Sample specific gene regulatory network prediction method, apparatus and device, and storage medium

Abstract

The invention provides a sample specific gene regulatory network prediction method, device and equipment and a storage medium, and relates to the field of artificial intelligence, the method comprises the following steps: obtaining dual expression data of regulatory factors and target genes corresponding to m samples; taking the dual expression data as input data, taking minimization of a reconstruction error of the input data as a target, and training a preset model to obtain a target model; obtaining an attention value matrix of the regulatory factors and the target genes under the single sample level corresponding to the m samples through the target model; and for each sample, based on the attention value matrix corresponding to the sample, determining a plurality of target regulatory factor-target gene regulatory relationships, and generating a sample specific gene regulatory network corresponding to the sample according to the plurality of target regulatory factor-target gene regulatory relationships, thereby realizing efficient prediction of the sample specific gene regulatory network.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method, apparatus, device, and storage medium for predicting sample-specific gene regulatory networks. Background Technology

[0002] A gene regulatory network (GRN) is a collection of interactions between multiple regulatory factors and multiple targets. Regulators can activate or inhibit targets. Therefore, this dependency between regulatory factors and targets makes GRNs a reliable model for studying biological phenomena.

[0003] Currently, the methods used to construct gene regulatory networks typically involve predicting gene regulatory networks at multiple sample levels.

[0004] However, since each biological sample (tissue or cell) has unique characteristics, the GRNs that affect the traits of these biological samples must also be specific. Therefore, GRNs at the multi-sample level are difficult to characterize GRNs at the single-sample level, and there is an urgent need for a method that can construct sample-specific (i.e., single-sample level) GRNs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, apparatus, device, and storage medium for predicting sample-specific gene regulatory networks, which can generate sample-specific gene regulatory networks for each sample based on the dual expression data of regulatory factors and target genes corresponding to m samples through a target model.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for predicting sample-specific gene regulatory networks, comprising: acquiring dual expression data of regulatory factors and target genes corresponding to m samples, where m is an integer greater than or equal to 1; using the dual expression data as input data, training a preset model with the goal of minimizing the reconstruction error of the input data to obtain a target model; obtaining the attention value matrix of regulatory factors and target genes at the single-sample level corresponding to each of the m samples through the target model; for each sample, determining multiple target regulatory factor-target gene regulatory relationships based on the attention value matrix corresponding to the sample, and generating a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulatory factor-target gene regulatory relationships.

[0008] In the first aspect, by inputting dual expression data of regulatory factors and target genes into a pre-set model for training, the goal is to minimize the reconstruction error of the input data, enabling the pre-set model to capture nonlinear regulatory relationships between genes. After training, the target model can output an attention value matrix reflecting the specific interaction strength between regulatory factors and target genes within a single sample. Based on this attention value matrix, multiple target regulatory factor-target gene regulatory relationships corresponding to each sample are screened, and a sample-specific gene regulatory network is constructed, achieving efficient prediction of sample-specific gene regulatory networks.

[0009] In some implementations, the preset model is a model based on a multi-head attention mechanism, which includes multiple attention heads, each corresponding to a parameter; the preset model is trained with dual-expression data as input data, with the goal of minimizing the reconstruction error of the input data, to obtain the target model, including: based on the dual-expression data, performing multiple rounds of iterative updates on the parameters corresponding to the multiple attention heads in the preset model to obtain the target model.

[0010] In some implementations, i is an integer greater than or equal to 1 and less than or equal to m, and for the i-th sample, each iteration update includes:

[0011] Obtain the embedding matrix and position vector corresponding to the i-th sample; using the embedding matrix and position vector corresponding to the i-th sample as input, obtain the sample attention value matrix corresponding to each of the multiple attention heads based on the multi-head attention mechanism; calculate the output attention value matrix corresponding to each of the multiple attention heads based on the sample attention value matrix corresponding to each of the multiple attention heads; reconstruct the sample based on the output attention value matrix corresponding to each of the multiple attention heads to obtain the predicted sample; update the parameters of the multiple attention heads using the difference between the predicted sample and the i-th sample as the loss.

[0012] In some implementations, the attention value matrix of the regulatory factor and the target gene at the single-sample level for each of the m samples is obtained through the target model, including: for the i-th sample: generating the attention value matrix of the i-th sample at each attention head through the target model; determining the attention value matrix of the regulatory factor and the target gene at the single-sample level for the i-th sample based on the average matrix of the multiple attention value matrices corresponding to the i-th sample.

[0013] In some implementations, the attention value matrix of the regulatory factor and target gene for the i-th sample at the single-sample level is determined based on the average matrix of multiple attention value matrices corresponding to the i-th sample, including: determining the attention value matrix of the regulatory factor and target gene for the i-th sample at the single-sample level according to the following formula:

[0014]

[0015] in, This represents the attention value matrix between the regulatory factor and the target gene for the i-th sample at the single-sample level. This represents the average of the multiple attention value matrices corresponding to the i-th sample. This represents the number of regulatory factors in the i-th sample. This represents the number of target genes in the i-th sample.

[0016] In some implementations, for each sample, based on the attention value matrix corresponding to the sample, multiple target regulator-target gene regulatory relationships are determined, and a sample-specific gene regulatory network corresponding to the sample is generated according to the multiple target regulator-target gene regulatory relationships, including: for each sample:

[0017] Calculate the standard score of the attention value matrix corresponding to the sample; obtain the statistical significance p-value corresponding to the sample based on the standard score, the statistical significance p-value corresponding to the sample includes the statistical significance p-values ​​corresponding to the regulatory relationships of multiple candidate regulators and target genes in the sample; generate the sample-specific gene regulation network corresponding to the sample based on the regulatory relationships of multiple target regulators and target genes in which the statistical significance p-values ​​meet the preset conditions.

[0018] In some implementations, the preset conditions include: the statistical significance p-value is less than a preset p-value correction threshold.

[0019] Secondly, the present invention provides a method for predicting sample-specific gene regulatory networks, comprising: an acquisition module for acquiring dual expression data of regulatory factors and target genes corresponding to m samples, where m is an integer greater than or equal to 1; a training module for training a preset model using the dual expression data as input data, with the goal of minimizing the reconstruction error of the input data, to obtain a target model; the acquisition module is further configured to acquire the attention value matrix of regulatory factors and target genes at the single-sample level corresponding to each of the m samples through the target model; and a generation module for each sample, which, based on the attention value matrix corresponding to the sample, determines multiple target regulatory factor-target gene regulatory relationships and generates a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulatory factor-target gene regulatory relationships.

[0020] In some implementations, the preset model is a multi-head attention mechanism model, which includes multiple attention heads, each corresponding to a parameter. The training module is specifically used to iteratively update the parameters corresponding to each of the multiple attention heads in the preset model multiple times based on dual-representation data, to obtain the target model.

[0021] In some implementations, i is an integer greater than or equal to 1 and less than or equal to m. For the i-th sample, in each iteration, the training module is specifically used to obtain the embedding matrix and position vector corresponding to the i-th sample; taking the embedding matrix and position vector corresponding to the i-th sample as input, and based on the multi-head attention mechanism, obtain the sample attention value matrix corresponding to each of the multiple attention heads; based on the sample attention value matrix corresponding to each of the multiple attention heads, calculate the output attention value matrix corresponding to each of the multiple attention heads; perform sample reconstruction based on the output attention value matrix corresponding to each of the multiple attention heads to obtain the predicted sample; and use the difference between the predicted sample and the i-th sample as the loss to update the parameters of the multiple attention heads.

[0022] In some implementations, the acquisition module is specifically used for the i-th sample to: generate the attention value matrix of the i-th sample at each attention head through the target model; and determine the attention value matrix of the regulatory factor and the target gene of the i-th sample at the single-sample level based on the average matrix of the multiple attention value matrices corresponding to the i-th sample.

[0023] In some implementations, the acquisition module is specifically used to determine the attention value matrix of the regulatory factor and target gene for the i-th sample at the single-sample level according to the following formula:

[0024]

[0025] in, This represents the attention value matrix between the regulatory factor and the target gene for the i-th sample at the single-sample level. This represents the average of the multiple attention value matrices corresponding to the i-th sample. This represents the number of regulatory factors in the i-th sample. This represents the number of target genes in the i-th sample.

[0026] In some implementations, the generation module is specifically used for each sample:

[0027] Calculate the standard score of the attention value matrix corresponding to the sample; obtain the statistical significance p-value corresponding to the sample based on the standard score, the statistical significance p-value corresponding to the sample includes the statistical significance p-values ​​corresponding to the regulatory relationships of multiple candidate regulators and target genes in the sample; generate the sample-specific gene regulation network corresponding to the sample based on the regulatory relationships of multiple target regulators and target genes in which the statistical significance p-values ​​meet the preset conditions.

[0028] In some implementations, the preset conditions include: the statistical significance p-value is less than a preset p-value correction threshold.

[0029] Thirdly, the present invention provides an electronic device including a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the method provided in the first aspect.

[0030] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to perform the method provided in the first aspect.

[0031] The beneficial effects of the second to fourth aspects can be referred to the first aspect, and will not be elaborated here. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This illustration shows a schematic diagram of the steps in a sample-specific gene regulatory network prediction method provided in an embodiment of this application.

[0034] Figure 2 A flowchart illustrating a sample-specific gene regulatory network prediction method provided in an embodiment of this application is shown.

[0035] Figure 3 A similarity diagram of the miRNA regulatory network in leukemia cells in Example 1 of this application is shown;

[0036] Figure 4 A similarity diagram of the TF regulatory network in a lymphoma sample from Embodiment 2 of this application is shown. Detailed Implementation

[0037] 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 some embodiments of the present invention, but not all embodiments.

[0038] A gene regulatory network (GRN) is a collection of multiple regulators interacting with multiple targets. It determines cell function and fate by regulating the expression levels of messenger RNA (mRNA) and proteins. Regulators can be DNA, RNA, and proteins, or any combination of these three molecules. Regulators can activate or inhibit targets. Therefore, this dependency between regulators and targets makes GRNs a reliable model for studying biological phenomena. Furthermore, GRN models help understand complex biological networks; for example, they allow for the identification of drugs for treating complex diseases. Inferring GRNs is one of the key issues in transcriptomics data. Constructing GRNs can help identify key factors influencing biological fate, such as phenotypic factors controlling cell differentiation and affecting disease development. Constructing GRNs can also help understand functional heterogeneity under different biological conditions and identify key genes driving these conditions.

[0039] Each biological sample (tissue or cell) possesses unique characteristics, therefore the GRNs influencing these biological sample traits must also be specific. Thus, constructing GRNs for each sample is a breakthrough in exploring sample heterogeneity and gene regulatory specificity. However, previous methods focused on identifying aggregated gene regulatory networks at multiple sample levels, neglecting the two key characteristics of sample heterogeneity and gene regulatory network specificity. Furthermore, existing sample-specific gene regulatory network prediction methods primarily employ linear interpolation and statistical perturbation strategies to identify linear regulatory relationships between regulatory factors and targets, often ignoring nonlinear regulatory relationships. With the widespread application of deep learning methods in transcriptome data, using deep learning algorithms to predict sample-specific gene regulatory networks holds great potential.

[0040] Current methods for predicting gene regulatory networks typically involve predicting networks at multiple sample levels. However, due to the heterogeneity of each biological sample, this inevitably prevents the construction of sample-specific gene regulatory networks. Furthermore, existing methods primarily employ linear interpolation and statistical perturbation strategies to identify linear regulatory relationships between regulatory factors and targets, often neglecting nonlinear regulatory relationships. Additionally, the high time complexity of existing sample-specific gene regulatory network prediction methods hinders efficient identification of such networks in large-scale datasets.

[0041] Based on this, embodiments of the present invention provide a method for predicting sample-specific gene regulatory networks, comprising: acquiring dual expression data of regulatory factors and target genes corresponding to m samples, where m is an integer greater than or equal to 1; using the dual expression data as input data, training a preset model with the goal of minimizing the reconstruction error of the input data to obtain a target model; obtaining the attention value matrix of regulatory factors and target genes at the single-sample level corresponding to each of the m samples through the target model; for each sample, determining multiple target regulatory factor-target gene regulatory relationships based on the attention value matrix corresponding to the sample, and generating a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulatory factor-target gene regulatory relationships.

[0042] By inputting dual expression data of regulatory factors and target genes into a pre-defined model for training, with the goal of minimizing the reconstruction error of the input data, the pre-defined model is able to capture the nonlinear regulatory relationships between genes. After training, the target model can output an attention value matrix reflecting the specific interaction strength between regulatory factors and target genes within a single sample. Based on this attention value matrix, multiple target regulatory factor-target gene regulatory relationships corresponding to each sample are screened, and a sample-specific gene regulatory network is constructed, achieving efficient prediction of sample-specific gene regulatory networks.

[0043] The sample-specific gene regulatory network prediction method provided by the embodiments of the present invention will be described below with reference to the accompanying drawings.

[0044] Figure 1 This illustration shows a schematic diagram of the steps in a sample-specific gene regulatory network prediction method provided in an embodiment of this application. Figure 2 This illustration shows a flowchart of a sample-specific gene regulatory network prediction method provided in an embodiment of this application. Figure 1 , Figure 2 As shown, the method may include:

[0045] S101: Obtain dual expression data of regulatory factors and target genes for m samples, where m is an integer greater than or equal to 1.

[0046] In some implementations, given transcriptome data can be acquired first. This given transcriptome data may be transcriptome data related to a pre-specified target disease, such as the disease under study, like leukemia or lymphoma. Correspondingly, the target disease-related transcriptome data may be expression profile data obtained from a public gene expression database based on the target disease. For example, a public gene expression database may include a Gene Expression Omnibus (GEO).

[0047] In some implementations, the acquired transcriptome data can be preprocessed to retain highly expressed regulatory factors and target genes. The regulatory factors can be transcription factors (TFs), microRNAs (miRNAs), circular RNAs (circRNAs), and long non-coding RNAs (lncRNAs), etc.

[0048] S102: Using the dual-expression data as input data, and aiming to minimize the reconstruction error of the input data, the preset model is trained to obtain the target model.

[0049] In some implementations, the preset model is a multi-head attention mechanism-based model (Transformer model), which includes multiple attention heads, each corresponding to a parameter. The preset model is trained using dual-representation data as input data, with the goal of minimizing the reconstruction error of the input data, to obtain the target model. This includes: iteratively updating the parameters corresponding to each of the multiple attention heads in the preset model based on the dual-representation data to obtain the target model. i is an integer greater than or equal to 1 and less than or equal to m. For the i-th sample, each iterative update includes: obtaining the embedding matrix and position vector corresponding to the i-th sample; using the embedding matrix and position vector corresponding to the i-th sample as input, obtaining the sample attention value matrix corresponding to each of the multiple attention heads based on the multi-head attention mechanism; calculating the output attention value matrix corresponding to each of the multiple attention heads based on the sample attention value matrix corresponding to each of the multiple attention heads; reconstructing the sample based on the output attention value matrix corresponding to each of the multiple attention heads to obtain the predicted sample; and updating the parameters of the multiple attention heads using the difference between the predicted sample and the i-th sample as the loss.

[0050] As an example, suppose we have dual expression data of regulatory factors and target genes from m samples. for:

[0051]

[0052] Where m is the number of samples, q is the number of regulatory factors, and p is the number of target genes.

[0053] For the i-th sample out of m samples, a single sample This can be represented by a linear embedding layer as follows:

[0054]

[0055] in, and These are the learnable parameters of the encoder. , . It is the embedded matrix, and d is the embedding dimension.

[0056] Each gene location j (from 1 to ...) Generate location embeddings The position embedding is obtained through an embedding table:

[0057]

[0058] single sample The embedding after adding position encoding in the embedding layer is:

[0059]

[0060] The preset goal of model training is to minimize reconstruction error. Therefore, in each iteration, the output of the multi-head attention module in the target model can be used. The predicted sample is obtained by decoding using a linear decoder. :

[0061]

[0062] in and These are the learnable parameters of the decoder, used to predict samples. , with the original input Compare and calculate the minimum mean square error (MSE) loss. :

[0063]

[0064] In each iteration, the MSE loss can be calculated. The learned parameters are updated through backpropagation and / or an Adaptive Moment Estimation (Adam) optimizer. The iteration count is determined when the preset number of iterations is reached (e.g., 1000), or when the minimum mean squared error (MSE) loss is calculated. If the value is less than a preset threshold, the iteration can be stopped, and the target model can be obtained.

[0065] In some embodiments, the sample attention value matrix corresponding to each of the multiple attention heads can be obtained through the following steps in each iteration:

[0066] by As input, a multi-head attention mechanism is used to obtain the query (Q) matrix for each attention head h (H heads in total). Key (K) matrix and the value (V) matrix :

[0067]

[0068]

[0069]

[0070] in, , , Each attention head h corresponds to the learning parameters of QKV, and .

[0071] based on , as well as Calculate the sample attention value matrix ,include:

[0072]

[0073] in, , It is an activation function that normalizes a numerical vector into a probability distribution vector, where the sum of the probabilities is 1.

[0074] In some embodiments, calculating the output attention value matrix corresponding to each of the multiple attention heads, based on the sample attention value matrix corresponding to each of the multiple attention heads, may include:

[0075] based on The attention value matrix output by each attention head for:

[0076]

[0077] The output of each attention head Concatenate the outputs of the multi-head attention module in the target model :

[0078]

[0079] S103: Obtain the attention value matrix of regulatory factors and target genes at the single-sample level for each of the m samples using the target model.

[0080] In some implementations, the attention value matrix of the regulatory factor and the target gene at the single-sample level for each of the m samples is obtained through the target model, including: for the i-th sample: generating the attention value matrix of the i-th sample at each attention head through the target model; determining the attention value matrix of the regulatory factor and the target gene at the single-sample level for the i-th sample based on the average matrix of the multiple attention value matrices corresponding to the i-th sample.

[0081] For the i-th sample, the i-th sample can be input into the target model, and the attention value matrix corresponding to the i-th sample can be calculated through the target model. For example, the attention value matrix. It can be calculated using the following formula:

[0082]

[0083] In some implementations, for the i-th sample, the attention value matrix between the regulatory factor and the target gene... yes Submatrix. The attention value matrix between regulators and target genes for the i-th sample at the single-sample level can be determined using the following formula:

[0084]

[0085] S104: For each sample, based on the attention value matrix corresponding to the sample, determine multiple target regulator-target gene regulatory relationships, and generate a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulator-target gene regulatory relationships.

[0086] In some implementations, for each sample, based on the attention value matrix corresponding to the sample, multiple target regulatory factor-target gene regulatory relationships are determined, and a sample-specific gene regulatory network corresponding to the sample is generated based on these multiple target regulatory factor-target gene regulatory relationships. This includes: for each sample: calculating the standard score of the attention value matrix corresponding to the sample; obtaining the statistical significance p-value corresponding to the sample based on the standard score, where the statistical significance p-value includes the statistical significance p-values ​​corresponding to each of the multiple candidate regulatory factor-target gene regulatory relationships in the sample; and generating a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulatory factor-target gene regulatory relationships among the multiple candidate regulatory factor-target gene regulatory relationships whose statistical significance p-values ​​satisfy a preset condition. The preset condition includes: the statistical significance p-value is less than a preset p-value correction threshold.

[0087] As an example, the standard score is the z-score, which is the attention value matrix for the i-th sample. The z-score can be calculated using the following formula. :

[0088]

[0089] in:

[0090]

[0091]

[0092] Assuming the z-scores follow a normal distribution, the p-values ​​for each z-score are as follows:

[0093]

[0094] in, It is the standard normal cumulative distribution function. The p-value correction method adopts the Benjamini-Hochberg (BH) method.

[0095] As an example, the preset p-value correction threshold can be set to 0.05. For each regulator-target gene pair, a regulatory relationship is identified when its significance p-value is less than this threshold. For the i-th sample, all regulator-target gene regulatory relationships are merged to construct the specific gene regulatory network for the i-th sample.

[0096] In this application, the beneficial effects of the present invention are manifested as follows: by training the Transformer model with the goal of minimizing the reconstruction error of the input data, the attention mechanism can accurately capture the nonlinear regulatory relationship between regulatory factors and target genes, improving the limitations of traditional linear interpolation or statistical perturbation methods. The entire training process is based on an unsupervised learning framework, avoiding dependence on prior labeled data. At the same time, the parallel computing characteristics of Transformer can effectively reduce the processing time of large amounts of sample data and improve efficiency. The trained target model can output a single-sample level attention value matrix based on the input sample (such as disease transcriptome data). The attention value matrix can quantify the specific regulatory intensity within the sample. Combined with the statistical process of Z-score conversion, p-value correction, and threshold screening, a specific gene regulation network for each sample (i.e., a disease sample-specific gene regulation network) can be constructed. This realizes the Transformer deep learning model based on the attention mechanism, which can efficiently and accurately construct a nonlinear gene regulation network at the precision level of a single sample (a biological tissue or cell), providing technical support for personalized diagnosis and treatment of complex human diseases in clinical practice.

[0097] The method provided in this application will be further described below through two embodiments.

[0098] Example 1:

[0099] This embodiment uses the prediction of leukemia cell-specific miRNA regulatory networks as an example to illustrate the method provided in this application.

[0100] In this embodiment, the target disease is leukemia, and the transcriptome data can be K562 leukemia miRNA and mRNA co-sequencing data collected from the public database GEO (GeneExpression Omnibus, https: / / www.ncbi.nlm.nih.gov / geo / ) (dataset number GSE114071).

[0101] After preprocessing the transcriptome data (e.g., logarithmic transformation, removal of duplicate genes and constant expression genes), expression profiles of 212 miRNAs and 15361 mRNAs from 19 K562 leukemia cells were obtained. Further preprocessing of the expression profile data can then be performed to retain highly expressed regulatory factors and target genes. For example, feature extraction can be performed on the expression profile data to retain mRNAs with the highest expression variance, ultimately yielding expression profiles of 212 miRNAs and 7680 mRNAs from 19 matched K562 leukemia cell samples. Therefore, in this embodiment, regulatory factors... and target genes Data on dual expression of regulatory factors and target genes .

[0102] Using the methods provided above, The following processing was performed: during Transformer model training, the number of iterations was set to 1000, the Adam optimizer was used to optimize the training model, the embedding dimension d was set to 16, the number of attention heads H was set to 2, and the p-value correction threshold was set to 0.05. Finally, the calculated sizes of the 19 K562 leukemia cell-specific miRNA regulatory networks were 5741, 5539, 5610, 5357, 5712, 5442, 5733, 5638, 5588, 5350, 5516, 5466, 5358, 5396, 5706, 5713, 5472, 5372, and 5473.

[0103] The regulatory network of the 19 K562 leukemia cell-specific miRNAs obtained above can be validated through the following three dimensions:

[0104] (1) Network similarity analysis

[0105] Given the miRNA regulatory networks of cells a and b in 19 K562 leukemia cells and Network similarity value The calculation is as follows:

[0106]

[0107] in, The logarithm of the same miRNA regulatory relationships in the two networks. This represents the logarithm of the miRNA regulatory relationship in the smallest of the two networks. The value ranges from [0, 1], and the larger the value, the more similar the two networks are.

[0108] (2) Cluster analysis

[0109] Network similarity analysis yielded network similarity matrices for 19 K562 leukemia cell-specific miRNA regulatory networks:

[0110]

[0111] The network similarity matrix can be used as a cell similarity matrix to obtain the cell distance (Distance, Dis) matrix, which is defined as follows:

[0112]

[0113] Based on the cell distance matrix Dis, hierarchical clustering can be used to perform cluster analysis on 19 leukemia cells.

[0114] (3) Verification Analysis

[0115] Obtain experimentally validated miRNA regulatory relationships collected from third-party databases, and calculate the percentage overlap between miRNA regulatory relationships in the leukemia cell-specific miRNA regulatory network and experimentally validated miRNA regulatory relationships. A higher overlap percentage indicates that more predicted miRNA regulatory relationships in leukemia cells have been experimentally confirmed. The third-party database can be either miRTarBase v10.0 or TarBase v9.0.

[0116] For example, the experimental verification of miRNA target genes can be obtained from two databases, miRTarBase v10.0 and TarBase v9.0, and a total of 320,010 miRNA target regulatory relationship pairs related to the K562 leukemia dataset were obtained.

[0117] Figure 3 A similarity diagram of the miRNA regulatory network in leukemia cells in Example 1 of this application is shown. Table 1 shows the cluster analysis results in Example 1. Table 2 shows the predicted and experimental verification relationships of leukemia miRNA regulation in Example 1.

[0118] Table 1

[0119] Cluster number Clustering of leukemia cells Cluster 1 Cell 1, Cell 2, Cell 5, Cell 7, Cell 8, Cell 16, Cell 17, Cell 18 Cluster 2 Cell 3, Cell 4, Cell 6, Cell 9, Cell 10, Cell 11, Cell 12, Cell 13, Cell 14, Cell 15 Cluster 3 Cell 19

[0120] Table 2

[0121] Network ID Predicting miRNA regulatory relationships Validating miRNA regulatory relationships 1 5741 280 2 5539 247 3 5610 252 4 5357 236 5 5712 263 6 5442 244 7 5733 263 8 5638 258 9 5588 257 10 5350 221 11 5516 229 12 5466 238 13 5358 223 14 5396 228 15 5706 250 16 5713 256 17 5472 242 18 5372 252 19 5473 240

[0122] In this embodiment, the similarity values ​​of the miRNA regulatory networks of 19 leukemia cells ranged from [0.85, 0.92], indicating that the miRNA regulatory networks of each leukemia cell are different. The hierarchical clustering method was used to classify the 19 leukemia cells into three categories, as shown in Table 1. As shown in Table 2, the predicted miRNA regulatory relationships and the experimentally verified miRNA regulatory relationships also differed in each leukemia cell.

[0123] Example 2:

[0124] This embodiment uses the prediction of a mouse lymphoma sample-specific TF regulatory network as an example to illustrate the method provided in this application.

[0125] In this embodiment, the target disease is lymphoma, and the transcriptome data can be K562 leukemia miRNA and mRNA co-sequencing data collected from the public database GEO (GeneExpression Omnibus, https: / / www.ncbi.nlm.nih.gov / geo / ) (dataset number GSE114071).

[0126] After preprocessing the transcriptome data (such as logarithmic transformation and removal of duplicate genes), expression profiles of 934 TFs and 12,545 mRNAs from 26 lymphoma samples were obtained. Further preprocessing of the expression profile data can then be performed to retain highly expressed regulatory factors and target genes. For example, feature extraction can be performed on the expression profile data to retain the top 20% of TFs and mRNAs with the largest variance in expression levels, ultimately yielding expression profiles of 187 TFs and 2,509 mRNAs from the 26 lymphoma samples. Therefore, in this embodiment, regulatory factors... and target genes Data on dual expression of regulatory factors and target genes .

[0127] Using the methods provided above, The process involved setting the number of iterations to 1000 during Transformer model training, using the Adam optimizer, setting the embedding dimension d to 16, the number of attention heads H to 2, and the p-value correction threshold to 0.05. Finally, the calculated TF regulatory network sizes specific to 26 lymphoma samples were 7476, 9662, 7402, 6523, 8575, 9874, 6942, 5795, 5169, 8499, 6495, 5386, 9849, 7236, 6795, 7582, 5630, 9273, 7419, 3390, 4913, 6488, 5197, 7287, 6625, and 5964.

[0128] The specific TF regulatory networks of the 26 lymphoma samples obtained above can be validated through the following three dimensions:

[0129] (1) Network similarity analysis

[0130] Given the TF regulatory networks of sample a and sample b from 26 lymphoma samples. and Network similarity value The calculation is as follows:

[0131]

[0132] in, For the logarithm of the same TF regulation relationship in the two networks, Let be the logarithm of the TF regulation relationship of the smallest of the two networks. The value ranges from [0, 1], and the larger the value, the more similar the two networks are.

[0133] (2) Cluster analysis

[0134] Network similarity analysis yielded network similarity matrices for miRNA regulatory networks specific to 26 lymphoma samples:

[0135]

[0136] The network similarity matrix can be used as a cell similarity matrix to obtain the cell distance (Distance, Dis) matrix, which is defined as follows:

[0137]

[0138] Based on the cell distance matrix Dis, hierarchical clustering can be used to perform cluster analysis on 26 lymphoma samples.

[0139] (3) Verification Analysis

[0140] The experimentally validated TF regulatory relationships are obtained from a third-party database. The percentage of overlap between the TF regulatory relationships in the mouse lymphoma sample-specific TF regulatory network and the experimentally validated TF regulatory relationships is calculated. A higher percentage of overlap indicates that more predicted TF regulatory relationships are validated in that lymphoma sample. The third-party database can be the CollicTRI database.

[0141] For example, experimental verification of TF target genes was obtained from the CollecTRI database, and a total of 1718 TF target regulatory pairs were obtained related to the mouse lymphoma dataset.

[0142] Figure 4 Table 3 shows a similarity diagram of the TF regulatory network in lymphoma samples from Example 2 of this application. Table 4 shows the cluster analysis results from Example 2. Table 5 shows the predicted and experimentally validated TF regulatory relationships in lymphoma samples from Example 2.

[0143] In this embodiment, the similarity values ​​of the TF regulatory networks of the 26 lymphoma samples ranged from [0.23, 0.84], indicating that the TF regulatory networks differed among the lymphoma samples. The hierarchical clustering method was used to divide the 26 lymphoma samples into 5 categories, as shown in Table 3. As shown in Table 4, the predicted TF regulatory relationships and the experimentally verified TF regulatory relationships also differed among the lymphoma samples.

[0144] Table 3

[0145] Cluster number Clustered lymphoma samples Cluster 1 Sample 1, Sample 3, Sample 4, Sample 18, Sample 24 Cluster 2 Sample 2, Sample 5, Sample 6, Sample 7, Sample 10, Sample 12, Sample 25, Sample 26 Cluster 3 Sample 8, Sample 9, Sample 14, Sample 19 Cluster 4 Sample 11, Sample 13, Sample 15, Sample 20, Sample 21, Sample 22, Sample 23 Cluster 5 Sample 16, Sample 17

[0146] Table 4

[0147] Network ID Predicting TF-regulatory relationships Verify the TF regulatory relationship 1 7476 54 2 9662 60 3 7402 44 4 6523 24 5 8575 52 6 9874 49 7 6942 29 8 5795 37 9 5169 44 10 8499 62 11 6495 26 12 5386 19 13 9849 87 14 7236 51 15 6795 28 16 7582 31 17 5630 43 18 9273 51 19 7419 21 20 3390 10 21 4913 28 22 6488 28 23 5197 21 24 7287 57 25 6625 63 26 5964 61

[0148] The results in Examples 1 and 2 demonstrate that gene regulatory networks exhibit sample (live tissue cell) specificity, thus necessitating prediction of gene regulatory networks at the single-sample accuracy level. Furthermore, sample-to-sample correlation networks can be constructed using sample-specific gene regulatory networks, providing novel methods for sample typing and annotation. In summary, the sample-specific gene regulatory network prediction method proposed in this invention can identify cancer sample-specific gene regulatory networks, providing technical support and understanding tools for personalized diagnosis and targeted therapy of human cancer in clinical practice, and possessing significant biological implications.

[0149] It should be noted that although this embodiment describes a sample-specific miRNA and TF regulatory network prediction case, it is also applicable to predicting single-sample-specific regulatory networks of other regulators (such as lncRNA, circRNA, and pseudogene) and target genes.

[0150] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting sample-specific gene regulatory networks, characterized in that, include: Obtain dual expression data of regulatory factors and target genes for m samples, where m is an integer greater than or equal to 1; Using the dual-expression data as input data, and aiming to minimize the reconstruction error of the input data, a preset model is trained to obtain the target model; The attention value matrix of regulatory factors and target genes at the single-sample level is obtained through the target model for each of the m samples. For each sample, based on the attention value matrix corresponding to the sample, multiple target regulatory factor-target gene regulatory relationships are determined, and a sample-specific gene regulatory network corresponding to the sample is generated according to the multiple target regulatory factor-target gene regulatory relationships.

2. The method according to claim 1, characterized in that, The preset model is a model based on a multi-head attention mechanism, which includes multiple attention heads, each of which corresponds to a parameter; The step of using the dual-expression data as input data and training a preset model to obtain a target model with the goal of minimizing the reconstruction error of the input data includes: Based on the dual-expression data, the parameters corresponding to each of the multiple attention heads in the preset model are updated in multiple rounds to obtain the target model.

3. The method according to claim 2, characterized in that, i is an integer greater than or equal to 1 and less than or equal to m. For the i-th sample, each iteration update includes: Obtain the embedding matrix and position vector corresponding to the i-th sample; Using the embedding matrix and position vector corresponding to the i-th sample as input, and based on the multi-head attention mechanism, obtain the sample attention value matrix corresponding to each of the multiple attention heads; Based on the sample attention value matrix corresponding to each of the multiple attention heads, calculate the output attention value matrix corresponding to each of the multiple attention heads; Based on the output attention value matrix corresponding to each of the multiple attention heads, sample reconstruction is performed to obtain the predicted sample; The parameters of the multiple attention heads are updated using the difference between the predicted sample and the i-th sample as the loss.

4. The method according to any one of claims 2 or 3, characterized in that, The step of obtaining the attention value matrix of regulatory factors and target genes at the single-sample level for each of the m samples through the target model includes: For the i-th sample: The target model is used to generate the attention value matrix of the i-th sample at each attention head; The attention value matrix of the regulatory factor and the target gene for the i-th sample at the single-sample level is determined by averaging the multiple attention value matrices corresponding to the i-th sample.

5. The method according to claim 4, characterized in that, The step of determining the attention value matrix of the regulatory factor and the target gene of the i-th sample at the single-sample level based on the average matrix of multiple attention value matrices corresponding to the i-th sample includes: The attention value matrix of the regulatory factor and target gene for the i-th sample at the single-sample level is determined according to the following formula: ； in, This represents the attention value matrix of the regulatory factor and the target gene for the i-th sample at the single-sample level. This represents the average matrix of the multiple attention value matrices corresponding to the i-th sample. This represents the number of regulatory factors in the i-th sample. This represents the number of target genes in the i-th sample.

6. The method according to claim 1, characterized in that, For each sample, based on the attention value matrix corresponding to the sample, multiple target regulatory factor-target gene regulatory relationships are determined, and a sample-specific gene regulatory network corresponding to the sample is generated according to the multiple target regulatory factor-target gene regulatory relationships, including: For each sample: Calculate the standard score of the attention value matrix corresponding to the sample; The statistical significance p-value corresponding to the sample is obtained based on the standard score. The statistical significance p-value corresponding to the sample includes the statistical significance p-values ​​corresponding to the regulatory relationships of multiple candidate regulatory factors and target genes in the sample. Based on the multiple candidate regulatory factor-target gene regulatory relationships, the target regulatory factor-target gene regulatory relationships in which the statistical significance p-value satisfies the preset conditions are generated to form a sample-specific gene regulatory network corresponding to the sample.

7. The method according to claim 6, characterized in that, The preset conditions include: The statistical significance p-value is less than the preset p-value correction threshold.

8. A method for predicting sample-specific gene regulatory networks, characterized in that, include: The acquisition module is used to acquire dual expression data of regulatory factors and target genes corresponding to m samples, where m is an integer greater than or equal to 1; The training module is used to train a preset model by taking the dual-expression data as input data and minimizing the reconstruction error of the input data to obtain the target model. The acquisition module is further configured to acquire the attention value matrix of regulatory factors and target genes at the single-sample level for each of the m samples through the target model; The generation module, for each sample, is used to determine multiple target regulator-target gene regulatory relationships based on the attention value matrix corresponding to the sample, and generate a sample-specific gene regulatory network corresponding to the sample based on the multiple target regulator-target gene regulatory relationships.

9. An electronic device, characterized in that, It includes a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions that can be executed by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on a computer-readable storage medium, and the computer program, when executed by a processor, performs the method of any one of claims 1-7.

Images