Reconstructing Temporal Gene Regulation Networks Using GCN-LSTM with Adaptive Layer Number

Abstract

This invention discloses a method for reconstructing a temporal gene regulation network based on an adaptive-layer GCN-LSTM, comprising: extracting gene features corresponding to different time points using gene expression matrices at different time points and a prior gene regulation network in an adaptive-layer GCN; inputting the gene features at different time points into an LSTM to obtain gene features at the next time point; and predicting the directed gene regulation network at the next time point in a gravity-inspired graph autoencoder (GAE) based on the gene features at the next time point and the prior gene regulation network. This invention employs an adaptive-layer GCN, which can adapt the number of GCN layers according to different dataset sizes; it utilizes LSTM to learn richer gene features between time points; and it uses a gravity-inspired graph autoencoder (GAE) to obtain the directed gene regulation network (GRN), resulting in higher cell type-specific inference accuracy.

Description

Technical Field

[0001] This invention relates to the technical field of reconstructing directed gene regulatory networks for time-series single-cell RNA sequencing (scRNA-Seq) datasets, and particularly to a GCN-LSTM method for reconstructing time-series gene regulatory networks based on adaptive layer number. Background Technology

[0002] Gene regulatory networks, which define the causal relationships between transcription factors (TFs) and target genes, play a crucial role in elucidating gene function, explaining biological processes, and analyzing biomarkers of complex diseases. Over the past two decades, inferring gene regulatory relationships and signal transduction from gene expression has been a hot research area. However, many existing tools primarily target bulk and single-cell static RNA sequencing data (…). Figure 1 A) and bulk single-cell time-series RNA sequencing data ( Figure 1 (B) There are relatively few tools specifically designed for time-series single-cell RNA sequencing data.

[0003] Compared to bulk RNA sequencing data, single-cell RNA sequencing data has significant advantages. For example, single-cell sequencing focuses on gene expression in a single cell, rather than the average expression across all cells, thus better reflecting cell specificity. Furthermore, single-cell sequencing data is typically larger, allowing models to learn more useful features. At the same time, time-series RNA sequencing data provides richer gene expression information than static sequencing data, enabling more accurate inferences about gene relationships.

[0004] TDL (Translation Deep Learning) 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 similarly to TDL, treating the image of the target transcription factor-gene pair as the master image and images of other gene pairs sharing the same gene as the target pair as neighbor images. Both images serve as input to the deep learning model to predict the regulatory relationships between target gene pairs. However, converting gene expression data into histograms is time-consuming, may introduce unexpected noise, and the image representation may mask some data features. Furthermore, these models do not fully consider the topology of prior gene regulatory networks, learning only first-order neighbor features directly connected to genes, failing to capture second-order and higher-order neighbor features indirectly connected to genes. In contrast, GMFGRN converts the gene expression matrix into a gene-cell heterogeneous graph, uses graph convolutional networks (GCNs) to learn low-dimensional embedding representations on the heterogeneous graph, and finally performs classification using a multilayer perceptron (MLP). However, GMFGRN does not actually make full use of time-series single-cell datasets. Instead, it simply connects the gene expression matrices at each time point, failing to fully utilize the complete time dimension information.

[0005] Invention application No. 202310765180.1 discloses a method for constructing a gene regulatory network based on dynamic cell differentiation. This method combines pseudo-temporal information with cell type and considers the distribution of cell density along the pseudo-temporal path, which helps to construct cell type-specific gene regulatory networks from the perspective of dynamic cell differentiation, making the constructed gene regulatory networks more accurate. However, it also has the following problems: it cannot effectively capture the features of second-order and higher-order neighbors indirectly connected to genes, and it cannot predict directed gene regulatory networks.

[0006] Invention application No. 202310564195.1 discloses a method for constructing a gene regulatory network based on cell dynamic differentiation. This method combines pseudo-temporal information with cell type and considers the distribution of cell density along the pseudo-temporal path, which helps to construct cell type-specific gene regulatory networks from the perspective of cell dynamic differentiation, making the constructed gene regulatory networks more accurate. However, it also has the following problems: it cannot effectively capture the features of second-order and higher-order neighbors indirectly connected to genes, and it cannot predict directed gene regulatory networks. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a GCN-LSTM-based reconstruction of temporal gene regulatory networks with adaptive layer count, which infers gene characteristics at the next time point and predicts directed gene regulatory networks (GRNs).

[0008] This invention provides a GCN-LSTM reconstruction of a temporal gene regulatory network based on adaptive layer number, comprising:

[0009] First aspect:

[0010] The steps of reconstructing a temporal gene regulatory network based on an adaptive layer GCN-LSTM include:

[0011] S1. Use gene expression matrices at different time points;

[0012] S2. Gene features at different time points are extracted from the gene expression matrices at different time points and the prior gene regulatory network in an adaptive-layer GCN.

[0013] S3. Input gene features from different time points into LSTM to obtain gene features at the next time point;

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

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

[0016] S21. Initialize the number of layers in the GCN model to the current number of layers in the GCN model, and initialize the optimal AUROC.

[0017] S22. Train the GCN model for the current layer and calculate the AUROC, then obtain the AUROC for the current layer.

[0018] S23. Determine whether the current AUROC is better than the best AUROC;

[0019] S24. If the current AUROC is better than the best AUROC, then update the best AUROC and the current layer number, and repeat steps S22-S24.

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

[0021] Furthermore, in step S3, gene features from different time points are input into an LSTM to obtain gene features for the next time point. The process includes:

[0022] Genetic characteristics at different time points The data is fed into the LSTM model. For a given time step t, the LSTM unit uses the current input vector E. t and the state vector h from the previous time step t-1 t-1 As input, the state vector h at the current time step is then output. t The state update formula is:

[0023] (1)

[0024] (2)

[0025] (3)

[0026] (4)

[0027] (5)

[0028] (6)

[0029] Among them, i t For the input gate, f t For the Gate of Oblivion, o t For output gate, s t Let σ(·) be the storage unit, σ(·) be the sigmoid activation function, and ⊙ be the element-wise multiplication. These are the parameters for the corresponding unit;

[0030] The last hidden state h t Input to a fully connected layer to generate gene features E for the next time point. t+1 .

[0031] Further, S4 includes:

[0032] S41. Encode the gene features at the next time point and the prior gene regulatory network using a gravity-inspired encoder to form gene latent parameter features.

[0033] S42, Potential for genes A gravity-inspired decoder is used to decode and obtain directed gene regulatory networks.

[0034] Furthermore, the encoded gene potential parameter features The formula is expressed as

[0035] (7)

[0036] Where Z is the potential vector, with n*d dimensions; For quality parameters, n*1 dimensional; , is the out-degree normalization of the prior gene regulatory network A, and I is the out-degree normalization variable.

[0037] Furthermore, the decoding obtains a directed gene regulatory network, expressed by the formula as follows:

[0038] (8)

[0039] Among them, a i-j Let r be the acceleration value from node i to node j. 2 m is the squared distance between node i and node j. j Let be the quality of node j.

[0040] Second aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0041] 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.

[0042] The beneficial effects of this invention are:

[0043] 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 vary depending on the dataset. Using a GCN with a fixed number of layers will prevent the model from achieving the best accuracy. This invention uses an adaptive GCN with different number of layers, which can adapt the number of GCN layers according to the size of the dataset.

[0044] 2. This invention utilizes the richer gene features learned between LSTM time points to infer a more accurate gene expression matrix for the next time point.

[0045] 3. This invention employs a gravity-inspired graph autoencoder (GAE) to predict directed gene regulatory networks (GRNs). The encoder uses a normalized out-degree version of the prior gene regulatory network (GRN), ensuring the asymmetry of the learned graph. The decoder uses a formula inspired by gravitational acceleration, rather than a simple dot product, ensuring that the decoded graph is directed.

[0046] 4. Compared with existing methods, the method of the present invention has higher accuracy in cell type-specific GRN inference.

[0047] 5. The method of this invention considers the topological structure of genes and does not require conversion to a three-dimensional histogram, saving time. This method utilizes prior GRN and the gene expression matrix at time point t+1, while previous methods such as TDL and dynDeepDRIM convert the data into a three-dimensional histogram before inputting it into the model. Compared to their approach, the model of this invention does not require conversion and runs faster. Attached Figure Description

[0048] Figure 1Schematic diagram of single-cell static RNA sequencing data and single-cell temporal RNA sequencing data;

[0049] Figure 2 This is a flowchart illustrating the principle of the control method of the present invention;

[0050] Figure 3 This is a schematic flowchart of the control method of the present invention;

[0051] Figure 4 These are comparative experimental diagrams of different methods in embodiments of the present invention;

[0052] Figure 5 These are ablation experiment diagrams for different methods in embodiments of the present invention;

[0053] Figure 6 This is a schematic diagram of the structure of the electronic device of the present invention.

[0054] Explanation of English and Chinese noun correspondences:

[0055] Long Short-Term Memory (LSTM) network

[0056] Long Short-Term Memory (LSTM) cell

[0057] Graph Convolutional Networks (GCNs)

[0058] Adaptive Depth GCN

[0059] Gene expression

[0060] Gene regulatory networks (GRN)

[0061] Area under the curve (AUROC)

[0062] Prior gene regulatory network (prior GRN)

[0063] Inferred gene regulatory networks

[0064] Image from encoder (GAE)

[0065] Gravity-Inspired Encoder

[0066] Gravity-Inspired Decoder Detailed Implementation

[0067] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0068] Existing methods for converting gene expression data into histograms are time-consuming and introduce unexpected noise, which may cause the image representation to obscure some data features. Furthermore, existing methods do not fully consider the topology of prior gene regulatory networks, and only learn the features of first-order neighbors directly connected to genes, while failing to capture the features of second-order and higher-order neighbors indirectly connected to genes.

[0069] To address the above problems, this invention provides a GCN-LSTM reconstruction of a temporal gene regulatory network based on adaptive layer number. Figure 2 The flowchart illustrating the principle of GCN-LSTM reconstruction of temporal gene regulatory networks based on adaptive layer number is provided for embodiments of the present invention. Figure 3 For based on Figure 2 The flowchart, the method includes:

[0070] S1. Obtain the gene expression matrix at different time points of single-cell time-series RNA sequencing data;

[0071] The gene expression matrix X at different time points is expressed by the formula:

[0072]

[0073] Where l is the total step size and t is the current time.

[0074] The prior gene regulatory network is represented by matrix A.

[0075] S2. Input the gene expression matrix X at different time points and the prior gene regulation network A into an adaptive-layer GCN to extract the corresponding gene feature matrices at different time points. .

[0076] The GCN adaptive layer number operation logic is as follows:

[0077] S21. Initialize the number of layers in the GCN model to the current number of layers in the GCN model, and initialize the optimal AUROC.

[0078] S22. Train the GCN model for the current layer and calculate the AUROC, then obtain the AUROC for the current layer.

[0079] S23. Determine whether the current AUROC is better than the best AUROC;

[0080] S24. If the current AUROC is better than the best AUROC, then update the best AUROC and the current layer number, and repeat steps S22-S24.

[0081] S25. If the current AUROC is not better than the best AUROC, then end the loop.

[0082] The pseudocode for GCN adaptive layer number operation is as follows:

[0083] initialize:

[0084] initial_layers = 1 # The initial number of GCN layers is 1

[0085] current_layers = initial_layers # The current number of GCN layers is equal to the initial number of layers

[0086] best_AUROC = 0 # Initialize the best AUROC to 0

[0087] while TRUE:

[0088] train_and_evaluate_model(current_layers) # Train the GCN model with the current number of layers and compute AUROC

[0089] current_AUROC = get_AUROC_from_model() # Get the AUROC of the current layer from the model.

[0090] # Determine if the current AUROC is better than the best AUROC, and update the best AUROC and the current layer number.

[0091] if current_AUROC > best_AUROC:

[0092] best_AUROC = current_AUROC

[0093] current_layers += 1 # Increment the current layer number by 1 and continue the calculation

[0094] else:

[0095] break # If the current AUROC is no longer better than the best AUROC, then end the loop.

[0096] output best_AUROC, current_layers - 1 # Subtract 1 because the current layer number was added once in the previous loop.

[0097] The gene feature matrix at different time points is obtained through an adaptive GCN with varying layers, as expressed by the formula: .

[0098] S3, Gene Feature Matrix at Different Time Points Input LSTM to obtain gene features E at the next time point t+1 ;

[0099] Gene feature matrix at different time points The data is fed into an LSTM model, which has a powerful ability to learn long-term dependencies in sequence data to capture the evolution patterns of weighted dynamic networks.

[0100] like Figure 2 As shown, the LSTM model can be described as a wrapper unit with multiple multiplication gates; for a certain time step t, the LSTM unit uses the current input vector E t and the state vector h from the previous time step t-1 t-1 As input, the current time step state vector h is then output. t :

[0101] The LSTM state update formula is:

[0102] (1)

[0103] (2)

[0104] (3)

[0105] (4)

[0106] (5)

[0107] (6)

[0108] Among them, i t For the input gate, f t For the Gate of Oblivion, o t For output gate, s t Let σ(·) be the storage unit, σ(·) be the sigmoid activation function, and ⊙ be the element-wise multiplication. These are the parameters for the corresponding unit;

[0109] Finally, the last hidden state h t Input to a fully connected layer to generate prediction results, E, the gene features at the next time point. t+1 .

[0110] The adaptive-layer GCN-LSTM model is formed by combining GCN with LSTM and adaptive-layer GCN.

[0111] S4, Genetic characteristics based on the next time point E t+1 The directed gene regulatory network (GRN) at the next time point can be predicted from the prior gene regulatory network (GRN) in a gravity-inspired GAE.

[0112] Traditional GAE encoders use a 2-layer GCN, which yields the latent vector representation of the nodes.

[0113]

[0114] Where the GCN has L layers; the weight matrix is ​​; the eigenvector is ; the original adjacency matrix is ​​A; and the symmetric normalization of A is . , , where the angle matrix 𝐷 of 𝐴+𝐼 is obtained because Because it is symmetrical, the traditional GAE cannot learn the directional information of the directed graph in the encoder part. The latent vector Z of the node is obtained after passing through 2 layers of GCN encoder.

[0115] In the decoder of a traditional GAE, the decoder part is a simple dot product of the latent vectors of node i and node j, which is then subjected to a sigmoid function for binary classification. The result of the dot product between node i and node j is mapped to the range of 0-1. Then, according to 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.

[0116] In traditional GAE, Z is a latent vector, because Therefore, the decoder cannot obtain the directionality between node i and node j; it can only determine whether node i and node j are connected.

[0117] Gravity-inspired GAE improves upon traditional GAE by incorporating the gravitational forces between celestial bodies, enabling it to perform directed link predictions.

[0118] The principle is as follows: Under the influence of universal gravitation, the gravitational force between the Earth and the Moon is the same, as shown in the formula for universal gravitation:

[0119]

[0120] If we consider the acceleration between the Earth and the Moon, the formula for the Earth's acceleration towards the Moon is:

[0121]

[0122] For the Moon, the formula for the acceleration of the Moon pointing towards the Earth is:

[0123]

[0124] Where F is the gravitational force; m1 is the mass of the Earth; m2 is the mass of the Moon; G is the gravitational constant; and r is the distance between the two objects.

[0125] Because m1 > m2, therefore The probability of Earth pointing towards the Moon is less than the probability of the Moon pointing towards Earth. When we abstract the Moon and Earth as nodes i and j, we find that when using the universal gravitation formula, the calculated gravitational values ​​for nodes i and j are the same; however, when using the acceleration formula, the calculated acceleration values ​​for nodes i and j are different. This means that we can use this formula to establish a directed relationship between nodes i and j. The formula for calculating the relationship between nodes i and j, abstracted from the acceleration formula, is as follows:

[0126]

[0127] In this context, because the larger the molecule, the more likely node i and node j are to be connected; therefore, in a gene regulatory network, the molecule is more likely to capture more influential genes. The molecule has a mass, which corresponds to the mass parameter M (mass) in the denominator of the gene regulatory network. Since the denominator is the square of the distance r, it is abstracted as... They believe that nodes that are structurally similar are more likely to be connected.

[0128] Based on traditional GAE, the influence of gravity-inspired effects are considered, such as Figure 2 As shown, the process includes:

[0129] 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 latent parameter vector.

[0130] S42, Gene parameter vector Gravity-inspired decoders are used to decode and obtain directed gene regulatory networks (GRNs).

[0131] In the gravity-inspired encoder part, the input prior gene regulatory network A and the gene expression matrix X (E) are used to input the encoder. t+1 After two layers of GCN, the output is d+1 dimensions, where the first d dimensions represent the latent vector Z, and the last dimension represents the mass parameter M (mass). The formula for the output result is:

[0132] (7)

[0133] Where Z is the potential vector, with n*d dimensions; The mass parameter vector is n*1 dimensional; , is the out-degree normalized version of the prior gene regulatory network A, where I is the out-degree normalization variable. It focuses on the out-degree of the A matrix and is also the reason why gravity-inspired encoders can encode directed graphs.

[0134] In the gravity-inspired decoder's decoding section, the universal gravitation formula is applied to the decoder, resulting in a difference between the final gene pair (i, j) and the gene pair (j, i), thus revealing the directionality of the gene regulatory network. The formula is expressed as:

[0135] (8)

[0136] Among them, a i-j Let r be the acceleration value from node i to node j. 2 m is the squared distance between node i and node j. j Let be the quality of node j.

[0137] Examples of experiments applying this method:

[0138] First, a dataset is established. This embodiment uses four real datasets and a simulated time-series single-cell RNA sequencing dataset, as shown in Table 1:

[0139] Table 1 Time-series single-cell RNA sequencing dataset

[0140] mESC1 3456 23481 9 GSE79578 mESC2 2717 24175 4 GSE65525 hESC1 1529 26178 5 E-MTAB-3929 hESC2 758 19189 6 GSE75748 Sim1 20000 10000 10 Simulation Sim2 20000 10000 8 Simulation Sim3 20000 10000 6 Simulation Sim4 20000 10000 4 Simulation

[0141] Table 1 shows the benchmark tests of the GCN-LSTM model. The real datasets include two datasets (mESC1 and mESC2) from mouse embryonic stem cells and two datasets (hESC1 and hESC2) from human embryonic stem cells.

[0142] The simulated dataset was generated by dyngen and includes 10,000 genes and 20,000 cells.

[0143] During benchmarking, the selection of positive and negative gene pairs is balanced, with a training set:test set ratio of 4:1.

[0144] Comparative experiment:

[0145] The performance of the adaptive layer GCN-LSTM was evaluated and compared with several unsupervised and supervised learning methods. When inferring gene regulatory networks, supervised learning methods utilize known intergeneric relationships and gene expression profiles to train an optimal model, while unsupervised methods rely solely on gene expression profiles to find relationships between data and features to infer the relationships between transcription factors and genes.

[0146] 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 the regulatory relationships between new gene pairs.

[0147] This invention selects two unsupervised learning methods (PCC and MI) and five supervised learning methods (TDL-3DCNN, TDL-LSTM, dynDeepDRIM, scTGRN and GMFGRN) for comparison.

[0148] We used an adaptive-layer GCN-LSTM to infer gene regulatory networks on real experimental data, and the experimental results are as follows: Figure 5 The experimental metric is AUROC, which is achieved through... Figure 5 It can be seen that the model of this invention outperforms other comparative models on four different datasets.

[0149] Specifically, on the mESC2 dataset, the model of this invention (0.843) improved accuracy by 7.6% compared to the best-performing scTGRN (0.767); on the hESC1 dataset, the model of this invention (0.821) improved accuracy by 2.3% compared to the best-performing GMFGRN (0.798); on the hESC2 dataset, the model of this invention (0.884) improved accuracy by 10.0% compared to the best-performing GMFGRN (0.784); and on the mESC1 dataset, the model of this invention (0.832) improved accuracy by 10.8% compared to the best-performing GMFGRN (0.724). Overall, the improvements were significant on the mESC1 and hESC2 datasets.

[0150] like Figure 6As shown, the model demonstrates a significant improvement over other comparative models on simulated datasets, specifically: a 15.8% improvement on the Sim1 dataset; a 7.3% improvement on the Sim2 dataset; a 5.7% improvement on the Sim3 dataset; and a 2.6% improvement on the Sim4 dataset. These results indicate that the model also outperforms other comparative models on simulated datasets, particularly showing a substantial improvement on the Sim1 dataset. This demonstrates that the model of this invention maintains strong performance when processing different types of data, showcasing its potential application value and effectiveness.

[0151] Ablation experiment:

[0152] To verify the contribution of the proposed module to the model performance, ablation experiments were conducted. The number of GCN layers was fixed at 2 for 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 a gravity-inspired GAE to directly infer the gene regulatory network as a dynamic dataset for ablation.

[0153] Table 2 shows the ablation experiment performance results under different configurations. The data in parentheses represent the number of GCN layers with the best results determined by the final model.

[0154] Table 2 Ablation Experiment Results

[0155] static data 0.7809 0.8313 0.7863 0.8806 0.8183 0.8327 0.8677 0.8828 Fixed number of GCN layers 0.7984 0.81702 0.8143 0.7944 0.8304 0.8444 0.8203 0.8504 OURS 0.8211（1） 0.8434（1） 0.8326（1） 0.8839（1） 0.8883（3） 0.8938（3） 0.8973（3） 0.9161（3）

[0156] Experimental conclusion:

[0157] Results on static data show that while the traditional GAE method achieves good performance, there is still room for improvement compared to the model of this invention. Experiments with a fixed number of GCN layers show that model performance is affected when the number of layers is limited, especially when dealing with complex data; the model of this invention demonstrates significant performance improvements on all datasets, especially reaching a maximum of 0.9161 on Sim4, indicating that the model of this invention has better adaptability and accuracy in gene regulatory network inference tasks.

[0158] This invention also provides a GCN-LSTM reconstruction temporal gene regulation device based on adaptive layer number, the device comprising:

[0159] The generation module includes a GCN-LSTM model. The generation module uses gene expression matrices at different time points and extracts corresponding gene features at different time points from the gene expression matrices at different time points and the prior gene regulation network in an adaptive-layer GCN. The gene features at different time points are then input into an LSTM to obtain gene features at the next time point.

[0160] The prediction module includes a gravity-inspired GAE model, and the generation module predicts the directed gene regulatory network for the next time point in the gravity-inspired GAE based on the gene characteristics of the next time point and the prior gene regulatory network.

[0161] The present invention also provides an electronic device, Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 6 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0162] S1. Use gene expression matrices at different time points;

[0163] S2. Gene features at different time points are extracted from the gene expression matrices at different time points and the prior gene regulatory network in an adaptive-layer GCN.

[0164] S3. Input gene features from different time points into LSTM to obtain gene features at the next time point;

[0165] S4. Based on the gene characteristics at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in gravity-inspired GAE.

[0166] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0168] S1. Use gene expression matrices at different time points;

[0169] S2. Gene features at different time points are extracted from the gene expression matrices at different time points and the prior gene regulatory network in an adaptive-layer GCN.

[0170] S3. Input gene features from different time points into LSTM to obtain gene features at the next time point;

[0171] S4. Based on the gene characteristics at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in gravity-inspired GAE.

[0172] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0173] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reconstructing temporal gene regulatory networks based on adaptive layer GCN-LSTM, characterized by the following steps: include: S1. Obtain the gene expression matrix at different time points of single-cell time-series RNA sequencing data; The gene expression matrix X at different time points is expressed by the formula: Where l is the total step size and t is the current time; The prior gene regulatory network is represented by matrix A; S2. Gene features at different time points are extracted from the gene expression matrices at different time points and the prior gene regulatory network in an adaptive-layer GCN. S3. Input gene features from different time points into LSTM to obtain gene features at the next time point; S4. Based on the gene characteristics at the next time point and the prior gene regulatory network, the directed gene regulatory network at the next time point is predicted in gravity-inspired GAE.

2. The method for regulating a network according to claim 1, characterized in that, The GCN adaptive layer number operation logic is as follows: S21. Initialize the number of layers in the GCN model to the current number of layers in the GCN model, and initialize the optimal AUROC. S22. Train the GCN model for the current layer and calculate the AUROC, then obtain the AUROC for the current layer. S23. Determine whether the current AUROC is better than the best AUROC; S24. If the current AUROC is better than the best AUROC, then update the best AUROC and the current layer number, and repeat steps S22-S24. S25. If the current AUROC is not better than the best AUROC, then end the loop.

3. The method for regulating a network according to claim 1, characterized in that, The process of inputting gene features from different time points in S3 into an LSTM to obtain gene features for the next time point includes: Genetic characteristics at different time points The data is fed into the LSTM model. For a given time step t, the LSTM unit uses the current input vector E. t and the state vector h from the previous time step t-1 t-1 As input, the state vector h at the current time step is then output. t The state update formula is: (1) (2) (3) (4) (5) (6) Among them, i t For the input gate, f t For the Gate of Oblivion, o t For output gate, s t Let σ(·) be the storage unit, σ(·) be the sigmoid activation function, and ⊙ be the element-wise multiplication. These are the parameters for the corresponding unit; The last hidden state h t Input to a fully connected layer to generate gene features E for the next time point. t+1 .

4. The method for regulating a network according to claim 3, characterized in that, 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 gene latent parameter features. S42, Potential for genes A gravity-inspired decoder is used to decode and obtain directed gene regulatory networks.

5. The method for regulating a network according to claim 4, characterized in that, The encoded gene potential parameter features The formula is expressed as (7) Where Z is the potential vector, with n*d dimensions; For quality parameters, n*1 dimensional; I is the out-degree normalization of the prior gene regulatory network A, and I is the out-degree normalization variable.

6. The method for controlling a network according to claim 4, characterized in that, The decoding process obtains a directed gene regulatory network, expressed by the formula: (8) Among them, a i-j Let r be the acceleration value from node i to node j. 2 m is the squared distance between node i and node j. j Let be the quality 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, it implements the steps of the control network method as described in any one of claims 1 to 6.

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, it implements the steps of the control network method as described in any one of claims 1 to 6.

Images