Disease-related alternative splicing isoform prediction model establishment and prediction method thereof
By using a shared graph convolutional neural network and an adversarial domain adaptation model, a disease-related alternative splicing isoform prediction model is constructed, which solves the problems of insufficient reliability and accuracy in existing technologies and achieves higher prediction reliability and accuracy.
Patent Information
- Application Number
- CN202310784979.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-29
- Publication Date
- 2025-12-26
- Estimated Expiration
- 2043-06-29
AI Technical Summary
Existing technologies have poor reliability and accuracy in predicting disease-related alternative splicing isomers and fail to effectively consider the association between isomers and diseases.
By combining a shared graph convolutional neural network model with an adversarial domain adaptation model, and by constructing a gene co-expression network and an alternative splicing isoform co-expression network, node features and association information are learned, and generative adversarial networks are used to achieve embedding space alignment, thereby constructing a disease-related alternative splicing isoform prediction model.
It improves the reliability and accuracy of predicting disease-related alternative splicing isomers, enabling more accurate prediction of pathogenic isomers and supporting the discovery of disease biomarkers and drug targets.
Smart Images
Figure CN116825178B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The embodiment of the present disclosure relates to the technical field of bioinformatics, in particular to a disease-related alternative splicing isoform prediction model and a prediction method thereof. BACKGROUND
[0002] Alternative splicing (AS) refers to a process of producing different mRNA isoforms by different splicing ways from a mRNA precursor. A gene can produce multiple isoforms with different functions through alternative splicing. Many studies have found that AS is related to some complex diseases (such as Alzheimer's disease). Traditional methods consider a gene as a single entity without distinguishing these isoforms with different functions. However, for a pathogenic gene, not all isoforms are related to the disease. Therefore, accurately predicting the pathogenic isoform in the pathogenic gene is of great significance for accurately understanding the pathogenesis and discovering disease markers and drug targets. With the development of high-throughput sequencing technology, many methods for predicting disease-related alternative splicing isoforms have been proposed, such as IsoResolve. IsoResolve is a method based on domain adaptation (DA) to predict isoform functional annotation. This method uses gene-level and isoform-level features and gene functional annotation as input, and isoform-level functional annotation label as output.
[0003] Although the above method has been successfully applied to real environment, only the analysis based on the functional label of the isoform is made, and the relevance between the isoform and the disease is ignored, which limits the current method in the field of isoform function prediction and restricts the development of related technologies.
[0004] Therefore, there is an urgent need for a disease-related alternative splicing isoform prediction model and a prediction method thereof with high reliability and accuracy. SUMMARY
[0005] Therefore, the embodiment of the present disclosure provides a disease-related alternative splicing isoform prediction model and a prediction method thereof, which at least partially solves the problem of poor reliability and accuracy in the prior art.
[0006] The embodiment of the present disclosure provides a disease-related alternative splicing isoform prediction model and a prediction method thereof, which comprises:
[0007] Step 1, collecting real RNA-seq data and labels of disease-related genes and corresponding alternative splicing isoforms;
[0008] Step 2, calculate the expression data of disease-related genes and corresponding alternative splicing isoforms according to the RNA-seq data;
[0009] Step 3, preprocess the expression data respectively to obtain sample data and split the sample data into a training set and a test set;
[0010] Step 4, use WGCNA to construct a gene co-expression network and an alternative splicing isoform co-expression network respectively, and construct a shared graph convolutional neural network model according to the gene co-expression network and the alternative splicing isoform co-expression network;
[0011] Step 5, construct an adversarial domain adaptation model and combine the shared graph convolutional neural network model to form an initial prediction model, train the initial prediction model using the training set, and obtain a target prediction model;
[0012] Step 6, calculate the evaluation parameters of the target prediction model using the test set and determine whether the target prediction model meets the preset requirements, if yes, execute Step 7, and if no, return to Step 5;
[0013] Step 7, input the alternative splicing isoform to be detected into the target prediction model to obtain a predicted pathogenic result.
[0014] According to a specific implementation manner of the embodiment of the present disclosure, the step 2 specifically comprises:
[0015] Step 2.1, download the human genome and its annotation file and sequence file from an initial database;
[0016] Step 2.2, index the RNA-seq data through bioinformatics software according to the annotation file and the sequence file, and perform alignment on the RNA-seq data through the bioinformatics software and the obtained index, and convert the sequence FASTQ file into a BAM file;
[0017] Step 2.3, establish a genome index according to the annotation file and the sequence file;
[0018] Step 2.4, calculate gene expression values and alternative splicing isoform expression values according to the genome index and the BAM file;
[0019] Step 2.5, extract the FPKM values corresponding to the gene expression values and the alternative splicing isoform expression values as the expression data of the disease-related genes and the corresponding alternative splicing isoforms, respectively.
[0020] According to a specific implementation manner of the embodiment of the present disclosure, the step 4 specifically comprises:
[0021] Step 4.1, perform data input and cleaning using the goodSamplesGenes method of WGCNA to remove missing values and outliers;
[0022] Step 4.2, using WGCNA to select the soft threshold value as the network construction threshold value;
[0023] Step 4.3, constructing the network according to the soft threshold value 7 through the adjacency method of WGCNA to obtain the adjacency matrix of the gene network and the isomer network;
[0024] Step 4.4, transforming the adjacency matrix of the gene network and the adjacency matrix of the isomer network into the topological overlap matrix TOM as the construction of the gene co-expression network and the alternative splicing isomer co-expression network through the TOMsimilarity method of WGCNA;
[0025] Step 4.5, defining the gene co-expression network as the source domain network, and the corresponding alternative splicing isomer co-expression network as the target domain network, and constructing a shared graph convolutional neural network model according to the source network and the target domain network.
[0026] According to a specific implementation manner of an embodiment of the present disclosure, the shared graph convolutional neural network model includes two hidden layers, and the number of neurons is 128 and 16 respectively, and each layer is connected with a Dropout layer.
[0027] According to a specific implementation manner of an embodiment of the present disclosure, the expression of the hidden layer is
[0028]
[0029] Wherein, X represents an input feature matrix, A represents an adjacency matrix, represents an adjacent matrix with self-loop, In∈Rn×n represents a unit matrix, and is a normalized adjacency matrix, Z (i-1) is the output of the i-1 layer, Z 0 =X, W i is a trainable parameter of the network, and σ(·) represents an activation function.
[0030] According to a specific implementation manner of an embodiment of the present disclosure, the adversarial domain adaptation model includes a domain classifier, a source domain classifier and a target classifier.
[0031] According to a specific implementation manner of an embodiment of the present disclosure, the step of training the initial prediction model by using the training set to obtain the target prediction model includes:
[0032] Setting a preset learning rate and an iteration number, inputting the training set into the initial prediction model for training until the evaluation index meets the preset condition or the iteration number is reached, and obtaining the target prediction model.
[0033] According to a specific implementation manner of an embodiment of the present disclosure, the expression of the average index is
[0034]
[0035] wherein True positive is the number of samples predicted as positive and true as positive, True negative is the number of samples predicted as negative and true as negative, and All samples is the total number of samples.
[0036] The disease-related alternative splicing isoform prediction model and the prediction scheme thereof in the embodiments of the present disclosure include: step 1, collecting real RNA-seq data and labels of disease-related genes and corresponding alternative splicing isoforms; step 2, calculating the expression data of the disease-related genes and the corresponding alternative splicing isoforms according to the RNA-seq data; step 3, pre-processing the expression data respectively to obtain sample data and splitting the sample data into a training set and a test set; step 4, constructing a gene co-expression network and an alternative splicing isoform co-expression network using WGCNA, and constructing a shared graph convolutional neural network model according to the two networks; step 5, constructing an adversarial domain adaptation model and combining the shared graph convolutional neural network model to form an initial prediction model, and training the initial prediction model using the training set to obtain a target prediction model; step 6, calculating the evaluation parameters of the target prediction model using the test set and determining whether the target prediction model meets the preset requirements, if yes, executing step 7, and if not, returning to step 5; and step 7, inputting a to-be-detected alternative splicing isoform into the target prediction model to obtain a predicted pathogenic result.
[0037] The embodiments of the present disclosure have the following beneficial effects: through the scheme of the present disclosure, the graph convolutional neural network is used to automatically learn the node features of the gene domain and the isoform domain, and learn the association information between nodes, and extract the features of the two levels of genes and isoforms. The adversarial learning regularization based on the generative adversarial network is used to realize the embedding space alignment and distribution alignment of the gene domain and the isoform domain, effectively migrate the knowledge of the gene domain to the isoform prediction model, and realize the prediction of the pathogenicity of the isoform based on the gene and isoform co-expression network and the gene pathogenic label, and the reliability is high and the accuracy is good. BRIEF DESCRIPTION OF DRAWINGS
[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art without creating laborious work.
[0039] Figure 1 A flowchart of a disease-related alternative splicing isoform prediction model and a prediction method thereof provided by the embodiments of the present disclosure. DETAILED DESCRIPTION
[0040] The embodiments of the present disclosure will be described in detail below with reference to the drawings.
[0041] The following detailed description of the embodiments of the present disclosure is provided for the purpose of understanding the present disclosure. It is obvious, however, that the embodiments described are only a part of the embodiments of the present disclosure, and are not all the embodiments of the present disclosure. The present disclosure can also be implemented or applied in other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present disclosure.
[0042] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the teachings provided herein one skilled in the art should be able to contemplate these and similar aspects of the present disclosure. For example, the device and / or method can be implemented using any number of aspects of the present disclosure described herein. In addition, the device and / or method can be implemented using other structure and / or functionality in addition to or other than one or more aspects of the present disclosure described herein.
[0043] It should also be noted that the drawings provided in the following embodiments are only schematic and are intended to provide the basic understanding of the present disclosure. In the drawings, only the components related to the present disclosure are shown, not the number, shape and size of the components as implemented, and the shape, number and ratio of the components can be arbitrarily changed in actual implementation, and the layout of the components can also be more complex.
[0044] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, one skilled in the art will understand that the described aspects can be practiced without these specific details.
[0045] The embodiments of the present disclosure provide a disease-related alternative splicing isoform prediction model and a prediction method thereof, which can be applied to the disease prediction process in a medical scene.
[0046] Referring to Figure 1 A flowchart of a disease-related alternative splicing isoform prediction model and a prediction method thereof provided by the embodiments of the present disclosure is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps:Figure 1 As shown, the method mainly comprises the following steps:
[0047] Step 1, collecting real RNA-seq data of disease-related genes and corresponding variable splicing isoforms and labels;
[0048] In practice, the embodiments of the present disclosure take Alzheimer's disease as an example, collect real RNA-seq data of Alzheimer's disease-related genes and corresponding variable splicing isoforms, and specifically 80 postmortem brain samples of old people without neurodegenerative diseases. All samples are collected from two places and sequenced in the same place. Specifically, total RNA is extracted from the temporal lobe cortex using Trizol reagent, and RIN is measured using Agilent Technologies 2100 Bioanalyzer. The cDNA library is prepared using TruSeq RNA Sample Prep Kit (Illumina, San Diego, CA), and then sequenced using 101 base pair end sequencing Illumina HiSeq 2000 sequencer. The average sequencing of this sample is 100 million reads, which is translated into about 70 reads per base of human transcriptome sequencing depth. And in order to reduce the interference of factors such as gender, age, RNA integrity number (RIN), the expression data is adjusted for covariates such as gender, age, sample source by regression to exclude confounding effects.
[0049] Step 2, calculating the expression data of disease-related genes and corresponding variable splicing isoforms according to the RNA-seq data;
[0050] On the basis of the above embodiments, the step 2 specifically comprises:
[0051] Step 2.1, downloading the human genome and its annotation file and sequence file from the initial database;
[0052] Step 2.2, according to the annotation file and sequence file, establishing an index through bioinformatics software and aligning the RNA-seq data through bioinformatics software and the obtained index, and converting the sequence FASTQ file to BAM file;
[0053] Step 2.3, establishing a genome index according to the annotation file and sequence file;
[0054] Step 2.4, calculating gene expression values and variable splicing isoform expression values according to the genome index and BAM file;
[0055] Step 2.5, extracting the corresponding FPKM values of the gene expression values and the variable splicing isoform expression values as the expression data of the disease-related genes and the corresponding variable splicing isoforms, respectively.
[0056] In specific implementation, the expression data of disease-related genes and corresponding variable splicing isomers are calculated according to the RNA-seq data, and the specific calculation steps are as follows:
[0057] A. Download the annotation GTF file and sequence FA file of the release-35 version of the GRCh38 human genome from the gencode database;
[0058] B. Index the GTF and FA obtained in step A using STAR, and align the RNA-seq data of 80 brain samples obtained from the clinic using STAR and the obtained index to convert the sequence FASTQ file to a BAM file.
[0059] C. Use the GTF and FA obtained in step A to establish a genomic index through the rsem-prepare-reference function option in RSEM-1.3.3.
[0060] D. Calculate gene expression values and variable splicing isomer expression values by using the index obtained in step C and the BAM file obtained in step B through the rsem-calculate-expression function option in RSEM-1.3.3.
[0061] F. Extract the FPKM value (Fragments Per Kilobase Million) from the result file obtained in step D with the suffix genes.results and isoforms.results as gene and isoform expression value data, respectively.
[0062] Step 3, the expression data is pre-processed respectively to obtain sample data and split it into training set and test set;
[0063] In specific implementation, considering that the collected expression data may have low expression conditions such as missing or errors, the expression data can be pre-processed respectively to obtain sample data, and the training set and test set are split under the condition that the positive and negative label ratios are the same, and the ratio of the number of genes in the training set to the number of genes in the test set is 7:3.
[0064] Step 4, using WGCNA to construct gene co-expression network and variable splicing isomer co-expression network respectively, and constructing shared graph convolutional neural network model according to the same;
[0065] Further, the step 4 specifically includes:
[0066] Step 4.1, using the goodSamplesGenes method of WGCNA for data input and cleaning to remove missing values and outliers;
[0067] Step 4.2, using WGCNA to select the soft threshold value as the network construction threshold value;
[0068] Step 4.3, according to the soft threshold value 7, the adjacency method of WGCNA is used to construct the network, and the adjacency matrix of the gene network and the isomer network is obtained;
[0069] Step 4.4, according to the adjacency matrix of the gene network and the adjacency matrix of the isomer network, the TOMsimilarity method of WGCNA is used to transform into the topological overlap matrix TOM as the construction of the gene co-expression network and the alternative splicing isomer co-expression network;
[0070] Step 4.5, the gene co-expression network is defined as the source domain network, and the corresponding alternative splicing isomer co-expression network is defined as the target domain network, and the shared graph convolutional neural network model is constructed according to the source network and the target domain network.
[0071] Optionally, the shared graph convolutional neural network model includes two hidden layers, and the number of neurons is 128 and 16 respectively, and each layer is connected with a Dropout layer.
[0072] Further, the expression of the hidden layer is
[0073]
[0074] Wherein, X represents an input feature matrix, A represents an adjacency matrix, represents an adjacent matrix with self-loop, In∈Rn×n represents a unit matrix, And is a normalized adjacency matrix, Z (i-1) is the output of the i-1 layer, Z 0 = X, W i is a trainable parameter of the network, and σ(·) represents an activation function.
[0075] In specific implementation, the process of using WGCNA to construct the gene co-expression network and the alternative splicing isomer co-expression network and constructing the shared graph convolutional neural network model can be as follows:
[0076] The goodSamplesGenes method of WGCNA is used for data input and cleaning, and missing values and outliers are removed;
[0077] The soft threshold value is selected by using WGCNA, and finally the soft threshold value of 7 is selected as the network construction threshold value;
[0078] The soft threshold value 7 obtained according to step B is used to construct a network by the adjacency method of WGCNA to obtain the adjacency matrix of the gene network and the isomer network;
[0079] In order to reduce the influence of noise and false correlation, the adjacency matrix obtained in step B is transformed into a topological overlap matrix TOM as a gene weighted network and an isomer weighted network by the TOMsimilarity method of WGCNA;
[0080] Then the gene co-expression network is defined as the source domain network, and the corresponding alternative splicing isomer co-expression network is defined as the target domain network. A graph convolutional neural network sharing parameters is constructed according to the source network and the target network, and the graph convolutional neural network includes two hidden layers (L=2), and the number of neurons is 128 and 16 respectively. A Dropout layer is connected after each layer to prevent overfitting, and the probability of randomly discarding neurons is set to 0.1. The hidden layer of the output network Z is defined as:
[0081]
[0082] wherein, is the adjacent matrix with self-loop (In∈Rn×n is the unit matrix), and is the normalized adjacency matrix. Z (i-1) is the output of the (i-1)th layer, Z 0 =X, W i is the trainable parameter of the network, and σ(·) represents an activation function.
[0083] Step 5, constructing an adversarial domain adaptation model and combining the shared graph convolutional neural network model to form an initial prediction model, and training the initial prediction model by using the training set to obtain a target prediction model;
[0084] Further, the adversarial domain adaptation model includes a domain classifier, a source domain classifier and a target classifier.
[0085] Further, the step of training the initial prediction model by using the training set to obtain the target prediction model includes:
[0086] The preset learning rate and the number of iterations are set, the training set is input into the initial prediction model for training until the evaluation index meets the preset condition or the number of iterations is reached, and the target prediction model is obtained.
[0087] Further, the expression of the average index is
[0088]
[0089] Where True positive is the number of samples predicted as positive and true as positive, True negative is the number of samples predicted as negative and true as negative, and All samples is the total number of samples.
[0090] In specific implementation, the adversarial domain adaptation model can be composed of a domain classifier, a source domain classifier and a target classifier. The domain classifier is used to determine the source of the source domain embedding and the target domain embedding generated by the shared graph convolutional neural network model. The source domain classifier is used to determine the pathogenic label of the source domain, and the target domain classifier is used to determine the pathogenic label of the target domain. The three classifiers learn class discrimination and domain-invariant node representation together, so as to realize classification of nodes in the target network.
[0091] Where the source domain label classifier and the target domain label classifier are a binary classification full connection layer. The domain classifier includes one hidden layer and one binary classification full connection layer, the number of neurons of the hidden layer is 40, a ReLU activation function is used and a Dropout layer is connected to prevent overfitting, and the probability of randomly discarding neurons is set to 0.1.
[0092] Three kinds of loss functions can be obtained through the three classifiers. The loss function of the domain classifier is a binary cross-entropy loss, and the calculation formula is:
[0093]
[0094] Where i is each sample, m i ∈{0,1} represents the domain label of the sample, from the source domain m i is 0, from the target domain m i is 1. The domain prediction labels of the samples by the domain classifier are respectively N s and N t represent the number of nodes of the source domain and the target domain respectively.
[0095] The source domain label loss function is a cross-entropy loss function, and the calculation formula is:
[0096]
[0097] Where i is each sample, y i represents the true label of the source domain, represents the predicted label of the source domain classifier, and N s represents the number of nodes of the source domain.
[0098] Since the target domain has no true label, an entropy loss is used as the loss function, and the calculation formula is:
[0099]
[0100] where i is each sample, denotes the predicted label of the target domain domain classifier, and Nt denotes the number of target domain nodes.
[0101] The overall loss function is calculated as follows
[0102] L = L s + γ1L DA + γ2L T
[0103] where γ1, γ2 are balance parameters. L s , L DA and L T respectively denote the source domain classifier loss, the domain classifier loss and the target domain classifier loss.
[0104] The neural network constructed by using (1) and (2) is trained by using the training set obtained in step 4. The optimizer used by the three classifiers is Adam, the learning rate is 3e-3, and the number of iterations is 200;
[0105] The evaluation index is accuracy, and the calculation formula is:
[0106]
[0107] where Truepositive is the number of samples predicted as positive and actually positive; Turenegative is the number of samples predicted as negative and actually negative; and Allsamples is the total number of samples.
[0108] Step 6, using the test set to calculate the evaluation parameters of the target prediction model and determining whether it meets the preset requirements, if yes, executing step 7, if no, returning to step 5;
[0109] In specific implementation, after the target prediction model is trained, the evaluation parameters of the target prediction model can be calculated using the test set and determining whether it meets the preset requirements, if yes, it means that the prediction accuracy of the model is sufficient enough for actual application of prediction, if no, it means that the prediction accuracy of the model is not enough enough, then the training step can be returned for retraining.
[0110] Step 7, inputting the variable splicing isoform to be detected into the target prediction model to obtain a predicted pathogenic result.
[0111] In specific implementation, when it is necessary to perform pathogenic prediction on a certain gene, the variable splicing isoform to be detected can be input into the target prediction model to obtain a predicted pathogenic result.
[0112] The disease-related alternative splicing isoform prediction model provided by the embodiment and the prediction method thereof automatically learn the node features of the gene domain and the isoform domain by using a graph convolutional neural network, and learn the association information between nodes, and extract the features of the gene and the isoform in two levels. The adversarial learning regularization based on the generative adversarial network is used to realize the embedding space alignment and distribution alignment of the gene domain and the isoform domain, effectively migrate the knowledge of the gene domain to the isoform prediction model, and realize the prediction of the pathogenicity of the isoform based on the gene and the isoform co-expression network and the gene pathogenic label, and the reliability is high and the accuracy is good.
[0113] The present scheme will be described below in combination with a specific embodiment. The present application is evaluated on the real data set AD obtained from step 1, and IsoResolve is used for comparison. In order to comprehensively compare the performance of the two methods, the AUROC and AUPRC results of the two methods are compared on the same number of positive samples (179) and different positive and negative sample ratio data sets. The positive and negative sample labels come from the article “Integrative functional genomic analysis of intron retention in human and mouse brain with Alzheimer’s disease.” The AUROC comparison results are shown in Table 1, and the AUPRC comparison results are shown in Table 2.
[0114] Table 1
[0115]
[0116] Table 2
[0117]
[0118] From the results in Table 2, although the AUROC and AUPRC decrease with the increase of the negative sample ratio, the AUROC and AUPRC of the present application are better than IsoResolve on different positive and negative ratio data sets, but it is noted that. This shows that the performance of the present application is better.
[0119] The units described in the embodiments of the present disclosure can be implemented in the form of software or in the form of hardware.
[0120] It should be understood that parts of the present disclosure can be realized by hardware, software, firmware or a combination thereof.
[0121] The above merely provides the specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present disclosure, which should be covered in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.
Claims
1. A disease-related alternative splicing isoform prediction model establishment and prediction method, characterized in that, The method comprises the following steps: Step 1, collecting real RNA-seq data of disease-related genes and corresponding variable splicing isoforms and labels; Step 2, calculating the expression data of disease-related genes and corresponding variable splicing isoforms according to the RNA-seq data; Step 3, preprocessing the expression data respectively to obtain sample data and splitting the sample data into a training set and a test set; Step 4, using WGCNA to construct a gene co-expression network and a variable splicing isoform co-expression network respectively, and constructing a shared graph convolutional neural network model according to the gene co-expression network and the variable splicing isoform co-expression network; The step 4 specifically comprises: Step 4.1, using the goodSamplesGenes method of WGCNA for data input and cleaning to remove missing values and outliers; Step 4.2, using WGCNA to select a soft threshold value as a network construction threshold value; Step 4.3, constructing a network according to the soft threshold value 7 through the adjacency method of WGCNA to obtain an adjacency matrix of the gene network and the isoform network; Step 4.4, transforming the adjacency matrix of the gene network and the adjacency matrix of the isoform network into a topological overlap matrix TOM as a gene co-expression network and a variable splicing isoform co-expression network through the TOMsimilarity method of WGCNA; Step 4.5, defining the gene co-expression network as a source domain network and the corresponding variable splicing isoform co-expression network as a target domain network, and constructing a shared graph convolutional neural network model according to the source network and the target domain network; Step 5, constructing an adversarial domain adaptation model and combining the shared graph convolutional neural network model to form an initial prediction model, and training the initial prediction model by using the training set to obtain a target prediction model; Step 6, calculating the evaluation parameters of the target prediction model by using the test set and determining whether the target prediction model meets the preset requirements, if yes, executing step 7, if not, returning to step 5; Step 7, inputting the variable splicing isoform to be detected into the target prediction model to obtain a predicted pathogenic result.
2. The method of claim 1, wherein The step 2 specifically comprises: Step 2.1, downloading the human genome and its annotation file and sequence file from an initial database; Step 2.2, establishing an index through bioinformatics software according to the annotation file and the sequence file, and aligning the RNA-seq data through the bioinformatics software and the obtained index to convert the sequence FASTQ file into a BAM file; Step 2.3, establishing a genome index according to the annotation file and the sequence file; Step 2.4, calculating gene expression values and variable splicing isoform expression values according to the genome index and the BAM file; Step 2.5, extracting the corresponding FPKM values of the gene expression values and the variable splicing isoform expression values as the expression data of the disease-related genes and the corresponding variable splicing isoforms.
3. The method of claim 2, wherein The shared graph convolutional neural network model comprises two hidden layers, and the number of neurons of each layer is 128 and 16 respectively, and a Dropout layer is connected after each layer.
4. The method of claim 3, wherein The expression of the hidden layer is ； where X denotes an input feature matrix, A denotes an adjacency matrix, =A+In denotes an adjacency matrix with self-loops, In∈Rn×n denotes an identity matrix, and is a normalized adjacency matrix, is the output of the layer, = X, is a trainable parameter of the network, and σ(·) denotes an activation function.
5. The method of claim 4, wherein The adversarial domain adaptation model comprises a domain classifier, a source domain classifier and a target classifier.
6. The method of claim 5, wherein The step of training the initial prediction model by using the training set to obtain the target prediction model comprises: The preset learning rate and the iteration number are set, the training set is input into the initial prediction model for training until the evaluation index meets the preset condition or the iteration number is reached, and a target prediction model is obtained.
7. The method of claim 6, wherein An expression of the evaluation index is ； wherein is the number of samples predicted as positive and true positive, is the number of samples predicted as negative and true negative, is the total number of samples.