A multi-task cell analysis method and system based on residual graph neural network
By constructing a multi-task cell analysis method based on residual graph neural networks, and combining a denoising autoencoder and graph neural networks, the problem of neglecting multi-task correlation in single-cell analysis is solved, and more efficient data dimensionality reduction, clustering and imputation effects are achieved.
Patent Information
- Application Number
- CN202311263903.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-27
- Publication Date
- 2025-12-19
- Estimated Expiration
- 2043-09-27
AI Technical Summary
Existing single-cell analysis methods neglect the correlation between multiple tasks, resulting in poor feature learning and performance optimization when processing single-cell transcriptome data.
A multi-task cell analysis method based on residual graph neural networks is adopted. By constructing a denoising autoencoder and a residual graph neural network model, and combining a dual self-supervised loss function, data dimensionality reduction, clustering and imputation are performed. The autoencoder and graph neural network are used to capture the relationships and features between cells.
It improves the performance of single-cell analysis tasks, particularly in terms of clustering accuracy, imputation accuracy, and computational efficiency, significantly enhancing the effectiveness of data analysis.
Smart Images
Figure CN117476107B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of bioinformatics single-cell data analysis, in particular to a multi-task cell analysis method and system based on residual graph neural network. BACKGROUND
[0002] Single-cell sequencing is a new sequencing technology that can measure gene expression within individual cells. Through the analysis of single-cell sequencing data, researchers can discover differences between different cell types, subtypes and states, reveal gene regulation networks, cell communication mechanisms, and explore disease mechanisms. Therefore, accurate analysis of single-cell data helps scientists to more comprehensively and accurately understand the properties, functions and diversity of cells. This technology has had a profound impact on the fields of biomedical research, medical diagnosis and drug development, and has helped to promote scientific progress and improve health care. Specifically, single-cell analysis mainly includes cell type clustering, gene imputation and dimensionality reduction of single-cell transcriptome data. However, due to the "dropout" phenomenon and noise problems commonly existing in single-cell transcriptome data, designing an accurate single-cell analysis method is a challenging task.
[0003] The key to improving single-cell analysis methods is to extract powerful and discriminative cell feature representations from single-cell transcriptome data. For example, Isabella et al. extended the previous method, i.e. hierarchical clustering significance, and proposed a model-based hypothesis testing method, which incorporates significance analysis into the clustering algorithm and allows statistical evaluation of clustering as different cell populations; George et al. proposed a method based on low-rank matrix approximation, which can impute missing values in single-cell transcriptome data while keeping biologically unexpressed genes (true biological zeros) at zero expression level; Wang et al. proposed a deep variational autoencoder (VASC) for single-cell transcriptome data, which is a deep multi-layer generative model for unsupervised dimensionality reduction and visualization of single-cell transcriptome data. VASC can explicitly model dropout events and find nonlinear hierarchical feature representations of the original data.
[0004] Although these methods have achieved satisfactory results on their respective tasks, they have ignored the correlation between tasks. However, recent studies have shown that multi-task learning has significant advantages in knowledge sharing, reducing overfitting, feature learning, and collaborative optimization, i.e. joint optimization of multiple tasks at the same time can improve the performance of each task. In this context, developing a suitable multi-task learning framework and designing an appropriate loss function to balance the weights between different tasks to capture the correlation between tasks and ultimately improve the performance of each task is of great significance for data analysis. SUMMARY
[0005] In view of the above problems, the present application is proposed.
[0006] Therefore, the technical problem solved by the present application is:
[0007] To solve the above technical problems, the present application provides the following technical scheme: a multi-task cell analysis method based on residual graph neural network, comprising: normalizing single cell transcriptome data, selecting the top 2000 genes with the highest transcription level from the normalized data, and obtaining new single cell transcriptome data.
[0008] According to the new single cell transcriptome data, a denoising autoencoder is constructed and trained, the original single cell transcriptome data is reduced in dimension, and a reduced feature representation is obtained.
[0009] An adjacency matrix is constructed using the reduced feature representation, and a residual graph neural network model is constructed.
[0010] The graph neural network model is connected to the denoising autoencoder, a double self-supervised model is constructed and trained.
[0011] According to the double self-supervised model, the clustering results, the interpolation results and the low-dimensional representation of the single cell transcriptome data are output.
[0012] As a preferred scheme of the multi-task cell analysis method based on residual graph neural network, wherein: the normalization includes a Seurat standard quality control process, and new single cell transcriptome data X is obtained.
[0013] As a preferred scheme of the multi-task cell analysis method based on residual graph neural network, wherein: the denoising autoencoder includes adding random Gaussian noise X corrupt , represented as:
[0014] X corrupt =X+e(noise)
[0015] Wherein, e(noise) represents the added Gaussian noise, X corrupt is input into the denoising autoencoder, and the feature learned by the lth layer of the encoder in the denoising autoencoder is represented as:
[0016]
[0017] Wherein represents the feature learned by the previous layer of the lth layer of the encoder, represents the weight of the corresponding layer in the encoder, represents the bias of the corresponding layer in the encoder, and sigma() represents the activation function Relu.
[0018] Let the output of the decoder be denoted as the reduced dimension feature E, and the output of each layer of the decoder be:
[0019]
[0020] where, denotes the weight corresponding to the decoder, denotes the bias corresponding to the decoder, denotes the output of the previous layer of the l-th layer of the decoder.
[0021] The reconstruction loss, i.e., the MSE loss, is:
[0022]
[0023] where N denotes the number of cells, x i denotes the original data of the i-th cell, denotes the reconstructed data of the i-th cell. The denoising loss of the ZINB as the denoising autoencoder is denoted as:
[0024]
[0025] ZINB(X|π, u, θ) = πδ0(X) + (1 - π)NB(X|u, θ)
[0026] where u denotes the mean of the negative binomial distribution, θ denotes the dispersion of the negative binomial distribution, and π denotes the weight of the probability mass at zero. δ0() denotes the Dirac function.
[0027] The matrix form of the three parameters is denoted as:
[0028]
[0029]
[0030]
[0031] The denoising loss function is the negative logarithm of the ZINB likelihood, denoted as:
[0032] L denoise = -log(ZINB(X|π, u, θ)
[0033] The total loss function of the denoising autoencoder is denoted as:
[0034] L auto = L restruction + L denoise
[0035] where, s represents the output of the last layer of the decoder i represents the hyperparameter for controlling the value of M, diag() represents the diagonal matrix function, exp() represents the exponential function, the sigmod() function represents the activation function, W u represents the matrix weight of u, W θ represents the matrix weight of θ, W π represents the matrix weight of π.
[0036] As a preferred scheme of the multi-task cell analysis method based on the residual graph neural network according to the present application, wherein: the adjacency matrix comprises, using the noise reduction autoencoder to obtain the gene weighted joint representation E of the single cell transcriptome data, for each cell cell i , using the features in E to find the k nearest neighbors of cell i The adjacency matrix is represented as:
[0037]
[0038] Wherein, t is a hyperparameter for controlling the size of the similarity, x i represents the original data of the i-th cell, x j represents the original data of the j-th cell.
[0039] As a preferred scheme of the multi-task cell analysis method based on the residual graph neural network according to the present application, wherein: the residual graph neural network model comprises, inputting X into a fully connected neural network with 256 hidden units to obtain Z 0 :
[0040] Z 0 = FC(X)
[0041] Wherein, FC() function represents a layer of fully connected neural network with 256 hidden units.
[0042] RGCN learns a new representation Z l , Z l The representation is calculated using the following convolution operation:
[0043]
[0044] Wherein σ represents the activation function, W l-1 represents the weight of the l-1 layer of the residual graph neural network, Z l-1 represents the output of the l-1 layer of the residual graph neural network, a l represents a hyperparameter, b l represents a hyperparameter, A represents the adjacency matrix, I represents the unit matrix, A normalized adjacency matrix is denoted.
[0045] As a preferred scheme of the multi-task cell analysis method based on the residual graph neural network, the double self-supervised model comprises connecting the decoder output of the denoising autoencoder and the output of the residual graph neural network with a double self-supervised loss function, using Student's t-distribution to calculate the similarity between the cell feature representation and the cluster center vector, and the cluster center is obtained by initial clustering of data using the K-means algorithm.
[0046] The calculation of Student's t-distribution is represented as:
[0047]
[0048] h i denotes the i-th row of E, u j denotes the j-th cluster center obtained by K-means clustering, v represents the degree of freedom of Student's t-distribution, q ij denotes a soft assignment.
[0049] The calculation of a target distribution is represented as:
[0050]
[0051] where f j =∑ i q ij denotes the frequency of the j-th soft clustering, f j' denotes the frequency belonging to other soft categories, j' represents other non-j categories, q ij ' represents the probability of cell i being assigned to category j'.
[0052] The objective function is represented as:
[0053]
[0054] KL(P||Q) represents the difference between the two distributions P and Q, through back propagation and optimization algorithm, the model gradually adjusts the distribution Q to make it closer to the distribution P, so as to realize the optimization of the target task.
[0055] As a preferred scheme of the multi-task cell analysis method based on the residual graph neural network, the output of the single cell transcriptome data clustering result, the interpolation result and the low-dimensional representation comprises: the output result of the residual graph neural network is the clustering result, the encoder output result of the denoising autoencoder is the gene interpolation result, and the decoder output result of the denoising autoencoder is the low-dimensional embedding representation of the feature.
[0056] A multi-task cell analysis system based on a residual graph neural network, characterized by comprising,
[0057] A preprocessing module normalizes single-cell transcriptome data, selects the top 2000 genes with the highest transcription level from the normalized data, and obtains new single-cell transcriptome data.
[0058] A model construction module constructs and trains a denoising autoencoder according to the new single-cell transcriptome data, reduces the dimension of the original single-cell transcriptome data, and obtains a reduced dimension feature representation. An adjacency matrix is constructed using the reduced dimension feature representation, a residual graph neural network model is constructed, and the graph neural network model and the denoising autoencoder are connected to construct a double self-supervised model and train it.
[0059] An output result module outputs the clustering result, the interpolation result, and the low-dimensional representation of the single-cell transcriptome data according to the double self-supervised model.
[0060] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the method as described above when executing the computer program.
[0061] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the method as described above.
[0062] The interpolation capability of the present application allows the discovery of relationships between genes and co-transcribed genes in the original data, the denoising autoencoder can simulate the "dropout" phenomenon in single-cell transcriptome data, which greatly improves the discriminability of the features extracted by the network, integrates the learning goals of multiple tasks into a unified function, and allows coordination between tasks, the process uses a double self-supervised module to connect the residual graph neural network module and the autoencoder model, effectively capturing the relationships between cells and the characteristics of single-cell transcriptome data. These two important types of information are complementary and can improve the performance of each single-cell analysis task. BRIEF DESCRIPTION OF DRAWINGS
[0063] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor. Among them:
[0064] Figure 1A residual graph neural network-based multi-task cell analysis method and system provided for the first embodiment of the present application are shown in the overall flowchart.
[0065] Figure 2 A residual graph neural network-based multi-task cell analysis method and system provided for the first embodiment of the present application are shown in the overall flowchart. DETAILED DESCRIPTION
[0066] In order to make the above objectives, features and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should fall within the protection scope of the present application.
[0067] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below.
[0068] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or selective embodiment that excludes other embodiments.
[0069] The present application is described in detail in conjunction with the schematic diagram, and in the detailed description of the embodiments of the present application, the cross-sectional view of the device structure is locally enlarged without the general proportion for the convenience of description, and the schematic diagram is only an example, which should not limit the scope of protection of the present application herein. In addition, the three-dimensional spatial dimensions of length, width and depth should be included in actual manufacture.
[0070] Meanwhile, in the description of the present application, it should be noted that the terms "up, down, in and out" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the indicated device or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first, second or third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.
[0071] Unless otherwise defined, the terms "mounting, connecting, associating" in the present application should be interpreted broadly, for example: it can be fixed connection, detachable connection or integral connection; it can also be mechanical connection, electrical connection or direct connection, it can also be indirectly connected through intermediate medium, or it can be the internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.
[0072] Embodiment 1
[0073] Reference Figure 1 、 Figure 2 For an embodiment of the present application, a multi-task cell analysis method based on residual graph neural network is provided, comprising:
[0074] S1: normalizing single cell transcriptome data, selecting the top 2000 genes with the highest transcription level from the normalized data, and obtaining new single cell transcriptome data.
[0075] It should be noted that the single cell transcription data is derived from the database after the single cell RNA sequencing experiment.
[0076] Further, the genes with the highest transcription level are usually more functionally meaningful in biology, as they can be important markers of cell state and cell type, and these genes can better reflect the differences and heterogeneity between cells. Selecting the top part of representative genes can reduce computational complexity and improve analysis speed.
[0077] Normalization includes a quality control process of Seurat standard normalization program to obtain new single cell transcriptome data X.
[0078] It should be noted that the normalization process is represented as:
[0079]
[0080] Wherein, X raw represents the original single cell transcriptome data, μ represents the mean of the original data, σ represents the standard deviation of the original data, X normalized represents the normalized single cell transcriptome data.
[0081] Further, the normalization tool mainly comes from SeuratV4 software, that is, using the PercentageFeatureSet function to filter cells with mitochondrial gene proportion greater than 5%, filter cells with UMI number greater than 2500 or less than 200, then use LogNormalize to normalize the original data, and Findvariablefeature function to find the top 2000 highly expressed genes.
[0082] S2: Construct and train a denoising autoencoder according to the new single-cell transcriptome data, and reduce the dimension of the original single-cell transcriptome data to obtain a reduced dimension feature representation.
[0083] The denoising autoencoder includes adding random Gaussian noise X corrupt , which is represented as:
[0084] X corrupt = X + e(noise)
[0085] where e(noise) represents the added Gaussian noise, and X corrupt is input into the denoising autoencoder. The feature learned by the lth layer of the encoder in the denoising autoencoder is represented as:
[0086]
[0087] where represents the feature learned by the previous layer of the lth layer of the encoder, represents the weight of the corresponding layer in the encoder, represents the bias of the corresponding layer in the encoder, and σ() represents the activation function Relu.
[0088] The output of the decoder is recorded as the reduced dimension feature E, and the output of each layer of the decoder is:
[0089]
[0090] where represents the corresponding weight of the decoder, represents the corresponding bias of the decoder, represents the output of the previous layer of the lth layer of the decoder.
[0091] The reconstruction loss, i.e., the MSE loss, is:
[0092]
[0093] where N represents the number of cells, x i represents the original data of the ith cell, represents the reconstructed data of the ith cell.
[0094] The ZINB is used as the denoising loss of the denoising autoencoder, which is represented as:
[0095]
[0096] ZINB(X|π, u, θ) = πδ0(X) + (1-π)NB(X|u, θ)
[0097] Wherein, u represents the mean of negative binomial distribution, theta represents the dispersion of negative binomial distribution, pi represents the weight of probability mass at zero point, delta0() represents Dirac function.
[0098] The matrix form of three parameters is represented as:
[0099]
[0100]
[0101]
[0102] Further, since the ZINB distribution model is highly sparse and over-dispersed, similar to the original single-cell RNA sequencing data, the ZINB model is used to construct a loss function to denoise the single-cell RNA sequencing data.
[0103] The denoising loss function is the negative logarithm representation of ZINB likelihood:
[0104] L denoise = -log(ZINB(X|pi,u,theta)
[0105] The total loss function of the denoising autoencoder is represented as:
[0106] L auto = L restruction + L denoise
[0107] It should be noted that the autoencoder in the present application includes 3 layers of encoder and 3 layers of decoder, the number of neurons of the decoder is 2000, 256, 32 respectively, the number of neurons of the decoder is 32, 256, 2000 respectively, and the activation function uses Relu. The denoising autoencoder is first pre-trained using the Adam algorithm, and the learning rate, batch size, noise value (noise level control) and iteration number of pre-training are set to 10 -4 , 64, 1 and 1000 respectively.
[0108] Wherein, represents the output of the last layer of the decoder, s i represents a hyperparameter used to control the value of M, diag() represents a diagonal matrix function, exp() represents an exponential function, and sigmod() function represents an activation function, W u represents the matrix weight of u, W θ represents the matrix weight of theta, and W π represents the matrix weight of pi.
[0109] S3: using the reduced dimension feature representation to construct an adjacency matrix, and constructing a residual graph neural network model.
[0110] The adjacency matrix includes, using the denoising autoencoder to obtain a gene-weighted joint representation E of the single-cell transcriptome data, for each cell cell i , using the features in E to find the k nearest neighbors of cell i The recent k neighbors, construct an adjacency matrix representation as:
[0111]
[0112] where t is a hyperparameter controlling the size of the similarity, x i represents the original data of the i-th cell, x j represents the original data of the j-th cell.
[0113] It should be noted that based on the constructed neighborhood graph S, the present application further selects the top 5 nearest neighbors of each cell to construct an undirected nearest neighbor graph. Finally, if two edges are connected, set to 1, and not connected, set to 0, this can obtain the adjacency matrix A.
[0114] The residual graph convolutional neural network model includes inputting X into a fully connected neural network (FC) with 256 hidden units to obtain Z 0 :
[0115] Z 0 = FC(X)
[0116] where the FC() function represents a layer of a fully connected neural network with 256 hidden units.
[0117] The RGCN learns a new representation Z l (l is the corresponding layer).
[0118] It should be noted that Z l can fuse two different types of information, namely the relationship between cells and the data itself.
[0119] Z l is calculated using the following convolution operation:
[0120]
[0121] where σ represents an activation function, W represents a weight matrix, W l-1 represents the weight of the (l-1)-th layer of the residual graph neural network, Z l -1 represents the output of the (l-1)-th layer of the residual graph neural network.
[0122] It should be noted that the two hyperparameters a l > 0 ensure that each layer can retain information from the input layer Z 0 , and bl >0 ensures that the weight matrix is adaptively decaying in the stacked layers, A denotes an adjacency matrix, and I denotes an identity matrix, denotes a normalized adjacency matrix.
[0123] Further, it means that in the graph neural network, the node representation of the current layer can be generated by multiplying the node representation of the previous layer with the normalized adjacency matrix. This process allows the nodes to propagate information in the graph and learn a higher-level representation. In the present invention, the number of neurons in the graph neural network is 256, and the number of layers in the residual neural network is set to 5 layers.
[0124] S4: connecting the graph neural network model with the denoising autoencoder, constructing a double self-supervised model and training.
[0125] The double self-supervised model includes connecting the decoder output of the denoising autoencoder with the output of the residual graph neural network with a double self-supervised loss function, using Student's t-distribution to calculate the similarity between the cell feature representation and the cluster center vector, and the cluster center is obtained by initial clustering of the data using the K-means algorithm.
[0126] The calculation of Student's t-distribution is represented as:
[0127]
[0128] h i denotes the i-th row of E, u j denotes the j-th cluster center obtained by K-means clustering, v denotes the degree of freedom of Student's t-distribution, and q ij denotes a soft assignment.
[0129] It should be noted that the soft assignment is the probability of assigning cell i to cluster j. In cluster analysis, there are usually two ways of hard assignment and soft assignment. Hard assignment refers to assigning each data point (e.g., cell) to a specific cluster, which is a binary decision, i.e., a data point either belongs to a cluster or not. This assignment method produces clear, non-fuzzy results. Soft assignment refers to assigning each data point to the probability or weight of each cluster, which allows a data point to have multiple cluster memberships, and this relationship is expressed in the form of probability. In order to improve the cohesion of the cluster, the present invention needs to make the feature representation of the cell closer to the center of the cluster.
[0130] The calculation of a target distribution is represented as:
[0131]
[0132] where f j =∑ i q ij , represents the frequency of the jth soft cluster, f j' represents the frequency belonging to other soft categories, j' represents other non-j categories, q ij ' represents the probability of cell i being assigned to category j'.
[0133] Further, inspired by the idea of self-training, the present application uses the distribution P to make the autoencoder module learn a powerful representation for the cell clustering task.
[0134] The objective function is represented as:
[0135]
[0136] Further, the target distribution P is generated by the distribution Q, and then the P distribution supervises the update of the distribution Q in turn. This process is commonly used to train models to make the generated distribution Q as close as possible to the target distribution P, so as to optimize the performance of the model. KL(P||Q) represents the difference between the two distributions P and Q, through back propagation and optimization algorithm, the model gradually adjusts the distribution Q to make it closer to the distribution P, so as to realize the optimization of the target task.
[0137] Further, this is a common training strategy for unsupervised learning or semi-supervised learning.
[0138] It should be noted that the training end condition is that the training is ended when the training is up to the 1000th round, otherwise the training will continue until the proportion of cells assigned to clusters is less than 0.0001 (which indicates that the network is close to convergence and the result tends to be stable). The calculation method of clustering is represented as:
[0139]
[0140] where Y curr represents the clustering label obtained in the current step, Y prev represents the clustering label obtained in the previous step, represents the number of cells different from Y curr and Y prev , and n represents the total number of all cells.
[0141] It should be emphasized that the real clustering label is only used for evaluation and does not participate in the training process.
[0142] S5: According to the double self-supervised model, output the clustering result, the interpolation result and the low-dimensional representation of the single cell transcriptome data.
[0143] The output module outputs the clustering result, the imputation result and the low-dimensional representation of the single-cell transcriptome data according to the dual self-supervised model.
[0144] In the above embodiment, the multi-task cell analysis system based on the residual graph neural network further includes:
[0145] The preprocessing module normalizes the single-cell transcriptome data, selects the top 2000 genes with the highest transcription level from the normalized data, and obtains new single-cell transcriptome data.
[0146] The model construction module constructs and trains a denoising autoencoder according to the new single-cell transcriptome data, reduces the dimension of the original single-cell transcriptome data, and obtains a reduced feature representation. An adjacency matrix is constructed using the reduced feature representation, and a residual graph neural network model is constructed. The graph neural network model and the denoising autoencoder are connected to construct a dual self-supervised model and train the same.
[0147] The output module outputs the clustering result, the imputation result and the low-dimensional representation of the single-cell transcriptome data according to the dual self-supervised model.
[0148] The computer device can be a server. The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The database of the computer device is used to store the data cluster data of the power monitoring system. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a method.
[0149] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-mentioned embodiments of each method. Any reference to memory, database or other medium used in each embodiment provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The database involved in each embodiment provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in each embodiment provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.
[0150] Embodiment 2
[0151] In one embodiment of the present application, a multi-task cell analysis method and system based on residual graph neural network are provided. In order to verify the beneficial effects of the present application, a comparative experiment is carried out for scientific demonstration.
[0152] A set of untreated single-cell transcriptome data (containing noise) is processed by the method of the present application and the traditional method respectively. In the present application, after normalization, denoising autoencoder training, construction of adjacency matrix and graph neural network, and double self-supervised model training, the trained model is used to cluster, interpolate and reduce the dimension of the single-cell transcriptome data, and compared with the traditional method.
[0153] Table 1 Comparison of clustering accuracy
[0154] Method Cluster accuracy Conventional method 0.85 Method of the invention 0.95
[0155] Table 2 Comparison of imputation effect on adjusted Land index
[0156] Method Adjusted rand index Conventional method 0.72 Method of the invention 0.86
[0157] Table 3 Comparison of calculation efficiency
[0158] Method Computation time (seconds) Conventional method 308 Method of the invention 147
[0159] From Tables 1-3, it can be seen that the present application has advantages in clustering accuracy, imputation accuracy and calculation efficiency.
[0160] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.
Claims
1. A multi-task cell analysis method based on residual graph neural networks, characterized in that, Comprise: Obtaining single-cell transcriptome data and normalizing, selecting the top 2000 genes with the highest transcription level from the normalized data to obtain new single-cell transcriptome data; According to the new single-cell transcriptome data, a denoising autoencoder is constructed and trained to reduce the dimension of the original single-cell transcriptome data and obtain a reduced feature representation; The denoising autoencoder comprises adding random Gaussian noise X corrupt is expressed as: X corrupt = X + e(noise) where e(noise) represents the added Gaussian noise, and X corrupt is input into the denoising autoencoder, and the feature representation learned by the l-th layer of the encoder in the denoising autoencoder is denoted as wherein represents the features learned by the previous layer of the lth layer of the encoder, represents the weights of the corresponding layer in the encoder, represents the bias of the corresponding layer in the encoder, and σ() represents the activation function Relu; The output of the last layer is denoted as the reduced feature E, and the output of each layer of the decoder is: wherein, represents a weight corresponding to the decoder, represents a bias corresponding to the decoder, represents an output of a previous layer of the l-th layer of the decoder; The reconstruction loss, i.e. the MSE loss, is: where N represents the number of cells, x i represents the original data of the i-th cell, represents the reconstructed data of the i-th cell; the denoising loss using ZINB as the denoising autoencoder is represented as: ZINB(X|π,u,θ)=πδ0(X)+(1-π)NB(X|u,θ) Where u represents the mean of the negative binomial distribution, θ represents the dispersion of the negative binomial distribution, π represents the weight of the probability particle at zero, and δ0() represents the Dirac function; The matrix representation of the three parameters is: The denoising loss function is the negative logarithmic representation of the ZINB likelihood: L denoise = -log(ZINB(X | π, u, θ) The total loss function of the denoising autoencoder is: L auto = L restruction + L denoise wherein, represents the output of the last layer of the decoder, s i represents a hyperparameter used to control the value of M, diag() represents a diagonal matrix function, exp() represents an exponential function, the sigmod() function represents an activation function, W u represents the matrix weight of u, W θ represents the matrix weight of θ, W π represents the matrix weight of π; An adjacency matrix is constructed using the reduced feature representation, and a residual graph neural network model is constructed; The graph neural network model and the denoising autoencoder are connected to construct a dual self-supervised model and train it; According to the dual self-supervised model, the clustering results, the interpolation results and the low-dimensional representation of the single-cell transcriptome data are outputted.
2. The residual graph neural network-based multi-task cell analysis method of claim 1, wherein: The normalization includes a normalization program SeuratSeurat standard quality control process to obtain new single-cell transcriptome data X.
3. The residual graph neural network-based multi-task cell analysis method of claim 2, wherein: The adjacency matrix comprises, using the denoising autoencoder to obtain a gene-weighted joint representation E of the single-cell transcriptome data, for each cell cell i , finding the k nearest neighbors to cell i cell in E, constructing an adjacency matrix representation denoted as: where t is a hyperparameter controlling the magnitude of the similarity, x i represents the original data of the i-th cell, x j represents the original data of the j-th cell.
4. The residual graph neural network-based multi-task cell analysis method of claim 3, wherein: The residual graph neural network model includes inputting X into a fully connected neural network with 256 hidden units to obtain Z 0 : Z 0 = FC(X) Where FC() function represents a layer of fully connected neural network with 256 hidden units; RGCN learns a new representation Z l , Z l The representation is computed using the following convolution operation: where σ denotes an activation function, W l-1 denotes the weights of the l-1th layer of the residual graph neural network, Z l-1 denotes the output of the l-1th layer of the residual graph neural network, a l denotes a hyperparameter, b l denotes a hyperparameter, A denotes an adjacency matrix, I denotes an identity matrix, denotes a normalized adjacency matrix.
5. The residual graph neural network-based multi-task cell analysis method of claim 4, wherein: The dual self-supervised model includes connecting the decoder output of the denoising autoencoder and the output of the residual graph neural network with a dual self-supervised loss function, using Student's t-distribution to calculate the similarity between the cell feature representation and the cluster center vector, and the cluster center is obtained by initial clustering of the data using K-means algorithm; The calculation of Student's t-distribution is: h i represents the i-th row of E, u j represents the j-th cluster center of K-means clustering, v represents the degrees of freedom of student t-distribution, q ij represents a soft assignment; The calculation of a target distribution is: where f j =∑ i q ij , represents the frequency of the jth soft cluster, f j' represents the frequency belonging to other soft classes, j' represents other non-j classes, q ij' represents the probability of cell i being assigned to class j'; The objective function is: KL(P||Q) represents the difference between the two distributions P and Q, through back propagation and optimization algorithm, the model gradually adjusts the distribution Q to make it closer to the distribution P, so as to realize the optimization of the target task.
6. The residual graph neural network-based multi-task cell analysis method of claim 5, wherein: The output of the single-cell transcriptome data includes the output results of the residual graph neural network as the clustering results, the encoder output results of the denoising autoencoder as the gene interpolation results, and the decoder output results of the denoising autoencoder as the low-dimensional embedding representation of the features.
7. A residual graph neural network-based multi-task cell analysis system using the method of any one of claims 1-6, characterized in that: The preprocessing module normalizes the single-cell transcriptome data, selects the top 2000 genes with the highest transcription level from the normalized data to obtain new single-cell transcriptome data; The model construction module constructs and trains a denoising autoencoder according to the new single-cell transcriptome data, reduces dimensions of the original single-cell transcriptome data, and obtains a feature representation after dimension reduction; constructs an adjacency matrix using the feature representation after dimension reduction, and constructs a residual graph neural network model; connects the graph neural network model and the denoising autoencoder, constructs a double self-supervised model, and trains the double self-supervised model. The output result module outputs a clustering result, an interpolation result, and a low-dimensional representation of single-cell transcriptome data according to the double self-supervised model.
8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 6.
9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 6.