Single-cell RNA sequencing data dimensionality reduction method, device and readable storage medium

Abstract

The present invention provides a single-cell RNA sequencing data dimensionality reduction method, device and readable storage medium. The method comprises: performing feature selection on single-cell sequencing data of N single cells to obtain a feature matrix X; constructing a connection matrix A according to the Mahalanobis distance between two single cells; constructing a graph autoencoder model based on A; inputting X into the graph autoencoder model to obtain a reconstructed feature matrix Y, a reconstructed connection matrix and extracting a hidden variable Z of the lowest dimensional layer in the graph autoencoder model; obtaining a total loss function according to X, A, Y, and Z; using a gradient descent method to obtain a minimized total loss function and a trained graph autoencoder model; extracting a hidden variable Z' of the lowest dimensional layer in the trained graph autoencoder model as a dimensionality reduction result. Through the present invention, it is achieved that the structural information between cells is maintained when carrying out a dimensionality reduction task on large, complex and high-dimensional data.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a single-cell RNA sequencing data dimensionality reduction method, device and readable storage medium. Background Art

[0002] Complex biological tissues and organisms are composed of cell groups with different shapes and functions. Single-cell RNA sequencing (scRNA-seq) technology sequences the RNA of each cell to obtain the expression levels of all genes in the cell. Unlike traditional batch sequencing, it has the ability to analyze the biological state of a single cell and is widely used in many biological fields such as tumor biology, embryonic development, and organ formation. In scRNA-seq data, the expression level of each gene in a cell can be regarded as a feature of the cell. The key step in mining biological information from it is to reduce the high-dimensional and complex scRNA-seq data to two dimensions to achieve visualization and intuitively observe the distribution of cell populations.

[0003] A scRNA-seq data set generally contains tens of thousands or even hundreds of thousands of cells, each of which contains tens of thousands of gene expression features. Current dimensionality reduction methods cannot fully mine key information from such large and high-dimensional data. Traditional dimensionality reduction methods have their own limitations. For example, linear methods such as PCA cannot accurately and effectively capture the features of original high-dimensional data in two-dimensional or three-dimensional space. In recent years, with the development of deep learning, some clustering methods based on deep learning have been proposed. However, in the face of large-scale, high-dimensional and noisy scRNA-seq data, the dimensionality reduction effect is still poor. At the same time, the current dimensionality reduction methods lack the mining of structural information between cells. For scRNA-seq data, the structural information between cells cannot be ignored. The structural information contained in cells with similar biological states contains a lot of biological information.

[0004] Therefore, a method is needed to mine the structural information between cells in high-dimensional, large-scale, and noisy scRNA-seq data to perform the dimensionality reduction task. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a single-cell RNA sequencing data dimensionality reduction method, device and readable storage medium.

[0006] In a first aspect, the present invention provides a single-cell RNA sequencing data dimensionality reduction method, the single-cell RNA sequencing data dimensionality reduction method comprising:

[0007] Perform feature selection on the single-cell sequencing data of N single cells to obtain a feature matrix X, where N is a positive integer;

[0008] Calculate the Mahalanobis distance between two single cells;

[0009] According to the Mahalanobis distance between two single cells, the connection matrix A is constructed;

[0010] Construct a graph autoencoder model based on the connection matrix A;

[0011] Input the feature matrix X into the graph autoencoder model;

[0012] Get the reconstructed feature matrix Y and reconstructed connection matrix output by the graph autoencoder model And extract the hidden variable Z of the lowest dimensional layer in the graph autoencoder model;

[0013] According to the feature matrix X, the connection matrix A, the feature matrix Y is reconstructed, and the connection matrix is ​​reconstructed And the hidden variable Z to get the total loss function;

[0014] The gradient descent method is used to obtain the minimized total loss function and the trained graph autoencoder model;

[0015] The hidden variable Z' of the lowest-dimensional layer in the trained graph autoencoder model is extracted as the dimensionality reduction result.

[0016] Optionally, the step of performing feature selection on the single-cell sequencing data of N single cells to obtain a feature matrix X includes:

[0017] M genes that meet the preset conditions are selected as features from the single-cell sequencing data of N single cells to obtain the feature matrix X, where N is a positive integer, and the preset conditions are:

[0018] d g >exp[-(m g -b)]+0.02

[0019] in, m g ={log2X ig |X ig >0}, I is unit 1, X ig is the expression level of the g-th gene in cell i, and b is the hyperparameter corresponding to M obtained by the dichotomy method.

[0020] Optionally, the step of constructing a connection matrix A according to the Mahalanobis distance between two single cells comprises:

[0021] Take each single cell as a node. For cell i, select the first t cells with the smallest Mahalanobis distance to cell i as the neighboring points of cell i to construct a graph and obtain a connection graph.

[0022] Based on the connection graph, a connection matrix A is constructed, where if cell j is a neighbor of cell i, then the element A in the i-th row and j-th column of the connection matrix A is ij is 1 if the value is true, otherwise it is 0.

[0023] Optionally, the propagation formula of the graph autoencoder model is:

[0024]

[0025] in, I N is the N-th order identity matrix, W (h -1) is the trainable parameter of the (h-1)th layer, and ReLU(·) is the nonlinear activation function.

[0026] Optionally, the feature matrix X, the connection matrix A, the reconstructed feature matrix Y, the reconstructed connection matrix And the steps to obtain the total loss function of the hidden variable Z include:

[0027] According to the feature matrix X and the reconstructed feature matrix Y, the content reconstruction loss function L is obtained. 1 ;

[0028] According to the connection matrix A and the reconstruction connection matrix Get the graph structure reconstruction loss function L 2 ;

[0029] According to the connection matrix A and the latent variable Z, the low-dimensional graph structure reconstruction loss function L is obtained. 3 ;

[0030] According to the total loss function calculation formula, the total loss function Loss is obtained. The total loss function calculation formula is:

[0031] Loss = L 1 +L 2 +λ·L 3

[0032] Among them, λ is a preset value.

[0033] Optionally, content reconstruction loss function L 1 for:

[0034]

[0035] Among them, X ij is the element in the i-th row and j-th column of the feature matrix X, Y ij is the element in the i-th row and j-th column of the reconstructed feature matrix Y.

[0036] Optionally, graph structure reconstruction loss function L 2 for:

[0037]

[0038] in, sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij , They are A, The element at row i and column j in .

[0039] Optional, low-dimensional graph structure reconstruction loss function L 3 for:

[0040]

[0041] Where H = sigmoid(Z T Z), sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij is the element in row i and column j of A, H ij is the element in H at the i-th row and j-th column.

[0042] In a second aspect, the present invention also provides a single-cell RNA sequencing data dimensionality reduction device, which includes a processor, a memory, and a single-cell RNA sequencing data dimensionality reduction program stored in the memory and executable by the processor, wherein when the single-cell RNA sequencing data dimensionality reduction program is executed by the processor, the steps of the single-cell RNA sequencing data dimensionality reduction method as described above are implemented.

[0043] In a third aspect, the present invention also provides a readable storage medium, on which a single-cell RNA sequencing data dimensionality reduction program is stored, wherein when the single-cell RNA sequencing data dimensionality reduction program is executed by a processor, the steps of the single-cell RNA sequencing data dimensionality reduction method as described above are implemented.

[0044] In the present invention, feature selection is performed on single-cell sequencing data of N single cells to obtain a feature matrix X, where N is a positive integer; the Mahalanobis distance between two single cells is calculated; a connection matrix A is constructed based on the Mahalanobis distance between two single cells; a graph autoencoder model is constructed based on the connection matrix A; the feature matrix X is input into the graph autoencoder model; and a reconstructed feature matrix Y and a reconstructed connection matrix output by the graph autoencoder model are obtained. And extract the hidden variable Z of the lowest dimensional layer in the graph autoencoder model; according to the feature matrix X, the connection matrix A, reconstruct the feature matrix Y, reconstruct the connection matrix and hidden variable Z to obtain the total loss function; the gradient descent method is used to obtain the minimized total loss function and the trained graph autoencoder model; the hidden variable Z' of the lowest dimensional layer in the trained graph autoencoder model is extracted as the dimensionality reduction result. Through the present invention, the similarity between cells is measured by Mahalanobis distance, the influence of the dimensions of different gene expression levels is removed, and the mutual interference between genes is eliminated; the graph autoencoder model utilizes the information of neighboring single cells, so that the structural information between cells is maintained when the dimensionality reduction task is carried out on large, complex and high-dimensional data. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the hardware structure of a single-cell RNA sequencing data dimensionality reduction device involved in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the process of an embodiment of a single-cell RNA sequencing data dimensionality reduction method of the present invention.

[0047] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0048] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0049] In a first aspect, an embodiment of the present invention provides a single-cell RNA sequencing data dimensionality reduction device, which can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0050] Reference Figure 1 , Figure 1The hardware structure diagram of the single-cell RNA sequencing data dimensionality reduction device involved in the embodiment of the present invention is shown in FIG. In the embodiment of the present invention, the single-cell RNA sequencing data dimensionality reduction device may include a processor 1001 (e.g., a central processing unit, CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components; the user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard); the network interface 1004 may optionally include a standard wired interface, a wireless interface (such as wireless fidelity WIreless-FIdelity, WI-FI interface); the memory 1005 may be a high-speed random access memory (random access memory, RAM), or a stable memory (non-volatile memory), such as a disk storage, and the memory 1005 may optionally be a storage device independent of the aforementioned processor 1001. Those skilled in the art will understand that Figure 1 The hardware structure shown in the figure does not constitute a limitation of the present invention, and may include more or less components than those shown in the figure, or combine certain components, or arrange the components differently.

[0051] Continue to refer to Figure 1 , Figure 1 The memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a single-cell RNA sequencing data dimensionality reduction program. The processor 1001 may call the single-cell RNA sequencing data dimensionality reduction program stored in the memory 1005 and execute the single-cell RNA sequencing data dimensionality reduction method provided in an embodiment of the present invention.

[0052] In a second aspect, an embodiment of the present invention provides a method for dimensionality reduction of single-cell RNA sequencing data.

[0053] In one embodiment, referring to Figure 2 , Figure 2 FIG. 1 is a flow chart of an embodiment of a method for reducing the dimension of single-cell RNA sequencing data according to the present invention. Figure 2 As shown in the figure, the single-cell RNA sequencing data dimensionality reduction methods include:

[0054] Step S10, performing feature selection on the single-cell sequencing data of N single cells to obtain a feature matrix X, where N is a positive integer;

[0055] In this embodiment, single-cell sequencing data of N single cells are selected from the single-cell RNA sequencing data set, and then feature selection is performed on the single-cell sequencing data of the N single cells to obtain a feature matrix X, where N is a positive integer.

[0056] Furthermore, in one embodiment, step S10 includes:

[0057] M genes that meet the preset conditions are selected as features from the single-cell sequencing data of N single cells to obtain the feature matrix X, where N is a positive integer, and the preset conditions are:

[0058] d g >exp[-(m g -b)]+0.02

[0059] in, m g ={log2X ig |X ig >0}, I is unit 1, X ig is the expression level of the g-th gene in cell i, and b is the hyperparameter corresponding to M obtained by the dichotomy method.

[0060] In this embodiment, N single-cell sequencing data are selected from the single-cell RNA sequencing data set. M = 2000 genes are predetermined as cell features, and the hyperparameter b corresponding to M = 2000 is found by binary search, so that d g satisfy:

[0061] d g >exp[-(m g -b)]+0.02

[0062] in, m g ={log2X ig |X ig >0}, I is unit 1, X ig is the expression level of the g-th gene in cell i, and b is the hyperparameter corresponding to M obtained by the dichotomy method.

[0063] Step S20, calculating the Mahalanobis distance between two single cells;

[0064] In this embodiment, in order to eliminate the differences in expression values ​​between genes, the Mahalanobis distance between two single cells is calculated using the Mahalanobis distance calculation formula.

[0065] Step S30, constructing a connection matrix A according to the Mahalanobis distance between two single cells;

[0066] In this embodiment, the similarity between two single cells can be determined according to the Mahalanobis distance between two single cells, so that the connection matrix A is constructed based on the similarity between two single cells.

[0067] Furthermore, in one embodiment, step S30 includes:

[0068] Take each single cell as a node. For cell i, select the first t cells with the smallest Mahalanobis distance to cell i as the neighboring points of cell i to construct a connection graph. Based on the connection graph, construct a connection matrix A, where if cell j is a neighboring point of cell i, then the element A in the i-th row and j-th column of the connection matrix A is ij is 1 if the value is true, otherwise it is 0.

[0069] In this embodiment, t is selected as the maximum value between (0.01×N) and 20.

[0070] Step S40, constructing a graph autoencoder model based on the connection matrix A;

[0071] In this embodiment, a graph autoencoder model is constructed based on the connection matrix A. Specifically, the propagation formula of the graph autoencoder model is:

[0072]

[0073] in, I N is the N-th order identity matrix, W (h-1) is the trainable parameter of the (h-1)th layer, and ReLU(·) is the nonlinear activation function. As an illustration, the dimensions of each layer of the graph autoencoder are 2000-512-256-128-68-10-2-10-68-128-512-2000. The hidden variable Z of the layer with a dimension of 2 will be extracted as the result of dimensionality reduction after the graph autoencoder model is trained, and the last layer is used as the reconstruction matrix Y.

[0074] Step S50, inputting the feature matrix X into the graph autoencoder model;

[0075] Step S60, obtaining the reconstructed feature matrix Y and the reconstructed connection matrix output by the graph autoencoder model And extract the hidden variable Z of the lowest dimensional layer in the graph autoencoder model;

[0076] Step S70, based on the feature matrix X, the connection matrix A, the feature matrix Y and the connection matrix And the hidden variable Z to get the total loss function;

[0077] In this embodiment, the content reconstruction loss function L is obtained according to the feature matrix X and the reconstructed feature matrix Y. 1 ;

[0078] According to the connection matrix A and the reconstruction connection matrix Get the graph structure reconstruction loss function L 2 ;

[0079] According to the connection matrix A and the latent variable Z, the low-dimensional graph structure reconstruction loss function L is obtained.3 ;

[0080] According to the total loss function calculation formula, the total loss function Loss is obtained. The total loss function calculation formula is:

[0081] Loss = L 1 +L 2 +λ·L 3

[0082] Among them, λ is a preset value.

[0083] Furthermore, in one embodiment, the content reconstruction loss function L 1 for:

[0084]

[0085] Among them, X ij is the element in the i-th row and j-th column of the feature matrix X, Y ij is the element in the i-th row and j-th column of the reconstructed feature matrix Y.

[0086] Furthermore, in one embodiment, the graph structure reconstruction loss function L 2 for:

[0087]

[0088] in, sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij , They are A, The element at row i and column j in .

[0089] Furthermore, in one embodiment, the low-dimensional graph structure reconstruction loss function L 3 for:

[0090]

[0091] Where H = sigmoid(Z T Z), sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij is the element in row i and column j of A, H ij is the element in H at the i-th row and j-th column.

[0092] Step S80, using the gradient descent method to obtain the minimized total loss function and the trained graph autoencoder model;

[0093] In this embodiment, the gradient descent method is used to adjust the parameters of the graph autoencoder model, so as to obtain the minimized total loss function and the graph autoencoder model parameters corresponding to the minimized total loss function.

[0094] Step S90, extracting the hidden variable Z' of the lowest dimensional layer in the trained graph autoencoder model as the dimensionality reduction result.

[0095] In this embodiment, after the trained graph autoencoder model is obtained according to step S80, the latent variable Z' of the lowest dimensional layer in the trained graph autoencoder model can be extracted as the dimensionality reduction result.

[0096] In this embodiment, feature selection is performed on the single-cell sequencing data of N single cells to obtain a feature matrix X, where N is a positive integer; the Mahalanobis distance between two single cells is calculated; a connection matrix A is constructed based on the Mahalanobis distance between two single cells; a graph autoencoder model is constructed based on the connection matrix A; the feature matrix X is input into the graph autoencoder model; and a reconstructed feature matrix Y and a reconstructed connection matrix output by the graph autoencoder model are obtained. And extract the hidden variable Z of the lowest dimensional layer in the graph autoencoder model; according to the feature matrix X, the connection matrix A, reconstruct the feature matrix Y, reconstruct the connection matrix and hidden variable Z to obtain the total loss function; the gradient descent method is used to obtain the minimized total loss function and the trained graph autoencoder model; the hidden variable Z' of the lowest dimensional layer in the trained graph autoencoder model is extracted as the dimensionality reduction result. Through this embodiment, the similarity between cells is measured by the Mahalanobis distance, the influence of the dimensions of different gene expression levels is removed, and the mutual interference between genes is eliminated; the graph autoencoder model utilizes the information of neighboring single cells, so that the structural information between cells is maintained when the dimensionality reduction task is carried out on large, complex and high-dimensional data.

[0097] In a third aspect, an embodiment of the present invention further provides a readable storage medium.

[0098] The readable storage medium of the present invention stores a single-cell RNA sequencing data dimensionality reduction program, wherein when the single-cell RNA sequencing data dimensionality reduction program is executed by a processor, the steps of the single-cell RNA sequencing data dimensionality reduction method as described above are implemented.

[0099] Among them, the method implemented when the single-cell RNA sequencing data dimensionality reduction program is executed can refer to the various embodiments of the single-cell RNA sequencing data dimensionality reduction method of the present invention, and will not be repeated here.

[0100] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0101] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0102] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device to execute the methods described in each embodiment of the present invention.

[0103] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for dimensionality reduction of single-cell RNA sequencing data, characterized in that: The single-cell RNA sequencing data dimensionality reduction method comprises: Perform feature selection on the single-cell sequencing data of N single cells to obtain a feature matrix X, where N is a positive integer; Calculate the Mahalanobis distance between two single cells; According to the Mahalanobis distance between two single cells, the connection matrix A is constructed; Construct a graph autoencoder model based on the connection matrix A; Input the feature matrix X into the graph autoencoder model; Obtain the reconstructed feature matrix Y and the reconstructed connection matrix A output by the graph autoencoder model, and extract the hidden variable Z of the lowest-dimensional layer in the graph autoencoder model; The total loss function is obtained based on the feature matrix X, the connection matrix A, the reconstructed feature matrix Y, the reconstructed connection matrix A and the latent variable Z; The gradient descent method is used to obtain the minimized total loss function and the trained graph autoencoder model; Extract the hidden variable Z' of the lowest-dimensional layer in the trained graph autoencoder model as the dimensionality reduction result; The step of constructing a connection matrix A according to the Mahalanobis distance between two single cells comprises: Take each single cell as a node. For cell i, select the first t cells with the smallest Mahalanobis distance to cell i as the neighboring points of cell i to construct a graph and obtain a connection graph. Based on the connection graph, a connection matrix A is constructed, where if cell j is a neighbor of cell i, then the element A in the i-th row and j-th column of the connection matrix A is ij is 1, otherwise it is 0; The step of obtaining the total loss function according to the feature matrix X, the connection matrix A, the reconstructed feature matrix Y, the reconstructed connection matrix A and the latent variable Z comprises: The content reconstruction loss function L1 is obtained according to the feature matrix X and the reconstructed feature matrix Y; According to the connection matrix A and the reconstructed connection matrix A, the graph structure reconstruction loss function L2 is obtained; According to the connection matrix A and the latent variable Z, the low-dimensional graph structure reconstruction loss function L3 is obtained; According to the total loss function calculation formula, the total loss function Loss is obtained. The total loss function calculation formula is: Loss = L1 + L2 + λ L3 Among them, λ is the preset value; The content reconstruction loss function L1 is: Among them, X ij is the element in the i-th row and j-th column of the feature matrix X, Y ij is the element in the i-th row and j-th column of the reconstructed feature matrix Y; The graph structure reconstruction loss function L2 is: in, sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij , They are A, The element in row i and column j of ; The low-dimensional graph structure reconstruction loss function L3 is: Where H = sigmoid(Z T Z), sigmoid(·) is a nonlinear activation function, T is the transpose operation of the matrix, A ij is the element in row i and column j of A, H ij is the element in H at the i-th row and j-th column.

2. The single-cell RNA sequencing data dimensionality reduction method according to claim 1, characterized in that: The step of performing feature selection on the single-cell sequencing data of N single cells to obtain a feature matrix X comprises: M genes that meet the preset conditions are selected as features from the single-cell sequencing data of N single cells to obtain the feature matrix X, where N is a positive integer, and the preset conditions are: d g >exp[-(m g -b)]+0.02 in, m g ={log2X ig |X ig >0}, I is unit 1, X ig is the expression level of the g-th gene in cell i, and b is the hyperparameter corresponding to M obtained by the dichotomy method.

3. The single-cell RNA sequencing data dimensionality reduction method according to claim 1, characterized in that: The propagation formula of the graph autoencoder model is: in, I N is the N-th order identity matrix, W (h-1) is the trainable parameter of the h-1th layer, and ReLU(·) is the nonlinear activation function.

4. A single-cell RNA sequencing data dimensionality reduction device, characterized in that: The single-cell RNA sequencing data dimensionality reduction device includes a processor, a memory, and a single-cell RNA sequencing data dimensionality reduction program stored in the memory and executable by the processor, wherein when the single-cell RNA sequencing data dimensionality reduction program is executed by the processor, the steps of the single-cell RNA sequencing data dimensionality reduction method as described in any one of claims 1 to 3 are implemented.

5. A readable storage medium, characterized in that: The readable storage medium stores a single-cell RNA sequencing data dimensionality reduction program, wherein when the single-cell RNA sequencing data dimensionality reduction program is executed by a processor, the steps of the single-cell RNA sequencing data dimensionality reduction method according to any one of claims 1 to 3 are implemented.

Images