Prediction method of innate immune gene expression based on interpretable deep learning

Abstract

The invention discloses an interpretable deep learning-based innate immune gene expression prediction method, which comprises the following steps of: 1, collecting and preprocessing a DNA (Deoxyribose Nucleic Acid) promoter sequence and epigenetic modification data, and dividing a preprocessed data set into a training set and an independent test set so as to verify the prediction performance of a model; 2, constructing a human innate immune gene expression prediction network based on residual connection, a convolutional neural network and an attention mechanism; 3, performing model training on a plurality of cell lines to obtain an optimal human innate immune gene expression prediction model, and evaluating the prediction effect of the model on the independent test set; and 4, carrying out post interpretability analysis on the model through an integral gradient method, and identifying a key motif which is most likely to influence gene expression. According to the method, DNA sequence information and epigenetic modification information are extracted more effectively, and a new tool is provided for deeply understanding a regulation mechanism of an innate immune gene and developing individualized medical treatment.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically is a method for predicting innate immune gene expression based on interpretable deep learning. Background Art

[0002] The innate immune system is the earliest defense system in the human body to respond to pathogen invasion. It is a natural defense function formed by humans in the long-term process of phylogenetic development and evolution. It cooperates with the acquired immune system to protect the human body from pathogens. However, the innate immune system involves a large number of genes, signaling pathways and cell types. Its complexity and diversity make it very difficult to fully understand the expression and regulation mechanisms of these genes. Although high-throughput sequencing technologies (such as RNA-seq) have greatly promoted gene expression research, data analysis and interpretation still face challenges, such as the processing of noisy data and the detection of rare cell types. In the field of synthetic biology, the use of promoter sequences to predict gene expression is a key technology that can help scientists design and optimize gene circuits to achieve precise control of target gene expression. In addition, epigenetic modifications in the promoter region are important factors in determining gene expression, just like genetic material DNA. They regulate the transcriptional activity of genes through various mechanisms, thereby affecting the level of gene expression.

[0003] In previous related studies, a variety of methods for gene expression prediction have been developed, which are mainly divided into traditional machine learning models and deep learning models. Traditional machine learning models have fast calculation speed and are easy to interpret, but cannot handle nonlinear relationships well. For high-dimensional data in DNA sequences, the fitting effect is not ideal, and this type of model only uses the sequence characteristics of DNA without combining biological prior information; deep learning models have excellent performance in processing large-scale, high-dimensional data, and can combine biological prior knowledge for gene expression prediction. However, most deep learning models use epigenetic information of a single type and a small amount, and do not take into account the relationship between DNA sequences and epigenetic modifications. Secondly, existing deep learning models can only give a predictive judgment on whether the input sequence is the expression of human innate immune genes, but cannot explain the specific basis for making the judgment, which makes the model less interpretable. Summary of the invention

[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and proposes a prediction method for innate immune gene expression based on interpretable deep learning, in order to improve the accuracy of human innate immune gene expression prediction and improve the interpretability of regression prediction of human innate immune gene mRNA abundance through nucleotide-level importance assessment.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0006] The method for predicting innate immune gene expression based on interpretable deep learning of the present invention is characterized in that it is carried out according to the following steps:

[0007] Step 1: Collection and preprocessing of gene promoter sequence data and epigenetic modification data;

[0008] Step 1.1: Get After preprocessing, we obtained the preprocessed sequence training data set. ,in, express The gene promoter sequence training samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of gene promoter sequence training samples in ; ;

[0009] Step 1.2: Get After preprocessing, we obtained the preprocessed human innate immunity gene promoter sequence test dataset. ,in, express Middle Gene promoter sequence test samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of samples tested for gene promoter sequences in , is the length of each gene promoter sequence sample;

[0010] Step 1.3: Use the pre-trained Enformer model to obtain and Each gene promoter sequence sample in After preprocessing the epigenetic modification data on the cell lines, any of the Epigenetic training datasets on cell lines and Epigenetic testing datasets on cell lines ,in, express Middle epigenetic modification training samples, and , express Middle Epigenetic modification signals; express Middle Epigenetic modification test samples, and , express Middle Epigenetic modification signals; , , Represents the number of epigenetic modification signals in the epigenetic modification sample corresponding to each gene promoter sequence sample;

[0011] Step 1.4: In the The true label on the cell line is ,in, express In the True labels on cell lines;

[0012] Step 2: construct a human innate immune gene expression prediction network, which includes: a sequence feature extraction network module SFEN, an epigenetic feature extraction network module EFEN, and a regression prediction module;

[0013] Step 2.1, constructing a sequence feature extraction network module SFEN, including: a plurality of feature extraction blocks, and each feature extraction block includes: a dimensionality reduction transition block and a dense connection block;

[0014] Will Input into SFEN, and after being processed by several feature extraction blocks in turn, the first Sequence feature tensor ;

[0015] Step 2.2, construct the appearance feature extraction network module EFEN, including: several convolution blocks;

[0016] Will Input into EFEN, and after being processed by several convolution blocks in sequence, the first In the cell lines Apparent feature tensor ;

[0017] Step 2.3: and After splicing, it is input into the regression prediction module for processing to obtain the In the cell lines Prediction value of mRNA abundance of gene promoter sequence samples ;

[0018] Step 2.4: Use formula (1) to establish Mean squared error loss function for each cell line :

[0019] (1)

[0020] Step 3: and Input into the human innate immune gene expression prediction network, use the gradient descent method to iteratively train the human innate immune gene expression prediction network, and calculate the mean square error loss function To update the network parameters, when the number of training iterations reaches the set number or the mean square error loss function The training is stopped when convergence occurs, thereby obtaining the optimal human innate immune gene expression prediction model IGEP;

[0021] Step 4: and Input into IGEP for testing and get A test dataset of human innate immune gene promoter sequences in cell lines The optimal mRNA abundance prediction value ,in, Indicates In the cell lines Human innate immune gene promoter sequence test sample The optimal mRNA abundance prediction value of

[0022] Step 5: Use the integrated gradient method to calculate Importance score of nucleotides ;

[0023] Step 6: Based on ,from The length extracted is Sequence fragment set ,in, Indicates from The extracted sequence fragments; Sequence fragment set The total number of sequence fragments in

[0024] Step 7: Use dimensionality reduction method to Perform dimensionality reduction to obtain the reduced dimensionality representation ; Then use clustering algorithm to Perform clustering to obtain the position weight matrix of the sequence motif set ,in, Indicates The position weight matrix of sequence motifs, is the total number of sequence motifs in the sequence motif set;

[0025] Step 8: Obtain the position weight matrix of the known motif database ,in, express Middle The position weight matrix of known motifs, represents the number of sequence motifs in the known motif database;

[0026] Step 9: Calculate using formula (5) and Euclidean distance :

[0027] (5)

[0028] In formula (5), represents the row index corresponding to the nucleotide in the position weight matrix, and ; represents the number of columns of the position weight matrix, and ; and Corresponding to and In the Line The element value of the column;

[0029] Step 10: Use the sequence identification method to process the position weight matrix of all known motifs whose Euclidean distance is close to 0 to obtain a motif set that is highly correlated with gene expression. ={ },in, express The sequence motifs, express The total number of sequence motifs in .

[0030] The method for predicting innate immune gene expression based on interpretable deep learning described in the present invention is also characterized in that the dimension reduction transition block in step 2.1 comprises: a two-dimensional convolutional layer Conv2D and a maximum pooling layer MaxPooling2D; the dense connection block consists of It is composed of densely connected two-dimensional convolution blocks, each of which consists of a batch normalization layer BN, an activation function layer ReLU, and a convolution kernel. The two-dimensional convolution layer, batch normalization layer BN, activation function layer ReLU and convolution kernel are The two-dimensional convolutional layers are constructed in sequence;

[0031] Residual connections are established between the convolution blocks in step 2.2, and each convolution block consists of a one-dimensional convolution layer Conv1D, an activation function layer ReLU, a channel attention SE layer, and a maximum pooling layer;

[0032] The regression prediction module in step 2.3 is composed of a fully connected layer Dense, an activation function layer ReLU, a random inactivation layer Dropout, and an activation function layer Softplus in sequence.

[0033] Further, the step 5 is performed according to the following steps:

[0034] Step 5.1, select the length Gene promoter sequence samples as the baseline input sequence, and ,in, For the all-zero vectors;

[0035] Step 5.2: Calculate using formula (2) arrive The linear interpolation sequence of :

[0036] (2)

[0037] In formula (2), is the interpolation weight coefficient, and ;

[0038] Step 5.3: Calculate using formula (3) No. Quantity Gradient relative to IGEP output , thus obtaining Gradient relative to IGEP output ;

[0039] (3)

[0040] Step 5.4: Calculate using formula (4) Previous Importance score of each nucleotide , thus obtaining Importance score of nucleotides ;

[0041] (4).

[0042] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute any of the prediction methods, and the processor is configured to execute the program stored in the memory.

[0043] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes any step of the prediction method when executed by a processor.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention directly maps the human innate immune gene promoter sequence data and the epigenetic modification data of the promoter region to the predicted probability of whether the human innate immune gene is expressed through an end-to-end deep neural network. Compared with traditional biological methods, it saves a lot of time and labor costs, and can capture nonlinear patterns in sequences while processing large-scale genomic data and high-dimensional feature space, significantly improving the efficiency and accuracy of gene expression prediction.

[0046] 2. When constructing a human innate immune gene expression prediction network, the present invention intersperses the channel attention mechanism with the residual connected convolutional neural network, which allows the model to more accurately extract key feature information while avoiding information loss and redundancy, reduces the impact of unimportant feature information on model judgment, and improves the model's ability to extract key features and the generalization performance of the model.

[0047] 3. The present invention uses a pre-trained model to extract a variety of epigenetic modification information in the promoter region of human innate immune genes, and fully combines it with the DNA promoter sequence information, greatly improving the prediction performance of the human innate immune gene expression prediction network.

[0048] 4. The present invention uses an integrated gradient algorithm to output nucleotide-level importance scores in human innate immune gene promoter sequences, and then identifies key sequence motifs that affect gene expression, making the model highly interpretable. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a framework diagram of the human innate immune gene expression prediction network of the present invention;

[0050] Figure 2 It is a structural schematic diagram of the sequence feature extraction network module SFEN of the present invention;

[0051] Figure 3 It is a structural diagram of the appearance feature extraction network module EFEN of the present invention;

[0052] Figure 4 It is a comparison diagram of the Spearman correlation coefficient of the regression prediction network of the present invention. DETAILED DESCRIPTION

[0053] In this embodiment, a prediction method for innate immune gene expression based on interpretable deep learning is provided. The overall network framework is as follows: Figure 1 As shown, the following steps are included:

[0054] Step 1: Collection and preprocessing of gene promoter sequence data and epigenetic modification data;

[0055] Step 1.1: Get The human gene promoter sequence data that is not related to human innate immunity is only 60bp long and cannot be used as the input feature of the convolutional neural network. A human gene promoter sequence that is not related to human innate immunity is extended along the transcription start site to both sides, and the length of the extended sequence is L. In this example, the sequence extension length L is set to 2048bp; due to data redundancy, the extended sequence is deduplicated to obtain the preprocessed sequence training data set ,in, express The gene promoter sequence training samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of gene promoter sequence training samples in ; .

[0056] Step 1.2: Get Human gene promoter sequences related to human innate immunity are also extended to L in length and deduplicated to obtain the preprocessed human innate immunity gene promoter sequence test dataset. ,in, express Middle Gene promoter sequence test samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of samples tested for gene promoter sequences in , The length of each gene promoter sequence sample.

[0057] Step 1.3: Use the pre-trained Enformer model to obtain and Each gene promoter sequence sample in After preprocessing the epigenetic modification data on the cell lines, any of the Epigenetic training datasets on cell lines and Epigenetic testing datasets on cell lines ,in, express Middle epigenetic modification training samples, and , express Middle Epigenetic modification signals; express Middle Epigenetic modification test samples, and , express Middle Epigenetic modification signals; , , Indicates the number of epigenetic modification signals in the epigenetic modification sample corresponding to each gene promoter sequence sample. According to the principle that the more epigenetic modification signals, the better, three cell lines with the most epigenetic modification signals were selected, namely human chronic myeloid leukemia cells (K562), human liver cancer cells (HepG2), and human breast cancer cells (MCF-7), and epigenetic modification information was obtained from them. Therefore, in this example, the cell line type Set to 3; the number of genetic modification signals varies depending on the cell line. Set to 383, 241, and 113 respectively.

[0058] Step 1.4: In the The true label on the cell line is ,in, express In the True labels on cell lines.

[0059] Step 2: Construct a human innate immune gene expression prediction network, such as Figure 1 As shown, it includes: a sequence feature extraction network module SFEN, an appearance feature extraction network module EFEN, and a regression prediction module;

[0060] Step 2.1: Construct a sequence feature extraction network module SFEN to extract sequence features from the one-hot encoding data of gene promoter sequences. Considering that the DNA promoter sequence information encoding data is relatively sparse, in order to extract sequence features more comprehensively, a densely connected convolutional neural network is used, which has the advantages of alleviating gradient disappearance and strengthening feature propagation. Figure 2 As shown, it includes: several feature extraction blocks, and each feature extraction block includes: a dimensionality reduction transition block and a dense connection block. In this example, the number of feature extraction blocks is set to 3; wherein the dimensionality reduction transition block includes: a two-dimensional convolution layer Conv2D and a maximum pooling layer MaxPooling2D, which are used for feature extraction and dimensionality reduction. In this example, the two-dimensional convolution layer Conv2D has 128 The convolution kernel is used to extract low-level features and detect sequence motifs; the window size of the maximum pooling layer MaxPooling2D is , used to find the most significant activation signal in the sliding window; the densely connected block consists of It is composed of densely connected two-dimensional convolution blocks, each of which consists of a batch normalization layer BN, an activation function layer ReLU, and a convolution kernel. The two-dimensional convolution layer, batch normalization layer BN, activation function layer ReLU and convolution kernel are The two-dimensional convolutional layers are constructed in sequence; in this example, the number of densely connected two-dimensional convolutional blocks is Set to 5. Among them, the size is The 2D convolutional layer is used to reduce the number of cascaded channels to a fixed size of The two-dimensional convolution layer is used as a traditional convolution operation, and the batch normalization layer BN is used to reduce internal covariate shift and avoid overfitting.

[0061] Will Input into SFEN, and after being processed by several feature extraction blocks in turn, the first Sequence feature tensor .

[0062] Step 2.2: Construct an epigenetic feature extraction network module EFEN to extract epigenetic features from epigenetic modification signals in the promoter sequence region, such as Figure 3As shown, it includes: several convolution blocks; wherein residual connections are established between the convolution blocks to avoid gradient disappearance during training, and each convolution block is composed of a one-dimensional convolution layer Conv1D, an activation function layer ReLU, a channel attention SE layer and a maximum pooling layer; in this example, the number of convolution blocks is set to 3, and the one-dimensional convolution layer in each convolution block has 256 convolution kernels of size 5.

[0063] Since epigenetic modification information is rich and diverse, and the correlation between features is not obvious, directly using feature channels may lead to information loss or redundancy. The attention mechanism can be added after each convolution operation to more accurately extract key features and reduce the impact of unimportant features on the final result.

[0064] Will Input into EFEN, and after being processed by several convolution blocks in sequence, the first In the cell lines Apparent feature tensor .

[0065] Step 2.3, construct a regression prediction module, which is composed of a fully connected layer Dense, an activation function layer ReLU, a random inactivation layer Dropout, and an activation function layer Softplus in sequence;

[0066] Will and After splicing, it is input into the regression prediction module for processing to obtain the In the cell lines Prediction value of mRNA abundance of gene promoter sequence samples .

[0067] Step 2.4: Use formula (1) to establish Mean squared error loss function for each cell line :

[0068] (1)

[0069] Step 3: and Input into the human innate immune gene expression prediction network, use the gradient descent method to iteratively train the human innate immune gene expression prediction network, and calculate the mean square error loss function To update the network parameters, when the number of training iterations reaches the set number or the mean square error loss function The training is stopped when convergence occurs, thereby obtaining the optimal human innate immune gene expression prediction model IGEP. In this example, the optimizer of the prediction network uses RMSprop, the learning rate is set to 1e-3, the batch size is set to 32, and the number of iterations is set to 100.

[0070] like Figure 4 As shown, three other gene expression prediction deep learning algorithms reported in recent journal literature were used to compare their performance with the above-mentioned human innate immune gene expression prediction network of the present invention. Five-fold cross validation was performed on three cell line data sets, and the comparison indicator was the Spearman correlation coefficient SCC. The results show that the human innate immune gene expression prediction network of the present invention achieved the best prediction performance.

[0071] Step 4: and Input into IGEP for testing and get A test dataset of human innate immune gene promoter sequences in cell lines The optimal mRNA abundance prediction value ,in, Indicates In the cell lines Human innate immune gene promoter sequence test sample The optimal mRNA abundance prediction value.

[0072] Step 5: Use the integrated gradient method to calculate Importance score of nucleotides , perform post hoc interpretability analysis on the model.

[0073] Step 5.1, select the length Gene promoter sequence samples as the baseline input sequence, and ,in, For the all-zero vectors;

[0074] Step 5.2: Calculate using formula (2) arrive The linear interpolation sequence of :

[0075] (2)

[0076] In formula (2), is the interpolation weight coefficient, and ;

[0077] By changing Generate a series of sequences , these sequences gradually change from Transition to ,exist and Linear interpolation is performed between .

[0078] Step 5.3: Calculate using formula (3) No. Quantity Gradient relative to IGEP output , thus obtaining Gradient relative to IGEP output ;

[0079]

[0080] This gradient represents the model's interpolation sequence Relative to the input test sample sensitivity.

[0081] Step 5.4: Calculate using formula (4) Previous Importance score of each nucleotide , thus obtaining Importance score of nucleotides ;

[0082]

[0083] By obtaining the importance score of each nucleotide in the entire sequence, the contribution of the nucleotide to the prediction result can be reflected.

[0084] Step 6: Based on ,from The length extracted is Sequence fragment set ,in, Indicates from The extracted sequence fragments; Sequence fragment set The total number of sequence fragments in ; in this example, the length of the sequence fragment Set to 5.

[0085] Step 7: Since the number of extracted sequence fragments is relatively large, the dimensionality reduction method is used right Perform dimensionality reduction to obtain the reduced dimensionality representation , and then use the clustering algorithm For the sequence fragment after dimensionality reduction Perform clustering to obtain the position weight matrix of the sequence motif set ,in Indicates The position weight matrix of sequence motifs, is the total number of sequence motifs in the sequence motif set.

[0086] Step 8: Obtain the position weight matrix of the known motif database ,in express Middle The position weight matrix of known motifs, It indicates the number of sequence motifs in the known motif database. In this example, the known motif database is the JASPAR database.

[0087] Step 9: Calculate using formula (5) and Euclidean distance :

[0088] (5)

[0089] In formula (5), represents the row index corresponding to the nucleotide in the position weight matrix, and ; represents the number of columns of the position weight matrix, and ; and Corresponding to and In the Row, No. Column element value;

[0090] Euclidean distance can be used to measure the difference between two sets of values. The smaller the value, the more similar the two matrices are. In this example, The smaller the value of, the greater the position weight matrix and The smaller the difference at each base position, the more it means that the sequence motif found by the present invention highly overlaps with the known sequence motif.

[0091] Step 10: Use the sequence identification method to visualize the position weight matrix of all known motifs with Euclidean distance close to 0 to obtain a set of motifs highly correlated with gene expression. ={ },in, express The sequence motifs, express The total number of sequence motifs in .

[0092] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0093] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A prediction method for innate immune gene expression based on interpretable deep learning, characterized in that: The steps are as follows: Step 1: Collection and preprocessing of gene promoter sequence data and epigenetic modification data; Step 1.1: Get After preprocessing, we obtained the preprocessed sequence training data set. ,in, express The gene promoter sequence training samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of gene promoter sequence training samples in ; ; Step 1.2: Get After preprocessing, we obtained the preprocessed human innate immunity gene promoter sequence test dataset. ,in, express Middle Gene promoter sequence test samples, and , express Middle The one-hot encoded vector of nucleotides, , express The total number of samples tested for gene promoter sequences in , is the length of each gene promoter sequence sample; Step 1.3: Use the pre-trained Enformer model to obtain and Each gene promoter sequence sample in After preprocessing the epigenetic modification data on the cell lines, any of the Epigenetic training datasets on cell lines and Epigenetic testing datasets on cell lines ,in, express Middle epigenetic modification training samples, and , express Middle Epigenetic modification signals; express Middle Epigenetic modification test samples, and , express Middle Epigenetic modification signals; , , Represents the number of epigenetic modification signals in the epigenetic modification sample corresponding to each gene promoter sequence sample; Step 1.4, order In the The true label on the cell line is ,in, express In the True labels on cell lines; Step 2: construct a human innate immune gene expression prediction network, which includes: a sequence feature extraction network module SFEN, an epigenetic feature extraction network module EFEN, and a regression prediction module; Step 2.1, constructing a sequence feature extraction network module SFEN, including: a plurality of feature extraction blocks, and each feature extraction block includes: a dimensionality reduction transition block and a dense connection block; Will Input into SFEN, and after being processed by several feature extraction blocks in turn, the first Sequence feature tensor ; Step 2.2, construct the appearance feature extraction network module EFEN, including: several convolution blocks; Will Input into EFEN, and after being processed by several convolution blocks in sequence, the first In the cell lines Apparent feature tensor ; Step 2.3: and After splicing, it is input into the regression prediction module for processing to obtain the In the cell lines Prediction value of mRNA abundance of gene promoter sequence samples ; Step 2.4: Use formula (1) to establish Mean squared error loss function for each cell line : (1) Step 3: and Input into the human innate immune gene expression prediction network, use the gradient descent method to iteratively train the human innate immune gene expression prediction network, and calculate the mean square error loss function To update the network parameters, when the number of training iterations reaches the set number or the mean square error loss function The training is stopped when convergence occurs, thereby obtaining the optimal human innate immune gene expression prediction model IGEP; Step 4: and Input into IGEP for testing and get A test dataset of human innate immune gene promoter sequences in cell lines The optimal mRNA abundance prediction value ,in, Indicates In the cell lines Human innate immune gene promoter sequence test sample The optimal mRNA abundance prediction value of Step 5: Use the integrated gradient method to calculate Importance score of nucleotides ; Step 6: Based on ,from The length extracted is Sequence fragment set ,in, Indicates from The extracted sequence fragments; Sequence fragment set The total number of sequence fragments in Step 7: Use dimensionality reduction method to Perform dimensionality reduction to obtain the reduced dimensionality representation ; Then use clustering algorithm to Perform clustering to obtain the position weight matrix of the sequence motif set ,in, Indicates The position weight matrix of sequence motifs, is the total number of sequence motifs in the sequence motif set; Step 8: Obtain the position weight matrix of the known motif database ,in, express Middle The position weight matrix of known motifs, represents the number of sequence motifs in the known motif database; Step 9: Calculate using formula (5) and Euclidean distance : (5) In formula (5), represents the row index corresponding to the nucleotide in the position weight matrix, and ; represents the number of columns of the position weight matrix, and ; and Corresponding to and In the Line The element value of the column; Step 10: Use the sequence identification method to process the position weight matrix of all known motifs whose Euclidean distance is close to 0 to obtain a motif set that is highly correlated with gene expression. ={ },in, express The sequence motifs, express The total number of sequence motifs in .

2. The method for predicting innate immune gene expression based on interpretable deep learning according to claim 1, characterized in that: The dimensionality reduction transition block in step 2.1 includes: a two-dimensional convolution layer Conv2D and a maximum pooling layer MaxPooling2D; the dense connection block consists of It is composed of densely connected two-dimensional convolution blocks, each of which consists of a batch normalization layer BN, an activation function layer ReLU, and a convolution kernel. The two-dimensional convolution layer, batch normalization layer BN, activation function layer ReLU and convolution kernel are The two-dimensional convolutional layers are constructed in sequence; Residual connections are established between the convolution blocks in step 2.2, and each convolution block consists of a one-dimensional convolution layer Conv1D, an activation function layer ReLU, a channel attention SE layer, and a maximum pooling layer; The regression prediction module in step 2.3 is composed of a fully connected layer Dense, an activation function layer ReLU, a random inactivation layer Dropout, and an activation function layer Softplus in sequence.

3. The method for predicting innate immune gene expression based on interpretable deep learning according to claim 2, characterized in that: The step 5 is performed as follows: Step 5.1, select the length Gene promoter sequence samples as the baseline input sequence, and ,in, For the all-zero vectors; Step 5.2: Calculate using formula (2) arrive The linear interpolation sequence of : (2) In formula (2), is the interpolation weight coefficient, and ; Step 5.3: Calculate using formula (3) No. Quantity Gradient relative to IGEP output , thus obtaining Gradient relative to IGEP output ; (3) Step 5.4: Calculate using formula (4) Previous Importance score of each nucleotide , thus obtaining Importance score of nucleotides ; (4)。 4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the prediction method according to any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the prediction method according to any one of claims 1 to 3 are executed.

Images