Single cell subcompartment detection method based on graph convolutional network

Abstract

The present invention proposes a single cell sub-compartment detection method based on a graph convolutional network, which is used to solve the technical problems of low detection accuracy and incomplete sub-compartment type detection in the existing technology for sub-compartment structure detection, and the implementation steps are: obtaining a training sample set and a test sample set; constructing a single cell sub-compartment detection network model based on a graph convolutional network; iteratively training the single cell sub-compartment detection network model; obtaining the detection results of the single cell sub-compartment; the graph convolutional network performs multiple graph convolution operations based on the interaction matrix and the feature matrix, and can use the feature information of the sequence interval itself in addition to the existing technology; the feature mapping network automatically learns the sub-compartment label of each sequence interval through iterative optimization, avoiding the problem that the existing technology needs to pre-set the number of sub-compartment types. It can be used for chromatin sub-compartment structure detection, comparison of chromatin structure differences of different cells, and cell feature extraction.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a chromatin structure detection method, and specifically to a single-cell chromatin sub-compartment structure detection method, which can be used to identify chromatin sub-compartment structures at the single-cell level, compare chromatin structure differences of different cells, and extract cell features. Background Art

[0002] The genome of eukaryotic organisms is mainly concentrated in the cell nucleus in the form of chromatin. The expression of genes in chromatin is not only related to the nucleotide sequence of the gene segment, but also closely related to the three-dimensional structure formed in the three-dimensional space. Biological functions such as genome replication, regulatory function, expression, transcription, nucleotide mutation, embryonic development and flow of non-coding RNA are all realized within the spatial structure of the cell nucleus. Nowadays, the rapid development of chromosome conformation capture technology, especially the advent of Hi-C technology, has provided a lot of cell-wide genome interaction data for gene expression research and provided technical support for analyzing the structure of chromatin. The processing flow of Hi-C data can be divided into sequencing data positioning, filtering, binning, standardization and other parts. After processing, the Hi-C data will be constructed into a Hi-C interaction matrix according to the pre-set bin size. At present, there are also many Hi-C data processing integrated software such as HiC-Pro, Juicer, HiC-bench, etc.

[0003] With the development of Hi-C technology, the capture of chromatin interaction information has become more and more accurate. Researchers have found that by analyzing the high-resolution interaction matrix, more detailed sub-compartment structures can be detected. Sub-compartments are the result of further division of the chromatin compartment structure and belong to a link in the chromatin hierarchical structure. The genes within the sub-compartments interact more frequently than genes in different sub-compartments. At first, researchers found that the detected sub-compartments can divide the A compartment of chromatin into two categories, A1 and A2, and the B compartment into two categories, B1 and B2. With the latest research, it is found that sub-compartments can be divided into more. There are more than three different sub-compartments in the A compartment, and the B compartment can be divided into more than four different sub-compartments. Different sub-compartments are enriched with specific biological information. It is conceivable that with the improvement of data set quality and the introduction of more accurate methods, more types of sub-compartments will be discovered. The discovery of sub-compartments has enriched researchers' understanding of the chromatin compartment structure, which is conducive to researchers' further study of the three-dimensional structure of chromatin and the process of DNA transcription and expression, but there are still many gaps in related knowledge that need to be explored. Therefore, the detection of sub-compartments has gradually become one of the important contents of chromatin structure research.

[0004] At present, subcompartment detection methods mainly focus on designing different feature indicators for Hi-C interaction matrix for feature extraction, mapping each bin to low-dimensional space through the designed indicator features, and then clustering using different clustering methods. These methods rely on artificially designed feature indicators, the selection of dimensionality reduction methods, and the performance of different clustering methods. In terms of verification, there are mainly network centrality verification and neural network training verification: network centrality verification is to divide the original interaction matrix according to the subcompartment labels detected by each bin, and divide it into different subcompartment interaction matrices according to the number of subcompartment types. Each subcompartment interaction matrix only leaves the bins of the same type and the number of interactions between them, and calculates different network centrality indicators through the subcompartment interaction matrix to verify the compactness of the detected subcompartment structure; neural network training verification is to train a simple classification neural network based on the obtained subcompartment labels, obtain the prediction accuracy through the same input features and neural network structure, and verify the accuracy of the detected subcompartment structure.

[0005] The latest research constructs the Hi-C interaction matrix as an interaction graph based on the graph idea. It only needs to regard the Hi-C interaction matrix as the adjacency matrix of the interaction graph, so as to construct an interaction graph with bins as nodes and the number of interactions between bins as weights. Using the currently rapidly developing graph embedding method to directly perform feature extraction, feature dimensionality reduction and other operations on the interaction graph can better extract effective features, and then cluster to obtain a more refined and accurate sub-compartment structure. For example, in 2020, Ashoor H et al. published a paper titled "Graph embedding and unsupervised learning predict genomic sub-compartments from Hi-C chromatin interaction data" in Nature Communications, which disclosed a sub-compartment detection method based on Hi-C data. This method first constructs the Hi-C interaction matrix as an interaction graph, learns the low-dimensional vector representation of each bin in the interaction graph through the LINE method, and finally clusters it through the k-means method to obtain six different sub-compartment structures. This method does not require the pre-design of various feature indicators. At the same time, based on the LINE method, it can extract the first-order and second-order similarities of each bin in the interaction graph, make better use of the interaction information between bins, and detect more types and more accurate sub-compartment structures. However, since the LINE method is used to learn the low-dimensional space representation of each bin in an unsupervised manner, this representation only depends on the interaction relationship information and the number of interactions between bins in the interaction graph, which affects the further improvement of its detection accuracy; at the same time, the k-means method is used for downstream clustering, and the number of clusters clustered by the k-means method and the accuracy of the clusters are very dependent on the choice of k value. In this method, since each cluster represents a sub-compartment structure, different k values ​​directly affect the detection accuracy of the final model sub-compartment, thereby limiting the scope of application of this method in sub-compartment structure detection. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and to propose a single-cell sub-compartment detection method based on a graph convolutional network, so as to solve the technical problem of low detection accuracy caused by missing information and incomplete detection of sub-compartment types in the prior art.

[0007] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0008] (1) Obtain training sample set and test sample set:

[0009] Obtain M single-cell Hi-C data containing cell labels in the data set, and preprocess each single-cell Hi-C data to obtain M interaction matrices A = {A1, A2, ..., A m ,…,A M} and feature matrix X = {X1,X2,…,X m ,…,X M}, and then Q preprocessed interaction matrices and their corresponding feature matrices form a training sample set H train , the remaining Y preprocessed interaction matrices and their corresponding feature matrices form the test sample set H test , where M ≥ 500, A m represents the mth preprocessed interaction matrix, X m Indicates A m The corresponding feature matrix, Q>M / 2, M=Q+Y;

[0010] (2) Constructing a single-cell subcompartment detection network model O based on graph convolutional networks:

[0011] (2a) Constructing the structure of the single-cell sub-compartment detection network model O based on graph convolutional network:

[0012] A single cell subcompartment detection network model O is constructed, which includes a sequentially connected feature mapping network, a graph pooling layer, a graph convolution network and a multi-layer perceptron MLP, wherein: the feature mapping network includes a plurality of sequentially connected graph convolution layers, a plurality of fully connected layers and a Softmax normalized activation function; the graph convolution network includes a plurality of graph convolution layers and a plurality of activation functions; the MLP includes an input layer, a hidden layer and an output layer;

[0013] (2b) Define the loss function L of the single-cell subcompartment detection network model O:

[0014]

[0015]

[0016] Among them, C represents the cross entropy loss function, θ is the weight parameter, b represents the randomly selected training sample number, and B is H train The number of randomly selected training samples, S b is the set of allocation matrices s t The b-th allocation matrix in , Denotes the allocation matrix S b The sum of the absolute values ​​of the values ​​in the wth row, a (b) 、c( b ) are the sample label predicted by the model and the actual sample label of the b-th training sample respectively;

[0017] (3) Iterative training of the single-cell sub-compartment detection network model O:

[0018] (3a) The number of initial iterations is t, the maximum number of iterations is T, T ≥ 120, and the single cell subcompartment detection model in the tth iteration is O t , the weights of the feature mapping network, graph convolution network and MLP in the tth iteration are ω 1t ,ω 2t ,ω 3t , and let t = 0, O t =O;

[0019] (3b) From the training sample set H train B training samples randomly selected from the dataset are used as the single cell subcompartment detection network model O t The input is forward propagated to obtain the distribution matrix set s of B training samples t And the cell label prediction value set a t ;

[0020] (3c) Using the loss function L and assigning the matrix set s t , cell label set a t , calculate O t The loss value L t ; Using the back propagation method, through L t Calculate O t The parameter gradient λ t , and then use the gradient descent method to pass λ t Weights ω for feature mapping network, graph convolution network and MLP 1t ,ω 2t ,ω 3t Update to obtain the single cell sub-compartment detection network model O of this iteration t ;

[0021] (3d) Determine whether t=T. If so, obtain the trained single-cell sub-compartment detection model O*. Otherwise, set t=t+1 and execute step (3b).

[0022] (4) Obtaining the detection results of single cell sub-compartments:

[0023] The test sample set H test As the input of the trained single-cell subcompartment detection model O*, forward propagation is performed to obtain the distribution matrix S of each test sample in the Y test sample set y , and the cell label prediction value of the cell corresponding to the test sample; for the distribution matrix S y Probabilistic statistics are performed to obtain the sub-compartment label to which each sequence interval bin of the cells corresponding to the test sample belongs.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] 1. The single-cell sub-compartment detection network model constructed by the present invention includes a graph convolutional network. In the process of training the model and obtaining the detection results of single-cell sub-compartments, multiple graph convolutional layers in the graph convolutional network perform multiple graph convolution operations based on the interaction matrix and the feature matrix, and can simultaneously use the interaction structure information between bins and the feature information of the bins themselves, so as to obtain a more accurate final feature matrix, avoiding the influence of the prior art that only uses the interaction structure information between bins on the detection accuracy, and effectively improving the accuracy of the model sub-compartment structure detection.

[0026] 2. The single-cell sub-compartment detection network model constructed by the present invention includes a feature mapping network. In the process of training the model and obtaining the detection results of single-cell sub-compartments, the feature mapping network performs graph convolution and nonlinear transformation on the sample input to obtain an effective allocation matrix. The allocation matrix is ​​normalized by the Softmax activation function to represent the probability of each bin belonging to different sub-compartments. It can continuously optimize and learn with the iteration of the model, avoiding the problem of the prior art that the number of sub-compartment types needs to be pre-set, so that the model can detect more comprehensive sub-compartment types, further improving the accuracy of the model sub-compartment structure detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of the implementation of the present invention;

[0028] Figure 2 It is a schematic diagram of the structure of the single cell sub-compartment detection network model constructed by the present invention;

[0029] Figure 3 It is a comparison chart of simulation results of the compactness of the sub-chamber structure of the present invention and the prior art. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the present invention does not belong to the subject matter that is not granted with a patent right as stipulated in Article 25 of the Patent Law, and also complies with the provisions of Article 2, Paragraph 2 of the Patent Law:

[0031] Reference Figure 1 , this example includes the following steps:

[0032] Step 1) Obtain single-cell Hi-C dataset:

[0033] Obtain the Hi-C data of 1171 mouse cells at different cell cycles published by Nagano T et al. in 2017 from the 4D Nucleome database;

[0034] Step 2) Construct training sample set and test sample set:

[0035] (2a) The original interaction matrix of the acquired 1171 cell Hi-C data was constructed using the HiC-Pro tool. The sequence interval bin size was selected as 200kb, then W = 986, and the other parameters were the default values ​​of the tool, resulting in 1171 original interaction matrices of size 986 × 986. In order to facilitate the comparison between different values ​​and the calculation of the data enhancement process, each original interaction matrix was z-score standardized by row to obtain the standardized interaction matrix Z of each single cell Hi-C data. m ;

[0036] (2b) For each interaction matrix Z after standardization m Perform convolution operation and calculate the interaction matrix G after convolution operation m Perform a random walk operation. During the convolution operation, the convolution kernel size is f = 5, and the step size is 1; during the random walk, the restart probability is r = 0.5, and the random walk will continue to iterate until The converged matrix R m It is the interaction matrix A after preprocessing m , according to the matrix A m Generate 986 one-hot codes with a dimension of 986 and arrange them in rows to get A m The corresponding feature matrix X m The 936 preprocessed interaction matrices and their corresponding feature matrices form the training sample set H train The remaining 235 preprocessed interaction matrices and their corresponding feature matrices form the test sample set H test .

[0037] Step 3) Construct a single cell subcompartment detection network model O based on graph convolutional network:

[0038] (3a) Construct the structure of the single-cell sub-compartment detection network model O based on graph convolutional network:

[0039] A single cell subcompartment detection network model O is constructed, which includes a sequentially connected feature mapping network, a graph pooling layer, a graph convolutional network, and a multi-layer perceptron MLP. The structure is as follows: Figure 2 As shown, where:

[0040] The feature mapping network contains three sequentially connected graph convolutional layers and three fully connected layers. The number of channels of the last fully connected layer is U, U = λW, λ∈[0.1,0.5]. If λ = 0.1, then U = 99.

[0041] The graph convolutional network contains three graph convolutional layers and three Relu activation functions, with the graph convolutional layer at the beginning and the activation function at the end. The graph convolutional layer and the Relu activation function are cross-connected.

[0042] MLP, the number of hidden layers it contains is u and u≥2, the number of channels in the output layer is the number of true cell labels, and the output layer is connected to the Sigmoid activation function, where the number of true cell labels is 3 and u=2.

[0043] (3b) Define the loss function L of the single-cell subcompartment detection network model O:

[0044]

[0045]

[0046] Where C represents the cross entropy loss function, θ is the weight parameter, θ = 0.05, b represents the randomly selected training sample number, and B is H train The number of randomly selected training samples, B = 30, S b is the set of allocation matrices s t The b-th allocation matrix in , Denotes the allocation matrix S b The sum of the absolute values ​​of the values ​​in the wth row, a (b) 、c (b) are the sample label predicted by the model and the actual sample label of the bth training sample, respectively. Adding a regularization term based on the cross entropy loss function has two main functions: on the one hand, it can make the distribution matrix obtained by the feature mapping network more realistic, and the probability of each bin belonging to different sub-compartments should be a one-hot vector; on the other hand, it can improve the stability of model training and ensure that the random setting of parameter initial values ​​will not have too much impact on the model training process.

[0047] Step 4) Iteratively train the single cell sub-compartment detection network model O:

[0048] (4a) The number of initial iterations is t, the maximum number of iterations is T, T = 200, and the single cell subcompartment detection model for the tth iteration is O t , the weights of the feature mapping network, graph convolution network and MLP in the tth iteration are ω 1t ,ω 2t ,ω 3t , and let t = 0, O t =O;

[0049] (4b) From the training sample set H train 30 randomly selected training samples were used as the single cell subcompartment detection network model O t Forward propagation of input:

[0050] (4b1) The three graph convolutional layers in the feature mapping network are based on the interaction matrix A in each training sample. m For the feature matrix X m Perform graph convolution operation to obtain feature matrix The three fully connected layers in the feature mapping network map the feature matrix Perform nonlinear transformation to obtain the feature matrix Softmax normalized activation function for feature matrix Normalize by row to get the allocation matrix S of size W×U m , U = 99. The specific calculation of the graph convolution operation is as follows:

[0051]

[0052] Where W (0) , W (1) , W (2) They are the weight matrices of the first, second, and third graph convolutional layers, I m is the identity matrix of size W×W, For the matrix The diagonal matrix of ;

[0053] Through cubic graph convolution operations, the feature vectors obtained from similar bins will tend to have more similar representations. After nonlinear mapping and normalization of each feature vector, the normalized feature vectors of the bins under the same sub-compartment structure will tend to be consistent, thereby automatically obtaining the sub-compartment label to which each bin belongs. This process can continuously optimize learning with the iteration of the model, avoiding the problem of the existing technology requiring the pre-setting of the number of sub-compartment types, enabling the model to detect more comprehensive sub-compartment types, further improving the accuracy of the model's sub-compartment structure detection.

[0054] (4b2) The interaction matrix A of the graph pooling layer on each training sample output by the feature mapping network m , feature matrix X m , allocation matrix S m Perform pooling operation to obtain the interval interaction matrix after pooling and the feature matrix

[0055] (4b3) The interaction matrix of multiple graph convolutional layers in the graph convolutional network for each training sample output by the graph pooling layer and the feature matrix Perform graph convolution operation to obtain the final feature matrix The specific calculation of the graph convolution operation is as follows:

[0056]

[0057] Where W (3) , W (4) , W (5) They are the weight matrices of the first, second, and third graph convolutional layers, For the matrix The diagonal matrix of ;

[0058] The three graph convolution layers perform graph convolution operations based on the interaction matrix and the feature matrix. They can simultaneously use the interaction structure information between bins and the feature information of the bins themselves to obtain a more accurate final feature matrix. This avoids the impact of the existing technology that only uses the interaction structure information between bins on detection accuracy, and effectively improves the accuracy of the model sub-compartment structure detection.

[0059] (4b4) The final feature matrix of the input layer in the last MLP to the output of the graph convolutional network Perform mean operation on columns and obtain the one-dimensional vector representation J of the corresponding training sample m ; Multiple hidden layers for J m Perform a nonlinear transformation, where the output vector J of the u-th hidden layer to the u-1 hidden layer is u-1 Perform nonlinear transformation and output vector J u ; Output vector J of the last hidden layer from the output layer u A nonlinear transformation is performed to obtain the cell label prediction value of the corresponding training sample. Finally, the cell label prediction value is mapped to the range of [0,1] through the Sigmoid activation function to obtain the final cell label prediction value, and then the distribution matrix set s of 30 training samples is obtained. t And the cell label prediction value set a t .

[0060] (4c) Using the loss function L and assigning the matrix set s t , cell label set a t , calculate O t The loss value L t ; Using the back propagation method, through L t Calculate O t The parameter gradient λ t , and then use the gradient descent method to pass λ t Weights ω for feature mapping network, graph convolution network and MLP 1t ,ω 2t ,ω 3t Update to obtain the single cell sub-compartment detection network model O of this iterationt , the update formulas are:

[0061]

[0062]

[0063]

[0064] Where η represents the learning rate, η = 0.01, ω 1t ′、ω 2t ′、ω 3t ′ respectively represents ω 1t ,ω 2t ,ω 3t The updated results, Represents partial derivative calculation.

[0065] (4d) Determine whether t=T. If so, obtain the trained single-cell sub-compartment detection model O*. Otherwise, set t=t+1 and execute step (4b).

[0066] Step 5) Obtain the detection results of single cell sub-compartments:

[0067] The test sample set H test As the input of the trained single-cell subcompartment detection model O*, forward propagation is performed to obtain the distribution matrix S of each test sample in the 235 test sample set. y , and the cell label prediction value of the cell corresponding to the test sample; for the distribution matrix S y Statistics are performed by row, and the sub-compartment label corresponding to the column number of the position with the maximum probability in each row is obtained; the corresponding sequence interval bin is obtained through the row number, and the obtained sub-compartment label is used as the sub-compartment label of the sequence interval bin, thereby obtaining the sub-compartment label of each sequence interval bin of the cells corresponding to the test sample.

[0068] The technical effect of the present invention is further illustrated by simulation experiments as follows:

[0069] 1. Simulation conditions and contents:

[0070] The simulation experiments were carried out under the conditions of Intel(R) Core(TM) i7-9400F CPU, main frequency 3.2GHz, memory 32G, and Python3.7.2.

[0071] The compactness of the sub-compartment structure and the accuracy of the sub-compartment labeling of the present invention and the prior art were compared and simulated, and the results are as follows: Figure 3 , as shown in Table 1:

[0072] (1) The preprocessed interaction matrix is ​​divided into different subcompartment interaction matrices. In each subcompartment interaction matrix, only the bins with the same subcompartment labels and the number of interactions between them are retained. The subcompartment interaction matrix is ​​used as the adjacency matrix to construct a subcompartment interaction graph. The closeness centrality and betweenness centrality of each node in these subcompartment interaction graphs are calculated to verify the compactness of the detected subcompartment structure.

[0073] (2) A simple classification neural network is trained based on the sub-compartment labels obtained according to the present invention and the prior art, and the prediction accuracy is obtained by using the same input features and neural network structure to verify the accuracy of the detected sub-compartment labels.

[0074] 2. Analysis of simulation results:

[0075] Reference Figure 3 ,in Figure 3 (a) is a box plot of the distribution of the closeness centrality results of nodes in the sub-compartment network. Figure 3 (b) is a box plot showing the distribution of the betweenness centrality results of nodes in the sub-compartment network.

[0076] Combination Figure 3 It can be seen that the maximum closeness centrality of the present invention is 0.65, the minimum is 0.45, the average is 0.55, the maximum betweenness centrality is 8, the minimum is 2.6, and the average is 5.5; the maximum closeness centrality of the prior art is 0.65, the minimum is 0.38, the average is 0.51, the maximum closeness centrality is 7.4, the minimum is 2.6, and the average is 4.8; the performance of the present invention in the two network centrality indicators is better than that of the prior art, and the overall distribution is denser and the average is higher. It proves that the sub-compartments detected by the present invention are more accurate, the sub-compartment structures found are tighter, and there are more connections between bins.

[0077] The comparison results of the prediction accuracy of the simple neural network trained by the present invention and the prior art are shown in Table 1.

[0078] Table 1

[0079] method Prediction accuracy Prior art 75％ The present invention 87％

[0080] From Table 1, it can be seen that when the input features and the network model structure are the same, using the sub-compartment labels obtained by the present invention as training labels for model training can achieve a higher prediction accuracy than the prior art, proving that the sub-compartment labels detected by the present invention are more accurate.

Claims

1. A single cell subcompartment detection method based on graph convolutional network, characterized in that: The steps include: (1) Obtain training sample set and test sample set: Obtain M single-cell Hi-C data containing cell labels in the data set, and preprocess each single-cell Hi-C data to obtain M interaction matrices A = {A1, A2, ..., A m ,…,A M } and feature matrix X = {X1,X2,…,X m ,…,X M }, and then Q preprocessed interaction matrices and their corresponding feature matrices form a training sample set H train , the remaining Y preprocessed interaction matrices and their corresponding feature matrices form the test sample set H test , where M ≥ 500, A m represents the mth preprocessed interaction matrix, X m Indicates A m The corresponding feature matrix, Q>M / 2, M=Q+Y; (2) Constructing a single-cell subcompartment detection network model O based on graph convolutional networks: (2a) Constructing the structure of the single-cell sub-compartment detection network model O based on graph convolutional network: A single cell subcompartment detection network model O is constructed, which includes a sequentially connected feature mapping network, a graph pooling layer, a graph convolution network and a multi-layer perceptron MLP, wherein: the feature mapping network includes a plurality of sequentially connected graph convolution layers, a plurality of fully connected layers and a Softmax normalized activation function; the graph convolution network includes a plurality of graph convolution layers and a plurality of activation functions; the MLP includes an input layer, a hidden layer and an output layer; (2b) Define the loss function L of the single-cell subcompartment detection network model O: Among them, C represents the cross entropy loss function, θ is the weight parameter, b represents the randomly selected training sample number, and B is H train The number of randomly selected training samples, S b is the set of allocation matrices s t The b-th allocation matrix in , Denotes the allocation matrix S b The sum of the absolute values ​​of the values ​​in the wth row, a (b) 、c (b) are the sample label predicted by the model and the actual sample label of the b-th training sample respectively; (3) Iterative training of the single-cell sub-compartment detection network model O: (3a) The number of initial iterations is t, the maximum number of iterations is T, T ≥ 120, and the single cell subcompartment detection model in the tth iteration is O t , the weights of the feature mapping network, graph convolution network and MLP in the tth iteration are ω 1t ,ω 2t ,ω 3t , and let t = 0, O t =O; (3b) From the training sample set H train B training samples randomly selected from the dataset are used as the single cell subcompartment detection network model O t The input is forward propagated to obtain the distribution matrix set s of B training samples t And the cell label prediction value set a t ; (3c) Using the loss function L and assigning the matrix set s t , cell label set a t , calculate O t The loss value L t ; Using the back propagation method, through L t Calculate O t The parameter gradient λ t , and then use the gradient descent method to pass λ t Weights ω for feature mapping network, graph convolution network and MLP 1t ,ω 2t ,ω 3t Update to obtain the single cell sub-compartment detection network model O of this iteration t ; (3d) Determine whether t=T. If so, obtain the trained single-cell sub-compartment detection model O*. Otherwise, set t=t+1 and execute step (3b). (4) Obtaining the detection results of single cell sub-compartments: The test sample set H test As the input of the trained single-cell subcompartment detection model O*, forward propagation is performed to obtain the distribution matrix S of each test sample in the Y test sample set y , and the cell label prediction value of the cell corresponding to the test sample; for the distribution matrix S y Probabilistic statistics are performed to obtain the sub-compartment label to which each sequence interval bin of the cells corresponding to the test sample belongs.

2. The single cell subcompartment detection method based on graph convolutional network according to claim 1, characterized in that: The preprocessing of each single cell Hi-C data in step (1) is implemented as follows: Obtain M single-cell Hi-C data containing cell labels in the data set, and preprocess each single-cell Hi-C data. (1a) The genome of the cell corresponding to each of the M single-cell Hi-C data obtained from the data set is divided into W sequence interval bins of the same size, and the number of interactions between the sequence interval bins is calculated based on the number of interactions between the fragments in each single-cell Hi-C data. Based on the number of interactions between each sequence interval bin and all other sequence interval bins, an original interaction matrix of size W×W is constructed. The array elements in the original interaction matrix are the size of the interaction number between the sequence interval bins corresponding to the row and column coordinates; then, each original interaction matrix is ​​z-score normalized row by row to obtain the normalized interaction matrix Z of each single-cell Hi-C data. m ; (1b) For each interaction matrix Z after standardization m Perform convolution operation and calculate the interaction matrix G after convolution operation m Perform random walk operation to realize data enhancement of each interaction matrix after standardization operation, and obtain M interaction matrices of size W×W after preprocessing A={A1,A2,…,A m ,…,A M }, then the interaction matrix A m Generate W one-hot codes with dimension W from the row number, and arrange the W one-hot codes by row to get A m The corresponding feature matrix X is of size W×W m , then the characteristic matrix corresponding to A is X={X1,X2,…,X m ,…,X M }; Among them, for the interaction matrix Z m Each element in Perform convolution and the interaction matrix G after the convolution operation m Each element in The formulas for random walk are: 1≤if≤p≤i+f≤W 1≤jf≤q≤j+f≤W in, Yes The interaction matrix G obtained by convolution m The value of the i-th row and j-th column in F pq is the value of the pth row and qth column in the convolution kernel parameter matrix F, f is the convolution window size, is the interaction matrix Z m The value of the pth row and qth column in Yes The normalized interaction matrix D obtained by normalizing by row m The value of row i and column j in is the normalized interaction matrix D m The interaction matrix after k random walks with restarts, r is the restart probability, and I is the identity matrix.

3. The single cell subcompartment detection method based on graph convolutional network according to claim 1, characterized in that: The single cell subcompartment detection network model O based on graph convolutional network described in step (2a), wherein: The feature mapping network contains three graph convolutional layers and three fully connected layers. The number of channels of the last fully connected layer is U, U = λW, λ∈[0.1,0.5]; The graph convolutional network contains three graph convolutional layers and three Relu activation functions, with the graph convolutional layer at the beginning and the activation function at the end. The graph convolutional layer and the Relu activation function are cross-connected. MLP, the number of hidden layers it contains is u and u ≥ 2, the number of channels in the output layer is the number of true cell labels, and the output layer is connected to the Sigmoid activation function.

4. The single cell subcompartment detection method based on graph convolutional network according to claim 1, characterized in that: Step (3b) will be described in the training sample set H train B training samples randomly selected from the dataset are used as the single cell subcompartment detection network model O t The input is forward propagated, and the implementation steps are: (3b1) will be from the training sample set H train B training samples randomly selected from the dataset are used as the single cell subcompartment detection model O t The input of the feature mapping network is multiple graph convolution layers based on the interaction matrix A in each training sample. m For the feature matrix X m Perform graph convolution operation to obtain feature matrix Multiple fully connected layers on the feature matrix Perform nonlinear transformation to obtain the feature matrix Softmax normalized activation function for feature matrix Normalize by row to get the allocation matrix S of size W×U m , where U is the dimension of the feature vector of the normalized sequence interval bin, B ≥ 20; (3b2) The interaction matrix A of the graph pooling layer on each training sample output by the feature mapping network m , feature matrix X m , allocation matrix S m Perform pooling operation to obtain the interval interaction matrix after pooling and the feature matrix in, is the allocation matrix S m The transposed matrix of (3b3) The interaction matrix of multiple graph convolutional layers in the graph convolutional network for each training sample output by the graph pooling layer and the feature matrix Perform graph convolution operation, and multiple activation functions transform the feature matrix obtained by the graph convolution layer to obtain the final feature matrix (3b4) The final feature matrix of the input layer in the MLP to the output of the graph convolutional network Perform mean operation on columns and obtain the one-dimensional vector representation J of the corresponding training sample m , multiple hidden layers perform nonlinear transformation on the one-dimensional vector representation of each training sample output by the graph convolutional network, and the output layer transforms the output vector J of the last hidden layer u A nonlinear transformation is performed to obtain the cell label prediction value of the corresponding training sample. Finally, the cell label prediction value is mapped to the range of [0,1] through the Sigmoid activation function to obtain the final cell label prediction value, and then the allocation matrix set s of B training samples is obtained. t And the cell label prediction value set a t .

5. The single cell subcompartment detection method based on graph convolutional network according to claim 1, characterized in that: The weights ω of the t-th iteration feature map network, graph convolution network, and MLP described in step (3c) 1t ,ω 2t ,ω 3t The update formulas are: Where η represents the learning rate, 0.0001≤η≤0.1, ω 1t ′、ω 2t ′、ω 3t ′ respectively represents ω 1t ,ω 2t ,ω 3t The updated results, Represents partial derivative calculation.

6. The single cell subcompartment detection method based on graph convolutional network according to claim 1, characterized in that: The pair allocation matrix S described in step (4) y Probability statistics are performed, which is specifically implemented as follows: For the allocation matrix S y Statistics are performed by row, and the sub-compartment label corresponding to the column number of the position with the maximum probability in each row is obtained; the corresponding sequence interval bin is obtained through the row number, and the obtained sub-compartment label is used as the sub-compartment label of the sequence interval bin, thereby obtaining the sub-compartment label of each sequence interval bin of the cells corresponding to the test sample.

Images