GCN-LSTM reconstruction time sequence gene regulation and control network based on adaptive layer number

Abstract

The invention discloses a GCN-LSTM reconstruction time sequence gene regulation and control network based on an adaptive layer number. The method comprises the steps that gene expression matrixes of different time points and a prior gene regulation and control network are used for extracting corresponding gene features of different time points from GCN of the adaptive layer number; inputting the gene features of different time points into LSTM to obtain the gene feature of the next time point; and a directed gene regulatory network of the next time point is predicted in the GAE inspired by gravity based on the gene characteristics of the next time point and the prior gene regulatory network. According to the method, the GCN with the self-adaptive layer number is adopted, and the GCN layer number can be self-adapted according to different sizes of data sets; according to the method, richer gene characteristics between time points are learned by virtue of LSTM, and a directed gene regulatory network (GRN) is obtained by adopting a graph auto-encoder (GAE) inspired by gravity, so that the method has higher cell type specific reasoning accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of reconstructing a directed gene regulatory network for a time-series single-cell RNA sequencing (scRNA-Seq) data set, and in particular to a GCN-LSTM based on an adaptive number of layers to reconstruct a time-series gene regulatory network. Background Art

[0002] Gene regulatory networks are causal regulatory relationships between transcription factors (TFs) and target genes. They play an important role in clarifying gene functions, explaining biological processes, and analyzing biomarkers of complex diseases. In the past two decades, people have focused on inferring gene regulatory relationships and signal transduction from gene expression as a hot research area. However, many existing tools are mainly targeted at bulk and single-cell static RNA sequencing data ( Figure 1 A) and bulk single-cell time-sequencing RNA sequencing data ( Figure 1 B) There are relatively few tools designed specifically for time-series single-cell RNA sequencing data.

[0003] Compared with bulk RNA sequencing data, single-cell RNA sequencing data has obvious advantages. For example, single-cell sequencing focuses on the gene expression of a single cell rather than the average of all cell expressions, so it can better reflect the specificity of the cell; in addition, the amount of single-cell sequencing data is usually large, which allows the model to learn more useful features from it. At the same time, time-series RNA sequencing data provides richer gene expression information than static sequencing data, and can more accurately infer the relationship between genes.

[0004] TDL is a model that converts time-series single-cell RNA sequencing data into three-dimensional histograms and then applies convolutional neural networks (CNNs) for regulatory classification. dynDeepDRIM encodes gene pairs in the same way as TDL. It regards the image of the target transcription factor-gene pair as the main image and the images of other gene pairs that share the same gene with the target gene pair as neighbor images. These two images are used as input to the deep learning model to predict the regulatory relationship between the target gene pairs. However, the process of converting gene expression data into histograms is time-consuming, may introduce some unexpected noise, and the image representation may mask some data features. In addition, these models do not fully consider the topological structure of the prior gene regulatory network, and only learn the first-order neighbor features directly connected to the gene, but fail to capture the second-order and higher-order neighbor features indirectly connected to the gene. In contrast, GMFGRN converts the gene expression matrix into a heterogeneous graph of genes and cells, uses a graph convolutional network (GCN) to learn a low-dimensional embedding representation on the heterogeneous graph, and finally performs classification through a multi-layer perceptron (MLP). However, GMFGRN does not actually utilize the time-series single-cell dataset, but simply connects the gene expression matrices at each time point together, failing to fully utilize the complete time dimension information.

[0005] The invention application with application number 202310765180.1 discloses a method for constructing a gene regulatory network based on dynamic cell differentiation. The application scheme combines pseudo-time information with cell types and takes into account the distribution of cell density along pseudo-time, which helps to construct a cell type-specific gene regulatory network from the perspective of dynamic cell differentiation, making the constructed gene regulatory network more accurate. However, it also has the following problems: it cannot effectively capture the second-order and higher-order neighbor features indirectly connected to genes, and cannot predict directed gene regulatory networks.

[0006] The invention application with application number 202310564195.1 discloses a method for constructing a gene regulatory network based on dynamic cell differentiation. This method combines pseudo-time information with cell types and takes into account the distribution of cell density along pseudo-time, which helps to construct a cell type-specific gene regulatory network from the perspective of dynamic cell differentiation, making the constructed gene regulatory network more accurate. However, it also has the following problems: it cannot effectively capture the second-order and higher-order neighbor features that are indirectly connected to genes, and cannot predict directed gene regulatory networks. Summary of the invention

[0007] In view of the above-mentioned problems, the purpose of the present invention is to provide a GCN-LSTM based on an adaptive number of layers to reconstruct a temporal gene regulatory network, use the GCN-LSTM with an adaptive number of layers to infer the gene characteristics of the next time point, and predict a directed gene regulatory network (GRN).

[0008] The embodiment of the present invention provides a GCN-LSTM reconstruction time series gene regulation network based on adaptive number of layers, including:

[0009] First aspect:

[0010] The GCN-LSTM based on adaptive layers reconstructs the temporal gene regulatory network. The steps include:

[0011] S1, Gene expression matrix using different time points;

[0012] S2, extract the gene expression matrix at different time points and the prior gene regulatory network into the corresponding gene features at different time points in the GCN with adaptive layers;

[0013] S3, the gene features at different time points are input into LSTM to obtain the gene features at the next time point;

[0014] S4. Based on the gene features at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in the gravity-inspired GAE.

[0015] Furthermore, the GCN adaptive layer operation logic is:

[0016] S21, the number of GCN model layers initialized is the current number of GCN model layers, and the optimal AUROC is initialized;

[0017] S22, train the GCN model of the current number of layers and calculate the AUROC to obtain the AUROC of the current number of layers;

[0018] S23, judging whether the current AUROC is better than the best AUROC;

[0019] S24, if the current AUROC is better than the best AUROC, update the best AUROC and the current number of layers, and repeat steps S22-S24;

[0020] S25. If the current AUROC is not better than the best AUROC, the loop ends.

[0021] Furthermore, the gene features at different time points in S3 are input into LSTM to obtain the gene features at the next time point, and the process includes:

[0022] Genetic features at different time points Feed it to the LSTM model. For a certain time step t, the LSTM unit takes the current input vector E t and the state vector h at the previous time step t-1 t-1 As input, it then outputs the state vector h of the current time step t , the state update formula is:

[0023]

[0024] Among them, i t is the input gate, f t For the forget gate, t is the output gate, s t is the storage unit, σ(·) is the sigmoid activation function, ⊙ is the element-by-element multiplication, {W x ,W h ,b} are the parameters of the corresponding unit;

[0025] The last hidden state h t Input the fully connected layer to generate the gene feature E at the next time point t+1 .

[0026] Further, the S4 includes:

[0027] S41, encode the gene features at the next time point and the prior gene regulatory network using a gravity-inspired encoder to form the gene potential parameter feature Z'

[0028] S42. Decode the gene potential Z' using a gravity-inspired decoder to obtain a directed gene regulatory network.

[0029] Furthermore, the gene potential parameter characteristic Z' formed by the encoding is expressed as

[0030]

[0031] Among them, Z is the potential vector, n*d dimension; is the quality parameter, n*1 dimension; is the out-degree normalization of the prior gene regulatory network A, and I is the out-degree normalization variable.

[0032] Furthermore, the decoding obtains a directed gene regulatory network, which is expressed as

[0033]

[0034] Among them, a i-j is the acceleration value from node i to node j, r 2 is the square of the distance between node i and node j, m j is the mass of node j.

[0035] A second aspect: An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method provided in the first aspect when executing the program.

[0036] A third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0037] Beneficial effects of the present invention:

[0038] 1. For time-series single-cell RNA sequencing datasets, the size of the prior gene regulatory network (GRN) and the number of gene expression matrices formed by different datasets are different. If a fixed number of layers of GCN is used, the model will not be able to achieve the best accuracy. The present invention uses a GCN with an adaptive number of layers, which can adapt the number of GCN layers according to the size of the dataset.

[0039] 2. The present invention uses LSTM to learn richer gene features between time points, and can use this to infer a more accurate gene expression matrix at the next time point.

[0040] 3. The present invention adopts a gravity-inspired graph autoencoder (GAE) to predict a directed gene regulatory network (GRN). The encoder uses the out-degree normalized version of the prior gene regulatory network (GRN) to ensure the asymmetry of the learned graph; the decoder uses a formula inspired by gravitational acceleration rather than a simple dot product to ensure that the decoded graph is directed.

[0041] 4. Compared with existing methods, the method of the present invention has higher cell type-specific GRN inference accuracy.

[0042] 5. The method of the present invention takes into account the topological structure of genes and does not need to be converted into a three-dimensional histogram, saving time. This method uses the prior GRN and the gene expression matrix at the t+1 time point. Previous methods such as TDL and dynDeepDRIM convert the three-dimensional histogram into the model. Compared with their approach, the model of the method of the present invention does not need to be converted and runs faster. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of single-cell static RNA sequencing data and single-cell time-series RNA sequencing data;

[0044] Figure 2 It is a principle flow chart of the control method of the present invention;

[0045] Figure 3 Schematic diagram of the process of the control method of the present invention;

[0046] Figure 4 This is a comparison experiment diagram of different methods in the embodiments of the present invention;

[0047] Figure 5These are ablation experiment diagrams of different methods according to the embodiments of the present invention;

[0048] Figure 6 It is a schematic structural diagram of the electronic device of the present invention.

[0049] Correspondence between Chinese and English nouns:

[0050] Long Short-Term Memory (LSTM)

[0051] Long Short-Term Memory Network Unit (LSTM Cell)

[0052] Graph Convolutional Network (GCN)

[0053] Adaptive Depth GCN

[0054] Gene expression

[0055] Gene Regulatory Network (GRN)

[0056] Area under the curve (AUROC)

[0057] Prior gene regulatory network (prior GRN)

[0058] Directed gene regulatory network (inferred GRN)

[0059] Graph Autoencoder (GAE)

[0060] Gravity-Inspired Encoder

[0061] Gravity-Inspired Decoder DETAILED DESCRIPTION

[0062] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0063] The existing methods convert gene expression data into histograms in a time-consuming process and introduce some unexpected noise, which may cause the image representation to mask some data features. In addition, the existing methods do not fully consider the topological structure of the prior gene regulatory network and only learn the first-order neighbor features directly connected to the gene, but fail to capture the second-order and higher-order neighbor features indirectly connected to the gene.

[0064] In view of the above problems, the present invention provides a GCN-LSTM reconstruction time-series gene regulatory network based on adaptive layers. Figure 2 The principle flow chart of the GCN-LSTM reconstruction of the temporal gene regulatory network based on the adaptive number of layers provided in the embodiment of the present invention is as follows: Figure 3 Based on Figure 2 Flow chart of the method comprising:

[0065] S1. Obtain the gene expression matrix at different time points of single-cell time-series RNA sequencing data;

[0066] The gene expression matrix X at different time points is expressed as:

[0067]

[0068] Among them, l is the total step length and t is the current time.

[0069] The prior gene regulatory network is represented as matrix A.

[0070] S2. Input the gene expression matrix X at different time points and the prior gene regulatory network A into the GCN with adaptive layers to extract the corresponding gene feature matrix at different time points.

[0071]

[0072] The operating logic of GCN adaptive layer number is:

[0073] S21, the number of GCN model layers initialized is the current number of GCN model layers, and the optimal AUROC is initialized;

[0074] S22, train the GCN model of the current number of layers and calculate the AUROC to obtain the AUROC of the current number of layers;

[0075] S23, judging whether the current AUROC is better than the best AUROC;

[0076] S24, if the current AUROC is better than the best AUROC, update the best AUROC and the current number of layers, and repeat steps S22-S24;

[0077] S25. If the current AUROC is not better than the best AUROC, the loop ends.

[0078] The pseudo code for running the GCN adaptive layer is:

[0079]

[0080] After the adaptive number of layers, GCN obtains the gene feature matrix at different time points, and the formula is expressed as

[0081]

[0082] S3. Gene feature matrix at different time points Input LSTM to obtain the gene feature E at the next time point t+1 ;

[0083] Gene feature matrix at different time points are fed into the LSTM model, which has a strong ability to learn long-term dependencies in sequential data to capture the evolving patterns of the weighted dynamic network.

[0084] like Figure 2 As shown in the figure, the LSTM model can be described as a package unit with multiple multiplication gate units; for a certain time step t, the LSTM unit takes the current input vector E t and the state vector h at the previous time step t-1 t-1 As input, it then outputs the current time step state vector h t :

[0085] The LSTM state update formula is:

[0086]

[0087] Among them, i t is the input gate, f t For the forget gate, t is the output gate, s t is the storage unit, σ(·) is the sigmoid activation function, ⊙ is the element-by-element multiplication, {W x ,W h ,b} are the parameters of the corresponding unit;

[0088] Finally, the last hidden state h t Input the fully connected layer to generate the prediction result, the gene feature E at the next time point t+1 .

[0089] The GCN and LSTM with adaptive number of layers together form a GCN-LSTM model with adaptive number of layers.

[0090] S4. Gene signature E based on the next time point t+1 The prior gene regulatory network (GRN) is used to predict the directed gene regulatory network (GRN) at the next time point in the gravity-inspired GAE.

[0091] In traditional GAE, the encoder uses a 2-layer GCN, and the potential vector representation of the node is obtained through the 2-layer GCN.

[0092]

[0093] Among them, the number of GCN layers is L; the weight matrix is ​​W; the feature vector is H; the original adjacency matrix is ​​A; the symmetric normalization of A is The angle matrix D of A+I is obtained because It is symmetric, so the traditional GAE cannot learn the direction information in the directed graph in the encoder part. What is obtained after the 2-layer GCN encoder is the potential vector Z of the node.

[0094] In the decoder of traditional GAE, the decoder part is a simple dot product of the potential vectors of node i and node j, which is then classified into two categories through the sigmoid function. The result of the dot product between node i and node j is mapped to between 0 and 1, and then based on a pre-selected threshold, if it is greater than the threshold, it is considered that there is an edge between node i and node j, and if it is less than the threshold, it is considered that there is no edge between node i and node j.

[0095] In traditional GAE, Z is the latent vector, because Therefore, in the decoder, the directionality of node i and node j cannot be obtained, and it can only be determined whether node i and node j are connected.

[0096] The gravity-inspired GAE is inspired by the universal gravitation between celestial bodies to improve the traditional GAE so that it can perform directed link prediction.

[0097] The principle is: under the action of universal gravitation, the universal gravitation between the earth and the moon is the same, such as the universal gravitation formula:

[0098]

[0099] If we look at the acceleration between the Earth and the Moon, for the Earth, the acceleration formula of the Earth pointing to the Moon is:

[0100]

[0101] For the moon, the acceleration formula of the moon pointing to the earth is:

[0102]

[0103] Among them, F is the universal gravitation; m1 is the mass of the earth; m2 is the mass of the moon; G is the gravitational constant; r is the distance between the two objects.

[0104] Because m1>m2, so a 1→2 2→1 ​, the possibility that the earth points to the moon is smaller than the possibility that the moon points to the earth; when the moon and the earth are abstracted into nodes i and j, it is found that when the universal gravitation formula is used, the calculated universal gravitation values ​​of nodes i and j are the same; but when the acceleration formula is used, the calculated acceleration values ​​of nodes i and j are different, which means that this formula can be used to make a directed relationship between nodes i and j; the calculation formula between nodes i and j abstracted by the acceleration formula is:

[0105]

[0106] Among them, because the larger the numerator, the more likely it is that node i and node j are connected; therefore, in the gene regulatory network, the numerator is more likely to capture more influential genes; there is a mass in the numerator, which corresponds to the denominator in the gene regulatory network as the mass parameter M (mass). Because the denominator is the square of the distance r, the abstraction is It is believed that nodes that are structurally similar are more likely to be connected.

[0107] Consider gravity-inspired effects on the basis of traditional GAE, such as Figure 2 As shown, the process includes:

[0108] S41, encode the gene features at the next time point and the prior gene regulatory network using a gravity-inspired encoder to form a gene potential parameter vector Z'

[0109] S42. Decode the gene parameter vector Z' using a gravity-inspired decoder to obtain a directed gene regulatory network (GRN).

[0110] In the encoder part inspired by gravity, the input prior gene regulatory network A and the gene expression matrix X(E t+1 ), the output of the two-layer GCN is d+1 dimensions, where the first d dimensions are represented as the potential vector Z, and the last 1 dimension is represented as the quality parameter M (mass). The output formula is:

[0111]

[0112] Among them, Z is the potential vector, n*d dimension; is the quality parameter vector, n*1 dimension; It is the out-degree normalized version of the prior gene regulatory network A. I is the out-degree normalized variable, which focuses on the out-degree of the A matrix and is also the reason why the gravity-inspired encoder encoding can be a directed graph.

[0113] In the decoding part of the decoder inspired by gravity, the universal gravitational formula is applied to the decoder, so that the final values ​​of the gene pair (i, j) and the gene pair (j, i) are different, thereby obtaining the directionality of the gene regulatory network, which is expressed as:

[0114]

[0115] Among them, a i-j is the acceleration value from node i to node j, r 2 is the square of the distance between node i and node j, m j is the mass of node j.

[0116] Experimental examples using this method:

[0117] First, a data set is established. This example uses four real data sets and a simulated time-series single-cell RNA sequencing data set, as shown in Table 1:

[0118] Table 1 Time series single-cell RNA sequencing dataset

[0119]

[0120]

[0121] Table 1 benchmarks the GCN-LSTM model. The real datasets include two datasets from mouse embryonic stem cells (mESC1 and mESC2) and two datasets from human embryonic stem cells (hESC1 and hESC2).

[0122] The simulated dataset was generated by dyngen and included 10,000 genes and 20,000 cells.

[0123] During the benchmark test, the selection of positive and negative gene pairs was balanced, with a training set:test set ratio of 4:1.

[0124] Comparative experiment:

[0125] The performance of GCN-LSTM with adaptive number of layers is evaluated and compared with several unsupervised and supervised learning methods. When inferring gene regulatory networks, supervised learning methods use known relationships between genes and gene expression profiles to train the optimal model, while unsupervised methods rely only on gene expression profiles to find the relationship between data and features to infer the relationship between transcription factors and genes.

[0126] With the development of single-cell RNA sequencing technology, more and more gene relationships have been experimentally verified and can be found in databases. Therefore, supervised learning can train models based on these known gene regulatory relationships and predict regulatory relationships between new gene pairs.

[0127] The present invention selects two unsupervised learning methods (PCC and MI) and five supervised learning methods (TDL-3DCNN, TDL-LSTM, dynDeepDRIM, scTGRN and GMFGRN) for comparison.

[0128] Using GCN-LSTM with adaptive layers on real experimental data to infer gene regulatory networks, the experimental results are as follows Figure 5 , the experimental indicator is AUROC, through Figure 5 It can be seen that the model of the present invention performs better than other comparison models on four different data sets.

[0129] Among them, on the mESC2 dataset, the model of the present invention is (0.843), which improves the accuracy by 7.6% over the best performing scTGRN (0.767); on the hESC1 dataset, the model of the present invention is (0.821), which improves the accuracy by 2.3% over the best performing GMFGRN (0.798); on the hESC2 dataset, the model of the present invention is (0.884), which improves the accuracy by 10.0% over the best performing GMFGRN (0.784); on the mESC1 dataset, the model of the present invention is (0.832), which improves the accuracy by 10.8% over the best performing GMFGRN (0.724). In general, the improvement on the mESC1 and hESC2 datasets is relatively large.

[0130] like Figure 6 As shown in the figure, the performance on the simulated data set shows a good improvement compared with other comparison models, with the specific improvements being as follows: 15.8% improvement on the Sim1 data set; 7.3% improvement on the Sim2 data set; 5.7% improvement on the Sim3 data set; and 2.6% improvement on the Sim4 data set. From these results, the model also performs better than other comparison models on the simulated data set, especially showing a significant improvement on the Sim1 data set. This shows that the model of the present invention can maintain strong performance when processing different types of data, demonstrating its potential application value and effectiveness.

[0131] Ablation experiment:

[0132] In order to verify the contribution of the proposed module to the model performance, ablation experiments were conducted. The number of GCN layers was fixed to 2 as the ablation using adaptive GCN layers and the gene expression matrix at time point t=0 was selected as a static dataset and directly input into the gravity-inspired GAE to directly infer the gene regulatory network as the ablation of the dynamic dataset.

[0133] The ablation test performance results under different configurations are given in Table 2. The data in () represents the number of GCN layers with the best results determined by the final model.

[0134] Table 2 Ablation experiment results

[0135]

[0136] Experimental conclusion:

[0137] The results of static data show that although the traditional GAE method has achieved good performance, there is still room for improvement compared with the model of the present invention. The experiment of fixing the number of GCN layers shows that when the number of layers is limited, the model performance is affected to a certain extent, especially when processing complex data; the model of the present invention shows obvious performance improvement on all data sets, especially on Sim4, reaching the highest 0.9161, which shows that the model of the present invention has better adaptability and accuracy in gene regulatory network inference tasks.

[0138] The present invention also provides a GCN-LSTM reconstructing temporal gene regulation device based on adaptive layer number, the device comprising:

[0139] The generation module includes a GCN-LSTM model. The generation module uses the gene expression matrix at different time points, extracts the gene expression matrix at different time points and the prior gene regulatory network in the GCN with an adaptive number of layers to obtain the corresponding gene features at different time points, and inputs the gene features at different time points into the LSTM to obtain the gene features at the next time point.

[0140] The prediction module includes a gravity-inspired GAE model. The generation module predicts the directed gene regulatory network at the next time point in the gravity-inspired GAE based on the gene features at the next time point and the prior gene regulatory network.

[0141] The present invention also provides an electronic device, Figure 6 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 6 As shown, the electronic device may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call the logic instructions in the memory, for example, to execute the following method:

[0142] S1, Gene expression matrix using different time points;

[0143] S2, extract the gene expression matrix at different time points and the prior gene regulatory network into the corresponding gene features at different time points in the GCN with adaptive layers;

[0144] S3, the gene features at different time points are input into LSTM to obtain the gene features at the next time point;

[0145] S4. Based on the gene features at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in the gravity-inspired GAE.

[0146] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0147] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including:

[0148] S1, Gene expression matrix using different time points;

[0149] S2, extract the gene expression matrix at different time points and the prior gene regulatory network into the corresponding gene features at different time points in the GCN with adaptive layers;

[0150] S3, the gene features at different time points are input into LSTM to obtain the gene features at the next time point;

[0151] S4. Based on the gene features at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in the gravity-inspired GAE.

[0152] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0153] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A GCN-LSTM based on adaptive layer number to reconstruct a sequential gene regulatory network, characterized in that the steps include: S1, Gene expression matrix using different time points; S2, extract the gene expression matrix at different time points and the prior gene regulatory network into the corresponding gene features at different time points in the GCN with adaptive layers; S3, the gene features at different time points are input into LSTM to obtain the gene features at the next time point; S4. Based on the gene features at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in the gravity-inspired GAE.

2. The control network according to claim 1, characterized in that: The GCN adaptive layer operation logic is: S21, the number of GCN model layers initialized is the current number of GCN model layers, and the optimal AUROC is initialized; S22, train the GCN model of the current number of layers and calculate the AUROC to obtain the AUROC of the current number of layers; S23, judging whether the current AUROC is better than the best AUROC; S24, if the current AUROC is better than the best AUROC, update the best AUROC and the current number of layers, and repeat steps S22-S24; S25. If the current AUROC is not better than the best AUROC, the loop ends.

3. The control network according to claim 1, characterized in that: The gene features at different time points in S3 are input into LSTM to obtain the gene features at the next time point. The process includes: Genetic features at different time points Feed it to the LSTM model. For a certain time step t, the LSTM unit takes the current input vector E t and the state vector h at the previous time step t-1 t-1 As input, it then outputs the state vector h of the current time step t , the state update formula is: Among them, i t is the input gate, f t For the forget gate, t is the output gate, s t is the storage unit, σ(·) is the sigmoid activation function, ⊙ is the element-by-element multiplication, {W x ,W h ,b} are the parameters of the corresponding unit; The last hidden state h t Input the fully connected layer to generate the gene feature E at the next time point t+1 .

4. The control network according to claim 3, characterized in that: The S4 includes: S41, encode the gene features at the next time point and the prior gene regulatory network using a gravity-inspired encoder to form the gene potential parameter feature Z' S42. Decode the gene potential Z' using a gravity-inspired decoder to obtain a directed gene regulatory network.

5. The control network according to claim 4, characterized in that: The gene potential parameter characteristic Z' formed by the encoding is expressed as Among them, Z is the potential vector, n*d dimension; is the quality parameter, n*1 dimension; is the out-degree normalization of the prior gene regulatory network A, and I is the out-degree normalization variable.

6. The control network according to claim 4, characterized in that: The decoding obtains a directed gene regulatory network, which is expressed as Among them, a i-j is the acceleration value from node i to node j, r 2 is the square of the distance between node i and node j, m j is the mass of node j.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the regulatory network according to any one of claims 1 to 6 are implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the regulatory network according to any one of claims 1 to 6 are implemented.

Images