neural network model training device. Since the device embodiment of the present invention corresponds to the above-mentioned method embodiment, any details not disclosed in the device embodiment can be referred to the above-mentioned method embodiment and will not be further described in this invention.

[0214] FIG4 is a schematic diagram of the structure of a training device for an unsupervised convolutional neural network model provided by an embodiment of the present disclosure, as shown in FIG4 , including:

[0215] A preprocessing unit 31 is used to preprocess the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins;

[0216] An input unit 32 is used to input the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0217] The training unit 33 is used to optimize the unsupervised convolutional neural network model based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0218] The present disclosure provides a training device for an unsupervised convolutional neural network model, which pre-processes the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins; the training spatial group slice image is input into the unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; the unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the related art, the embodiment of the present application can denoise the clustering results after pre-processing, and by learning the cell spatial position information and gene expression similarity, the clustering result can maintain the compactness and continuous integrity of the tissue anatomical structure.

[0219] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0220] The input unit 32 is further configured to:

[0221] Inputting the training space group slice images into the first batch of normalization layers of the unsupervised convolutional neural network model to perform a first normalization process to obtain the training space group slice images after the first normalization process;

[0222] Inputting the first normalized training space group slice image into the first convolutional layer for feature extraction to obtain the first feature-extracted training space group slice image;

[0223] Inputting the feature-extracted training space group slice images into the second batch normalization layer to obtain second-normalized training space group slice images;

[0224] Inputting the second normalized training space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted training space group slice image;

[0225] The training space group slice images after the second feature extraction are output to the third batch of normalized layers to obtain the target cluster categories output by the unsupervised convolutional neural network model.

[0226] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the apparatus further includes:

[0227] An acquisition unit 34 is used to acquire the target cluster category output by the unsupervised convolutional neural network model;

[0228] The fusion unit 35 is configured to, when it is determined that the number of the target cluster categories is greater than a preset convergence threshold, fuse the outlier bins based on the Euclidean distance until the number of the target cluster categories is less than or equal to the preset convergence threshold.

[0229] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the fusion unit 35 is further configured to:

[0230] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0231] Calculate the first mean of the bin of each cluster category after dimensionality reduction on the target dimension;

[0232] Calculate the Euclidean distance between different cluster categories based on the first mean of different cluster categories;

[0233] The outlier bins whose Euclidean distances meet the preset fusion conditions are merged.

[0234] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the apparatus further includes:

[0235] A creation unit 36 ​​is used to create the preset fusion condition before fusing the outlier bins whose Euclidean distances meet the preset fusion condition. The preset fusion condition includes: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0236] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the preprocessing unit 31 is further configured to:

[0237] Filtering the original spatial transcriptome data according to a preset filtering rule to obtain filtered spatial transcriptome data;

[0238] performing standardization processing on the filtered spatial transcriptome data to obtain standardized spatial transcriptome data;

[0239] Performing dimensionality reduction processing on the standardized spatial transcriptome data to obtain spatial transcriptome data after dimensionality reduction processing;

[0240] The spatial transcriptome data after the dimensionality reduction process is converted into the training spatial group slice image.

[0241] Furthermore, in a possible implementation of this embodiment, as shown in FIG5 , the apparatus further includes:

[0242] A pre-training unit 37 is configured to pre-train the pre-training space group slice images based on a pre-training process loss function before the input unit inputs the training space group slice images into the unsupervised convolutional neural network model to obtain a pre-trained unsupervised convolutional neural network model;

[0243] The input unit 32 is further configured to input the training space group slice images into the pre-trained unsupervised convolutional neural network model.

[0244] Corresponding to the aforementioned clustering method based on an unsupervised convolutional neural network model, the present invention also provides a clustering device based on an unsupervised convolutional neural network model. Since the device embodiments of the present invention correspond to the aforementioned method embodiments, details not disclosed in the device embodiments can be referred to the aforementioned method embodiments and will not be further described in this invention.

[0245] The embodiment of the present application further provides a clustering device based on an unsupervised convolutional neural network model, as shown in FIG6 , comprising:

[0246] An input unit 41 is used to input the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial location information and gene expression similarity;

[0247] The acquisition unit 42 is used to obtain the target cluster category output by the unsupervised convolutional neural network model.

[0248] Furthermore, in a possible implementation of this embodiment, as shown in FIG7 , the apparatus further includes:

[0249] The fusion unit 43 is used to, after obtaining the target cluster category output by the unsupervised convolutional neural network model, fuse the outlier bins based on the Euclidean distance when it is determined that the number of the target cluster categories is greater than a preset convergence threshold, until the number of the target cluster categories is less than or equal to the preset convergence threshold.

[0250] The present disclosure provides a clustering device based on an unsupervised convolutional neural network model, which pre-processes the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins; the training spatial group slice image is input into the unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; the unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model. Compared with the related art, the embodiment of the present application can denoise the clustering results after pre-processing, and by learning the cell spatial position information and gene expression similarity, the clustering result can maintain the compactness and continuous integrity of the tissue anatomical structure.

[0251] Furthermore, in a possible implementation of this embodiment, as shown in FIG7 , the fusion unit 43 is further configured to:

[0252] Perform dimensionality reduction on the bins of each cluster category to the target dimension;

[0253] Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively;

[0254] Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories;

[0255] Merge the outlier bins whose Euclidean distances meet the preset fusion conditions;

[0256] The preset fusion conditions include: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

[0257] In some embodiments, before inputting the spatial transcriptome data to be processed into the unsupervised convolutional neural network model, the apparatus further comprises: a training unit 44;

[0258] The training unit 44 is further configured to:

[0259] Preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes multiple bins;

[0260] Inputting the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity;

[0261] The unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

[0262] In some embodiments, the unsupervised convolutional neural network model includes: a first batch of normalization layers, a first convolutional layer, a second batch of normalization layers, an activation layer, a second convolutional layer, and a third batch of normalization layers;

[0263] The input unit 41 is further configured to:

[0264] Inputting the to-be-processed spatial transcriptome data into the first batch of normalization layers of the unsupervised convolutional neural network model for a first normalization process to obtain a spatial group slice image after the first normalization process;

[0265] Inputting the first normalized space group slice image into the first convolutional layer for feature extraction to obtain a first feature-extracted space group slice image;

[0266] Inputting the feature-extracted space group slice image into the second batch of normalized layers to obtain a second normalized space group slice image;

[0267] Inputting the second normalized space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, to obtain the second feature-extracted space group slice image;

[0268] The spatial group slice images after the second feature extraction are output to the third batch of normalization layers to obtain the target clustering categories output by the unsupervised convolutional neural network model.

[0269] It should be noted that the above explanation of the method embodiment is also applicable to the device of this embodiment, and the principles are the same, which is not limited in this embodiment.

[0270] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0271] FIG8 shows a schematic block diagram of an example electronic device 500 that can be used to implement an embodiment of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0272] As shown in FIG8 , the device 500 includes a computing unit 501 that can perform various appropriate actions and processes based on a computer program stored in a ROM (Read-Only Memory) 502 or a computer program loaded from a storage unit 508 into a RAM (Random Access Memory) 503. Various programs and data required for the operation of the device 500 can also be stored in the RAM 503. The computing unit 501, the ROM 502, and the RAM 503 are connected to each other via a bus 504. An I / O (Input / Output) interface 505 is also connected to the bus 504.

[0273] Various components in device 500 are connected to I / O interface 505, including: an input unit 506, such as a keyboard, mouse, etc.; an output unit 507, such as various types of displays, speakers, etc.; a storage unit 508, such as a magnetic disk, optical disk, etc.; and a communication unit 509, such as a network card, modem, wireless communication transceiver, etc. The communication unit 509 allows device 500 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0274] The computing unit 501 can be a variety of general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Unit), various specialized AI (Artificial Intelligence) computing chips, various computing units that run machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as the training method of an unsupervised convolutional neural network model. For example, in some embodiments, the training method of an unsupervised convolutional neural network model can be implemented as a computer software program that is tangibly embodied in a machine-readable medium, such as the storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 500 via the ROM 502 and / or the communication unit 509. When the computer program is loaded into the RAM 503 and executed by the computing unit 501, one or more steps of the method described above can be performed. Alternatively, in other embodiments, the computing unit 501 may be configured to execute the aforementioned training method of the unsupervised convolutional neural network model in any other appropriate manner (e.g., by means of firmware).

[0275] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System on Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0276] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0277] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0278] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0279] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0280] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact via a communication network. This client-server relationship is established by computer programs running on the respective computers, establishing a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host, a host product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosts and VPS services ("Virtual Private Servers" or simply "VPS"). The server may also be a server in a distributed system or a server integrated with blockchain.

[0281] It's important to note that artificial intelligence (AI) is the study of how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). This encompasses both hardware and software technologies. AI hardware technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily encompass computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graphs.

[0282] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0283] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A training method for an unsupervised convolutional neural network model, characterized in that: include: Preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes a plurality of bins; Inputting the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; The unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

2. The method according to claim 1, characterized in that The unsupervised convolutional neural network model includes: a first batch of standardized layers, a first convolutional layer, a second batch of standardized layers, an activation layer, a second convolutional layer, and a third batch of standardized layers; The step of inputting the training space group slice image into the unsupervised convolutional neural network model comprises: Inputting the training space group slice images into the first batch of standardized layers of the unsupervised convolutional neural network model for first standardized processing to obtain the training space group slice images after the first standardized processing; Inputting the first normalized training space group slice image into the first convolutional layer for feature extraction to obtain the first feature extracted training space group slice image; Inputting the feature-extracted training space group slice images into the second batch of standardized layers to obtain the second standardized training space group slice images; Inputting the second normalized training space group slice image into the activation layer for processing, and then using it as the input of the second convolutional layer, and obtaining the second feature extracted training space group slice image; The training space group slice images after the second feature extraction are output to the third batch of standardized layers to obtain the target clustering category output by the unsupervised convolutional neural network model.

3. The method according to claim 2, characterized in that The method further comprises: Obtaining a target clustering category output by the unsupervised convolutional neural network model; When it is determined that the number of the target cluster categories is greater than a preset convergence threshold, the outlier bins are merged based on the Euclidean distance until the number of the target cluster categories is less than or equal to the preset convergence threshold.

4. The method according to claim 3, characterized in that The fusion of outlier bins based on Euclidean distance includes: Perform dimensionality reduction on the bins of each clustering category to the target dimension; Calculate the first mean value of each bin of each cluster category on the target dimension after dimensionality reduction processing. value; Calculate the Euclidean distance between different cluster categories according to the first mean of different cluster categories; The outlier bins whose Euclidean distances meet the preset fusion conditions are merged.

5. The method according to claim 4, characterized in that Before fusing the outlier bins whose Euclidean distances meet the preset fusion condition, the method further includes: The preset fusion condition is created, and the preset fusion condition includes: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

6. The method according to any one of claims 1 to 5, characterized in that The preprocessing of the original spatial transcriptome data to obtain the spatial group slice images for training includes: Filtering the original spatial transcriptome data according to a preset filtering rule to obtain filtered spatial transcriptome data; Performing standardization processing on the filtered spatial transcriptome data to obtain standardized spatial transcriptome data; Performing dimensionality reduction processing on the standardized spatial transcriptome data to obtain spatial transcriptome data after dimensionality reduction processing; The spatial transcriptome data after the dimension reduction process is converted into the training spatial group slice image.

7. The method according to claim 1, characterized in that Before inputting the training space group slice image into the unsupervised convolutional neural network model, the method further includes: Pre-training the pre-training space group slice images based on the pre-training process loss function to obtain a pre-trained unsupervised convolutional neural network model; The step of inputting the training space group slice image into the unsupervised convolutional neural network model comprises: The training space group slice images are input into the pre-trained unsupervised convolutional neural network model.

8. A clustering method based on an unsupervised convolutional neural network model, characterized in that: include: Inputting the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; Obtain a target clustering category output by the unsupervised convolutional neural network model.

9. The method according to claim 8, characterized in that After obtaining the target clustering category output by the unsupervised convolutional neural network model, the method further includes: When it is determined that the number of the target cluster categories is greater than a preset convergence threshold, the outlier bins are merged based on the Euclidean distance until the number of the target cluster categories is less than or equal to the preset convergence threshold.

10. The method according to claim 9, characterized in that The fusion of outlier bins based on Euclidean distance includes: Perform dimensionality reduction on the bins of each clustering category to the target dimension; Calculate the first mean of the bins after dimensionality reduction processing on the target dimension respectively; Calculate the mean and variance of the Euclidean distance between different bins under different clustering categories; Merge the outlier bins whose Euclidean distances meet the preset fusion conditions; The preset fusion conditions include: the category to which the outlier bin belongs and other categories are spatially separable, and the bin corresponding to the minimum Euclidean distance in the set of Euclidean distances belongs to the set of outlier bins.

11. The method according to claim 8, characterized in that Before inputting the spatial transcriptome data to be processed into the unsupervised convolutional neural network model, the method further comprises: Preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes a plurality of bins; Inputting the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; The unsupervised convolutional neural network model is optimized based on the cell spatial information continuity loss function and the gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

12. The method according to any one of claims 8 to 11, characterized in that The unsupervised convolutional neural network model includes: a first batch of standardized layers, a first convolutional layer, a second batch of standardized layers, an activation layer, a second convolutional layer, and a third batch of standardized layers; The step of inputting the to-be-processed spatial transcriptome data into the unsupervised convolutional neural network model comprises: Inputting the to-be-processed spatial transcriptome data into the first batch of standardized layers of the unsupervised convolutional neural network model for a first standardized process to obtain a spatial group slice image after the first standardized process; Inputting the space group slice image after the first normalization process into the first convolutional layer for feature extraction to obtain the space group slice image after the first feature extraction; Inputting the feature-extracted space group slice image into the second batch of standardized layers to obtain a second standardized space group slice image; Inputting the space group slice image after the second normalization processing into the activation layer for processing, and then using it as the input of the second convolutional layer, and obtaining the space group slice image after the second feature extraction; The spatial group slice images after the second feature extraction are output to the third batch normalization layer to obtain the target clustering category output by the unsupervised convolutional neural network model.

13. A training device for an unsupervised convolutional neural network model, characterized in that: include: A preprocessing unit, used for preprocessing the original spatial transcriptome data to obtain a training spatial group slice image, wherein the original spatial transcriptome data includes a plurality of bins; An input unit is used to input the training space group slice image into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; A training unit is used to optimize the unsupervised convolutional neural network model based on a cell spatial information continuity loss function and a gene expression similarity loss function to obtain a trained unsupervised convolutional neural network model.

14. A clustering device based on an unsupervised convolutional neural network model, characterized in that: include: An input unit, used for inputting the spatial transcriptome data to be processed into an unsupervised convolutional neural network model, wherein the unsupervised convolutional neural network model simultaneously learns the bin spatial position information and gene expression similarity; An acquisition unit is used to acquire the target clustering category output by the unsupervised convolutional neural network model.

15. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 7, or the method of any one of claims 8 to 12.

16. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 7, or the method according to any one of claims 8 to 12.

17. A computer program product, characterized in that The method comprises a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 7 or the method according to any one of claims 8 to 12.

Images