A method of inferring a gene regulatory network

Abstract

The application discloses a method for inferring a gene regulation network, comprising the following steps: S1, constructing an input data set, comprising obtaining data and preprocessing the data, and finally generating a 4D tensor required by a training model input; S2, constructing an attention time convolution network, comprising a hollow causal convolution, an attention residual structure, an activation function, normalization, regularization, and dropout; inputting the 4D input tensor into the attention time convolution network; S3, constructing a gene regulation network and realizing gene function allocation. The application can be used for constructing a gene regulation network and realizing gene function allocation at the same time, improving the performance of the model when inferring the regulation relationship between genes, and in addition, the attention time convolution network can simultaneously learn the time characteristic information and the spatial characteristic information of the time course scRNA-seq data, thereby avoiding focusing on one kind of information and causing the loss of data information.

Description

Technical Field

[0001] This invention relates to the field of constructing single-cell gene regulatory networks, and in particular to a method for inferring gene regulatory networks. Background Technology

[0002] Gene regulatory networks consist of interactions between regulators that influence biological processes in organisms. These networks define and maintain cell-type-specific transcriptional states, which are fundamental to cell morphology and function. Gene regulatory networks control intracellular gene expression levels and activity through the binding of transcription factors to target genes, thus playing a role in various biological processes. However, biological processes remain highly dynamic over time, responding to changes in the environment and stimuli. Therefore, the binding of transcription factors to target genes also exhibits spatiotemporal dynamics. Researchers reconstruct gene regulatory networks by analyzing gene expression levels at different time points and modeling them based on time-series gene expression data, allowing them to infer causal interactions between genes. Furthermore, gene regulatory networks can explain how cells function at the genomic level and reveal the fundamental laws behind various life phenomena, which is of great importance to biological fields such as drug development and epidemiological research.

[0003] Currently, commonly used gene expression profiles mainly come from two mainstream technologies: bulk RNA sequencing (bulk RNA-seq) and single-cell RNA sequencing (scRNA-seq). Bulk RNA-seq yields average transcriptome data from all cells in the same sample, making it relatively easy to obtain. However, the specific information of individual cells is often masked in these samples. Due to the characteristics of bulk RNA-seq data, modeling the time-dependent changes in bulk gene expression profiles can provide valuable information for reconstructing gene regulatory networks and understanding dynamic biological processes. Several computational methods have been proposed to reconstruct gene regulatory networks from time-series bulk gene expression data. These methods can infer causal relationships between genes from gene expression relationships; however, bulk gene expression data only represents average expression levels and inappropriately assumes homogeneity of gene expression across cells, thus ignoring cellular heterogeneity.

[0004] As researchers delve deeper into the structure and function of organisms, it becomes increasingly clear that there are significant differences in transcriptome data between individual cells. scRNA-seq, a popular technique, captures the gene expression profile of individual cells, allowing for the reconstruction of cell-type-specific gene regulatory networks by taking cellular heterogeneity into account. scRNA-seq gene expression data can be categorized into time-series scRNA-seq and static scRNA-seq data; the former provides more information about gene interactions than the latter. Most existing methods use static scRNA-seq gene expression data to reconstruct gene regulatory networks, and these methods cannot be directly applied to time-series scRNA-seq gene expression data. A major challenge with static scRNA-seq gene expression data is the inability to track cells analyzed at specific time points, making it unclear which cell at the next time point is closely related to a specific cell at the previous time point, thus hindering the determination of the precise trajectory of genes. To address these issues, some methods utilize pseudo-time to reconstruct gene regulatory networks from scRNA-seq data, but these methods cannot be directly applied to time-series scRNA-seq gene expression data. Therefore, static scRNA-seq gene expression data are mostly used for interaction tasks, and it is difficult to infer causal relationships between genes.

[0005] Models developed using time-series scRNA-seq gene expression data can effectively address data-related issues and demonstrate superior performance, but these methods still have certain limitations. First, due to the large number of genes per cell in scRNA-seq gene expression data, the models are significantly affected by noise. Second, when predicting causal relationships between gene pairs, these methods focus only on extracting temporal or spatial features from the data, leading to the loss of data information. Summary of the Invention

[0006] To address the existing problems, this invention provides a method for inferring gene regulatory networks, the specific scheme of which is as follows:

[0007] A method for inferring gene regulatory networks includes the following steps:

[0008] S1, Construct the input dataset, including acquiring and preprocessing the data, and finally generating the 4D tensor required for training the model input;

[0009] S2, construct an attention-temporal convolutional network, including dilated causal convolution, attention residual structure, activation function, normalization, regularization, and dropout; input the 4D input tensor into the attention-temporal convolutional network;

[0010] S3, constructing gene regulatory networks and realizing gene function allocation.

[0011] Preferably, step S1 specifically includes the following steps:

[0012] S11. Obtain both real-time scRNA-seq and simulated-time scRNA-seq datasets. To define positive and negative transcription factor-gene pairs, use the ChIP-seq data from the datasets as a benchmark to infer their potential targets. Gene pairs with one or more significant peaks in the promoter region of the target gene are designated as positive transcription factor-gene pairs, and the rest as negative transcription factor-gene pairs. Furthermore, the number of positive and negative gene pairs is balanced when constructing the input dataset.

[0013] S12 is a specialized standardization of the gene expression matrix containing T time points; where the rows of the gene expression matrix represent genes and the columns represent cells.

[0014] S13 generates a 4D input tensor for each gene pair (a, b) at each time point;

[0015] Preferably, step S13 includes the following steps:

[0016] S131, Calculate a new gene “AVG” to represent the average expression of all genes in each cell;

[0017] S132, reconstruct the genome (a, b, AVG) by combining the gene “AVG” and the gene pair (a, b);

[0018] S133, the expression values ​​of the genome (a, b, AVG) in all cell samples are divided into 8 equal parts, thereby constructing an 8×8×8 3D matrix for the genome (a, b, AVG), where each entry (i, j, k) represents the probability that (a, b, AVG) co-occurs at the i-th, j-th, and k-th expression levels;

[0019] S134, perform the above operation on the gene expression matrix at each time point, and finally generate the 4D tensor required for the training model input. The final dimension of the input tensor is T×8×8×8.

[0020] Preferably, step S2 specifically includes the following steps:

[0021] S21, the attention-temporal convolutional network consists of n attention residual modules, wherein each attention residual block contains two residual modules and one attention module; the residual block is composed of dilated causal convolution, weight normalization, ReLU activation function, and Dropout in sequence;

[0022] The attention module consists of global average pooling, a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid function in sequence.

[0023] The attention module works as follows: First, global average pooling is performed on the feature vector to reduce the spatial features to 1×1×1; then, a non-linear activation function and two fully connected layers are used to construct connections between channels; next, normalized weights are obtained through a sigmoid activation function, and finally, weights are multiplied and applied channel-by-channel to each channel of the original feature vector to complete the weight allocation for different feature channels; the function used is as follows:

[0024]

[0025]

[0026]

[0027] Where D, H, and W represent the dimensions of the feature vector, T1 and T2 represent the mapping process of the fully connected layer, ReLU represents the non-linear activation function, σ represents the calculation of the weight values ​​of each channel, and X represents the matrix corresponding to the original feature vector for each channel.

[0028] S22, for an input x = {x1, x2, ..., x...} containing T time points, T-1 ,x T}, through x T The output y at time T is calculated using data from previous times and the data from previous times. T ; where the output y is calculated. T When, x needs to be T The data from the previous time steps are input together into the attention-temporal convolutional network; the dilated causal convolution in the attention residual module can perform interval sampling of the input data during convolution, increasing the receptive field during convolution; the size of the receptive field is related to the dilation factor, the larger the dilation factor, the larger the receptive field; for the i-th convolutional layer, its dilation factor is d = 2. i ;

[0029] S23, containing an input x = {x1, x2, ..., x...} at T time points. T-1 ,x T After the features are extracted by two residual blocks, they are input into the attention module to assign weights to different feature channels. The above process is the processing of one attention residual module. The 4D input tensor formed after preprocessing needs to be iteratively processed by n attention residual modules in the attention temporal convolutional network.

[0030] The input tensor of S24, 4D is trained through an attention-temporal convolutional network, and a loss function is calculated simultaneously. Training stops when the number of training iterations reaches a set number or the loss error is less than a set threshold. The optimal model obtained from iterative training is then used for test data, following the same steps. The 4D input tensor from the attention-temporal convolutional network is converted into compressed embeddings and integrated through a fully connected layer. Finally, a sigmoid function is used to generate the final binary classification prediction score, and the prediction score is used to determine whether there is a causal regulatory relationship between gene pairs (a, b). The output of the attention-temporal convolutional network is y = {y1, y2, ..., y...}. T-1 y T}

[0031] Preferably, in step S3, a gene regulatory network is constructed based on the causal regulatory relationship between gene pairs (a, b) determined by the model. Gene pairs connected by directed edges indicate that they have a causal regulatory relationship, while gene pairs not connected by edges indicate that they have no regulatory relationship.

[0032] Preferably, the process of gene function allocation in step S3 is as follows: inputting the data of this type of functional gene set into the attention temporal convolutional network, it is possible to infer the gene pairs with regulatory relationships in this type of gene set. Gene pairs with regulatory relationships indicate that they have the same gene function and can be allocated into functional blocks of the same type of genes.

[0033] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed, performs the method described in any of the above-mentioned embodiments.

[0034] The present invention also discloses a computer system including a processor, a storage medium storing a computer program, and the processor reading from the storage medium and running the computer program to perform the method described in any of the preceding claims.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. This invention is a method specifically designed for inferring gene regulatory networks from single-cell time-series data. It is a deep learning model that effectively addresses the problems associated with using other data and exhibits superior performance compared to other models. This invention can be used simultaneously to construct gene regulatory networks and realize gene function allocation.

[0037] 2. This invention uses the average expression of all genes in each cell as a reference gene for judging the relationship between gene pairs. This can effectively avoid the noise problem caused by the large number of genes in time-series scRNA-seq gene expression data, thereby improving the performance of the model in inferring the regulatory relationship between gene pairs.

[0038] 3. This invention utilizes the joint expression of gene pairs to construct a 4D input tensor, that is, to represent the joint expression of each gene pair as a histogram image. This is beneficial to leveraging the analytical advantages of deep learning models to infer the relationship between the gene pairs encoded in the image.

[0039] 4. This invention specifically designs an attention-temporal convolutional network for time-series scRNA-seq data. The attention residual module in the attention-temporal convolutional network increases the receptive field of the convolution while allowing the output of each convolution to contain a wider range of information. In addition, the attention-temporal convolutional network can simultaneously learn the temporal and spatial features of time-series scRNA-seq data, avoiding the loss of data information due to focusing on only one type of information. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a model framework diagram of the present invention;

[0042] Figure 2 The detailed structure of the attention-temporal convolutional network is shown in the model framework diagram of this invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] A method for inferring gene regulatory networks includes the following steps:

[0045] S1, Construct the input dataset, including acquiring and preprocessing the data, and finally generating the 4D tensor required for training the model input.

[0046] Specifically, step S1 includes the following steps:

[0047] S11: Obtain both real-time scRNA-seq and simulated-time scRNA-seq datasets. To define positive and negative transcription factor-gene pairs, use the ChIP-seq data from the datasets as a benchmark to infer their potential targets. Gene pairs with one or more significant peaks in the promoter region of the target gene are designated as positive transcription factor-gene pairs, and the rest as negative transcription factor-gene pairs. Furthermore, the number of positive and negative gene pairs is balanced when constructing the input dataset.

[0048] Specific implementation: Two real datasets from mouse embryonic stem cells and two from human embryonic stem cells were used. One mouse embryonic stem cell dataset contained 38 transcription factors, and the other three datasets contained 36 transcription factors. The number of time points in these datasets were 4, 5, 6, and 9, respectively. The four simulated datasets used each contained 100 transcription factors, 10,000 genes, and 20,000 cells, with the number of time points being 4, 6, 8, and 10, respectively. In addition, a mouse cerebral cortex scRNA-seq dataset containing three time points was used for gene function assignment. This dataset contained four baseline gene sets.

[0049] S12, perform specialized standardization on the gene expression matrix containing T time points; where the rows of the gene expression matrix represent genes and the columns represent cells. In the above embodiment, all time-series scRNA-seq datasets are standardized to obtain a normalized gene expression matrix.

[0050] S13 generates a 4D input tensor for each gene pair (a, b) at each time point.

[0051] Specifically, step S13 includes:

[0052] S131, calculate a new gene “AVG” to represent the average expression of all genes in each cell.

[0053] S132, reconstruct the genome (a, b, AVG) by combining the gene “AVG” and the gene pair (a, b).

[0054] S133, the expression values ​​of the genome (a, b, AVG) in all cell samples are divided into 8 equal parts, thereby constructing an 8×8×8 3D matrix for the genome (a, b, AVG), where each entry (i, j, k) represents the probability that (a, b, AVG) co-occurs at the i-th, j-th, and k-th expression levels.

[0055] S134, the above operations are performed on the gene expression matrix at each time point, finally generating the 4D tensor required for the training model input. The final dimension of the input tensor is T×8×8×8. In this embodiment, the optimal gene pair image resolution used is 8×8×8.

[0056] S2, construct an attention-temporal convolutional network, including dilated causal convolution, attention residual structure, activation function, normalization, regularization, and dropout; detailed network structure is as follows. Figure 2 As shown, the 4D input tensor is fed into the attention-based temporal convolutional network.

[0057] Specifically, step S2 includes the following steps:

[0058] S21, the attention-temporal convolutional network consists of n attention residual modules, where each attention residual module contains two residual modules and one attention module. In this embodiment, the number of attention residual modules is 10, the optimizer used is a stochastic gradient descent optimizer, the learning rate during training is continuously adjusted using a cosine annealing strategy, the adjustment range is 0.01-0.00001, the cross-entropy loss function used is BCEWithLogitsLoss, the batch size is 256, and the maximum period is set to 200.

[0059] The residual block is composed of dilated causal convolution, weight normalization, ReLU activation function, and Dropout in sequence.

[0060] The attention module consists of global average pooling, a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid function.

[0061] The attention module works as follows: First, global average pooling is performed on the feature vector to reduce the spatial features to 1×1×1; then, a non-linear activation function and two fully connected layers are used to construct connections between channels; next, normalized weights are obtained through a sigmoid activation function, and finally, weights are multiplied and applied channel-by-channel to each channel of the original feature vector to complete the weight allocation for different feature channels; the function used is as follows:

[0062]

[0063]

[0064]

[0065] Where D, H, and W represent the dimensions of the feature vector, T1 and T2 represent the mapping process of the fully connected layer, ReLU represents the non-linear activation function, σ represents the calculation of the weight values ​​of each channel, and X represents the matrix corresponding to the original feature vector for each channel.

[0066] S22, for an input x = {x1, x2, ..., x...} containing T time points, T-1 x T}, through x T The output y at time T is calculated using data from previous times and the data from previous times. T ; where the output y is calculated. T When, x needs to be T The data from the previous time steps are input into the attention-temporal convolutional network; the dilated causal convolution in the attention residual module can sample the input data at intervals during convolution, increasing the receptive field during convolution and allowing the output of each convolution to contain a wider range of information; in this embodiment, the range of T is 3-10.

[0067] The size of the receptive field is related to the hole factor; the larger the hole factor, the larger the receptive field. For the i-th convolutional layer, the hole factor is d = 2. i ;

[0068] S23, containing an input x = {x1, x2, ..., x...} at T time points. T-1 x T After the features are extracted by two residual blocks, they are input into the attention module to allocate weights for different feature channels. The above process is the processing of one attention residual module. The 4D input tensor formed after preprocessing needs to be iteratively processed by n attention residual modules in the attention temporal convolutional network. In this embodiment, the feature vector is reduced to 1×1×1 after passing through the attention module.

[0069] The input tensor of S24, 4D is trained through an attention-temporal convolutional network, and the loss function is calculated simultaneously. Training stops when the number of training iterations reaches a set number or the loss error is less than a set threshold. The optimal model obtained from iterative training is used for test data, and the steps are the same as above. The input tensor of 4D through the attention-temporal convolutional network is converted into a compressed embedding and integrated through a fully connected layer. Finally, the final prediction score for binary classification is generated through the sigmoid function, and the causal regulatory relationship between gene pairs (a, b) is determined based on the prediction score.

[0070] In this embodiment, to ensure the randomness of the partitioning results, random partitioning is adopted, with a training set to test set ratio of 8:2. Furthermore, to determine whether the model training has been optimized to the optimal level, this example uses TruePositive Rate (TPR) and False Positive Rate (FPR) to plot the Receiver Operating Characteristic curve (ROC curve) and calculates the area under the ROC curve (AUROC). Simultaneously, based on the Precision-Recall curve (PR curve) plotted using precision and recall, the area under the PR curve (AUPR) is calculated to determine the performance of the optimal model in inferring regulatory relationships between genes.

[0071] The output of the attention-temporal convolutional network is y = {y1, y2, ..., y3}. T-1 y T}

[0072] S3, constructing gene regulatory networks and realizing gene function allocation.

[0073] In step S3, a gene regulatory network is constructed based on the causal regulatory relationship between gene pairs (a, b) determined by the model. Gene pairs connected by directed edges indicate that they have a causal regulatory relationship, while gene pairs without connected edges indicate that they do not have a regulatory relationship. In this embodiment, a directed graph representing the gene regulatory network is drawn based on the regulatory relationship between gene pairs.

[0074] The process of gene function allocation in step S3 is as follows: The data of this type of functional gene set is input into an attention-based temporal convolutional network. Gene pairs with regulatory relationships within this set can be inferred. Gene pairs with regulatory relationships indicate that they have the same gene function and can be assigned to functional blocks of the same type of gene. In this embodiment, the dataset used for gene function allocation is a baseline gene set representing four gene functions: cognition, perception, synaptic signaling, and axonal function.

[0075] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed, performs the method described in any of the above-mentioned embodiments.

[0076] The present invention also discloses a computer system including a processor, a storage medium storing a computer program, and the processor reading from the storage medium and running the computer program to perform the method described in any of the preceding claims.

[0077] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in a generalized manner in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described functionality in different ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the invention.

[0078] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0079] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0080] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0081] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0082] 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.

[0083] Note: To more clearly illustrate the technical solution of this application, the applicant has submitted color illustrations of the specification along with this application document for reference.

Claims

1. A method for inferring gene regulatory networks, characterized in that, Includes the following steps: S1, Construct the input dataset, including acquiring and preprocessing the data, and finally generating the 4D tensor required for training the model input; S2, construct an attention-temporal convolutional network, including dilated causal convolution, attention residual structure, activation function, normalization, regularization, and dropout; input the 4D input tensor into the attention-temporal convolutional network; The attention-temporal convolutional network consists of n attention residual modules, where each attention residual module contains two residual blocks and one attention module; the residual blocks are composed of dilated causal convolution, weight normalization, ReLU activation function, and Dropout in sequence; The attention module consists of global average pooling, a fully connected layer, a ReLU activation function, a fully connected layer, and a Sigmoid function in sequence. S3, constructing gene regulatory networks and realizing gene function allocation; Step S1 specifically includes the following steps: S11. Obtain the real time-series scRNA-seq dataset and the simulated time-series scRNA-seq dataset respectively. To define positive and negative transcription factor-gene pairs, use the ChIP-seq data of the dataset as a benchmark to infer their potential targets. Gene pairs with one or more significant peaks in the promoter region of the target gene are regarded as positive transcription factor-gene pairs, and the rest are regarded as negative transcription factor-gene pairs. In addition, the number of positive and negative gene pairs selected is balanced when constructing the input dataset. S12 is a specialized standardization of the gene expression matrix containing T time points; where the rows of the gene expression matrix represent genes and the columns represent cells. S13 generates a 4D input tensor for each gene pair (a, b) at each time point; The specific steps of step S13 include: S131, Calculate a new gene "AVG" to represent the average expression of all genes in each cell; S132, Reconstruct the genome (a, b, AVG) by combining the gene "AVG" and the gene pair (a, b). S133, the expression values ​​of the genome (a, b, AVG) in all cell samples are divided into 8 equal parts, thereby constructing an 8×8×8 3D matrix for the genome (a, b, AVG), where each entry (i, j, k) represents the probability that (a, b, AVG) co-occurs at the i-th, j-th, and k-th expression levels; S134, perform the above operation on the gene expression matrix at each time point, and finally generate the 4D tensor required for the training model input. The final dimension of the input tensor is T×8×8×8.

2. The method according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21, the attention module works as follows: First, global average pooling is performed on the feature vector to reduce the spatial features to 1×1×1; then, a non-linear activation function and two fully connected layers are used to build connections between channels; then, normalized weights are obtained through the Sigmoid activation function, and finally, the weights are multiplied and applied to each channel of the original feature vector to complete the weight allocation for different feature channels. S22, for input containing T time points ,pass The output at time T is calculated using data from previous times and the data from previous times. Among them, the calculated output At that time, it is necessary to The data from the previous time steps are input together into the attention-temporal convolutional network; the dilated causal convolution in the attention residual module performs interval sampling of the input data during convolution, increasing the receptive field during convolution; the size of the receptive field is related to the dilation factor, the larger the dilation factor, the larger the receptive field; for the Lth convolutional layer, its dilation factor is d = ; S23, containing input at T time points After the features are extracted by two residual blocks, they are input into the attention module to assign weights to different feature channels. The above process is the processing of one attention residual module. The 4D input tensor formed after preprocessing needs to be iteratively processed by n attention residual modules in the attention temporal convolutional network. S24, the 4D input tensor is trained through an attention-temporal convolutional network, while simultaneously calculating the loss function. Training stops when the number of training iterations reaches a set number or the loss error is less than a set threshold. The optimal model obtained from iterative training is used for test data. The 4D input tensor from the attention-temporal convolutional network is converted into compressed embeddings and integrated through a fully connected layer. Finally, the final binary classification prediction score is generated using the sigmoid function, and the causal regulatory relationship between gene pairs (a, b) is determined based on the prediction score. The output of the attention-temporal convolutional network is... .

3. The method according to claim 1, characterized in that: In step S3, a gene regulation network is constructed based on the causal regulatory relationship between gene pairs (a, b) determined by the model. Gene pairs connected by directed edges indicate that they have a causal regulatory relationship, while gene pairs not connected by edges indicate that they have no regulatory relationship.

4. The method according to claim 1, characterized in that, The process of gene function allocation in step S3 is as follows: input the functional gene set data into the attention temporal convolutional network, infer the gene pairs with regulatory relationships in the functional gene set, and the gene pairs with regulatory relationships represent the same gene function and are allocated into functional blocks of the same type of gene.

5. A computer-readable storage medium, characterized in that: The medium contains a computer program, which, when run, performs the method as described in any one of claims 1 to 4.

6. A computer system, characterized in that: It includes a processor and a storage medium, on which a computer program is stored, and the processor reads from the storage medium and runs the computer program to perform the method as described in any one of claims 1 to 4.

Images