A method for identifying rice blast pathogenicity genes suitable for small sample and high-dimensional data

Abstract

The present invention belongs to the field of biometric identification technology, and relates to a method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data, wherein the method processes the collected data to obtain the expression amount of the sample corresponding to each gene, then performs differential analysis, identifies the differentially expressed genes, and constructs a PPI network based on the differentially expressed genes to determine the key genes related to rice blast; finally, a CGAN-SAE deep learning model is constructed, and the CGAN module is trained using the determined key genes, so that the CGAN module can generate simulated gene sample data that is difficult to distinguish from real data; the SAE module is then trained with simulated gene samples and real samples, so that the SAE module can capture global features, and the extracted features are further processed by PCA to identify the core genes in the rice blast infection process. The method solves the problem of gene screening in rice data sets with a small number of samples and high dimensions.

Description

Technical Field

[0001] The present invention belongs to the technical field of biometric identification, and in particular relates to a method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data. Background Art

[0002] Rice is an important food crop, mainly divided into two types: indica and japonica. It plays an important role in the global population, especially in Asia. As the world's population grows, the demand for increasing rice production is increasing. Therefore, research on high yield and disease resistance of rice has become an important field of agricultural science research.

[0003] Rice blast is a disease that is particularly harmful to rice and is caused by the fungus Magnaporthe oryzae. "The fungus undergoes extensive developmental changes that allow it to invade plant cells, establish complex infection structures, and proliferate within host cells without causing obvious disease symptoms." Not only targeting rice, the rice blast fungus has spread to plants such as wheat. "This is due to host jumps from strains of the fungus that infect pastures, indicating that M. oryzae has the ability to cause disease on a wide range of grass hosts and then perform host jumps under the strong selection pressure of widespread agricultural cultivation of cereals." In order to inhibit the spread of rice blast, it is particularly important to study the pathogenesis and infection mode of rice blast.

[0004] At present, the prevention of rice blast is mainly considered from the following aspects: in terms of nutrient management, the use of soil with high biological activity and good soil fertility can effectively prevent infection; in terms of agricultural measures, disease-resistant rice varieties are selected, and different rice varieties have different resistance to rice blast. Disease-resistant varieties generally have strong resistance to pathogens. For example, some varieties can prevent the invasion of pathogens, and some can inhibit the spread of pathogens in plants; in terms of chemical control, chemical agents, especially fungicides, are the simplest and most common method for preventing and controlling rice blast, but they also pose a great threat to human health and the environment.

[0005] It is well known that breeding disease-resistant varieties with resistance genes is the most effective and economical strategy to control the disease. The discovery of disease-resistant genes is the basis of disease-resistant breeding. Scientists use methods such as genetic analysis, molecular marker-assisted selection (MAS) and genome-wide association studies (GWAS) to find new disease-resistant genes.

[0006] For example, the R gene is a gene that can resist rice blast and can play an important role in molecular breeding. Some scholars have proposed strategies for the use of the rice blast R gene in molecular breeding, such as accurately understanding the use of the R gene in major modern rice varieties, cultivating core resistance germplasm with excellent comprehensive traits, and screening and using broad-spectrum durable resistance gene combinations. This method can not only speed up the breeding process, but also achieve gene transfer between species, broadening the source of disease-resistant genes.

[0007] In recent years, machine learning and deep learning have played an increasingly important role in the process of resistance gene screening. Machine learning algorithms can process large amounts of genomic data and identify patterns and features related to resistance. Deep learning models can capture complex relationships between high-dimensional data, which is particularly important for analyzing the interaction between genes and the environment. In the study of exploring key genes for rice blast, more and more researchers have combined machine learning models in order to further explore potential pathogenic factors. For example, researchers have applied a variety of machine learning methods such as random forests, support vector machines (SVMs), and Bayesian classifiers to screen and analyze genes. In addition, some researchers have also combined deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), to capture more complex gene expression patterns and time series data. These models can identify potential features in high-dimensional data, thereby improving the accuracy and efficiency of disease resistance gene screening.

[0008] At present, the research on rice blast mainly focuses on single-omics data, which is not enough to capture the complex cross-kingdom regulatory interactions between rice blast fungus and rice. In order to solve this limitation, researchers have begun to study machine learning models suitable for multi-omics. For example: Zhang et al. proposed a model algorithm called GERWR based on random walk of graph machine learning, which explored the pathogenic factors of rice blast from the perspective of multi-omics data analysis and revealed the intrinsic connection between different omics pathogenic factors. Zhao et al. proposed a method called weighted gene autoencoder multi-omics relationship prediction (WGAEMRP), which combines weighted gene co-expression network analysis (WGCNA) and graph autoencoder to predict the relationship between multi-omics data of rice blast fungus and rice.

[0009] In the study of cross-border regulation between rice blast fungus and rice, some researchers have put forward some preliminary insights and results. Zhao et al. constructed the sRNA regulatory network of rice infected by rice blast fungus based on multi-omics data and mined the key infection factors of rice blast fungus sRNA and protein. Zhang et al. trained a SVM model that can be used to predict differentially expressed sRNA, which has reference significance for the prediction of differentially expressed sRNA of homologous species of rice blast fungus and provides convenience for future research on rice blast fungus. Although key genes can be screened using existing models, most of the existing methods use machine learning models for gene mining, which has weak processing capabilities for high-dimensional data, especially in small sample data, has low capture of nonlinear relationships, and has average generalization ability. It is particularly prone to overfitting during model training, resulting in incorrect prediction results. Summary of the invention

[0010] The purpose of the present invention is to provide a method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data. The method uses gene chip expression data from GEO, finds out 891 genes whose expression levels change significantly before and after rice blast infection through differential analysis, constructs a PPI interaction network on the STRING database through differentially expressed genes, and screens out 30 key genes through the interaction network; finally, the key genes are further screened according to the constructed CGAN-SAE-based deep learning model, and finally 5 core genes are screened out. After verification, it was found that the expression levels of these 5 genes were very different from those of the normal group, indicating that these 5 genes are likely to be the core genes that rice blast affects rice gene expression. By deeply analyzing the regulatory network of rice blast pathogenicity genes, it is expected to provide a new theoretical basis for disease-resistant breeding strategies.

[0011] To achieve the above object, the present invention adopts the following technical solution:

[0012] A method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data includes the following steps:

[0013] Step 1. Data collection;

[0014] Step 2. Process the data collected in step 1 to obtain the expression level of each gene corresponding to the sample;

[0015] Step 3. Perform differential analysis on the data processed in step 2 to identify differentially expressed genes in the data;

[0016] Step 4. Construction and analysis of protein-protein interaction network;

[0017] A protein-protein interaction network was constructed based on the differentially expressed genes determined in step 3, and then the protein-protein interaction network was reconstructed using Cytoscape software integrated with the CytoHubba plug-in. The reconstructed protein-protein interaction network was analyzed to determine the key genes related to rice blast disease;

[0018] Step 5. Construct a CGAN-SAE deep learning model, and use the key genes related to rice blast determined in step 4 to train the CGAN module, so that the trained CGAN module can generate simulated gene sample data that is difficult to distinguish from real data; then input the newly generated simulated gene samples and the real samples obtained in step 1 into the SAE module for training, so that the trained SAE module can capture global features when performing feature extraction, and the extracted features are further processed by principal component analysis to reduce the redundancy of high-dimensional features, and at the same time, the features are scored for importance, so as to identify the core genes in the rice blast infection process;

[0019] The CGAN module includes a generator and a discriminator; wherein the network structure of the generator includes a first block layer, a second block layer, a third block layer, a fourth block layer, a linear layer, and a tanh activation function, which are arranged in sequence. The first block layer is a linear layer and is not regularized; the second block layer, the third block layer, and the fourth block layer are all linear layers and regularized, the fourth block layer is followed by a linear layer, and the tanh activation function is used after the linear layer; the network structure of the discriminator includes a first linear layer, a leaky relu activation function, a second linear layer, a first dropout layer, a leaky relu activation function, a third linear layer, a second dropout layer, a leaky relu activation function, and a fourth linear layer, which are arranged in sequence; wherein the dropout rate of the first dropout layer and the second dropout layer is 0.4;

[0020] The value function of the CGAN module is:

[0021] ;

[0022] In the formula, Maximize V(G,D) for the discriminator D; Minimize V(G,D) for generator G; is the discriminator for the real data x The predicted value of , D(G(z|y)) is the given condition y In the case of z The predicted values ​​of the generated samples; Indicates that for the real data distribution The real data extracted from x Expected value; For a random noise distribution P z The noise vector extracted from (z) z Expected value;

[0023] A gated axial attention layer is added to both the encoder and decoder networks of the SAE module. The network structure of the SAE encoder includes a first linear layer, a relu activation function, a second linear layer, and a gated axial attention layer arranged in sequence; the network structure of the decoder is the same as that of the encoder.

[0024] As a preferred embodiment of the present invention, during the differential analysis in step 3, the "limma" package of the R programming language is used as a tool for differential analysis, and a linear model is constructed to fit the expression level of the gene, thereby identifying the differentially expressed genes in the data.

[0025] As a preferred embodiment of the present invention, when constructing the protein-protein interaction network in step 4, the STRING database is used, and the differentially expressed genes obtained after the differential analysis are input into the STRING database.

[0026] As a preferred embodiment of the present invention, the loss function of the SAE module is L By reconstruction error and sparsity constraints The composition is expressed as:

[0027] ;

[0028] ;

[0029] ;

[0030] in, represents the weight of adjusting the sparsity constraint; n represents the number of samples; Represents the sample of the i-th input; represents the i-th reconstructed sample; m refers to the number of neurons in the hidden layer, refers to the preset sparse target, refers to the actual activation ratio of the jth neuron, The function represents the KL divergence.

[0031] Advantages and beneficial effects of the present invention:

[0032] (1) In view of the problem that current gene screening technology is difficult to adapt to small samples, high dimensions, and nonlinearity, the present invention proposes a solution based on the CGAN-SAE deep learning model. In order to adapt the CGAN module to the data set of the present invention, the network layer is adjusted so that the CGAN module can generate forged gene samples that are difficult to distinguish from real data, thereby increasing the diversity of the data set, expanding the training data set, alleviating the problem of insufficient samples in agricultural gene research, and helping the model learn a wider range of feature patterns, thereby improving the performance of the model on unseen data. In addition, more synthetic samples can help prevent the model from overfitting to limited real samples, making the model more robust and universal, and solving the problem of gene screening in rice data sets with a small number of samples and high dimensions. This method can achieve efficient modeling in small sample scenarios, providing strong support for the research of scarce disease-resistant resources.

[0033] (2) The present invention combines the advantages of CGAN and SAE, improves SAE, introduces a gated axial attention mechanism, and enhances SAE's ability to capture global features. The extracted features are further processed by principal component analysis (PCA) to reduce the redundancy of high-dimensional features. At the same time, the features are scored for importance. This method can quickly identify the core genes in the rice blast infection process.

[0034] (3) The SAE module of the present invention is combined with PCA to effectively improve the efficiency of feature extraction and data dimensionality reduction. It can automatically extract key features from high-dimensional genomic data without the need for manual feature design, greatly improving the efficiency of data dimensionality reduction. The features after dimensionality reduction retain information that has important contributions to the phenotype and significantly reduce data redundancy and noise interference.

[0035] (4) The core genes screened out in this invention provide important clues for future rice disease resistance breeding. The specific action mechanism of these genes in the rice immune system can be further explored, especially by verifying the function of these genes in rice resistance to rice blast through gene editing technology (such as CRISPR-Cas9). This method not only provides new potential targets for rice disease resistance breeding, but also demonstrates the broad prospects of combining bioinformatics with machine learning methods for plant disease gene screening.

[0036] (5) The successful case of applying CGAN in agricultural gene research can inspire more research on how to use generative models to solve bioinformatics problems, and may give rise to a series of new tools and technologies specifically for agricultural gene data analysis. Based on this approach, researchers can use CGAN to simulate the expression patterns of key genes in known disease-resistant varieties and generate a large number of potential disease-resistant samples in a short period of time. By generating a large number of potential disease-resistant samples, researchers can quickly screen and verify new disease-resistant candidate genes, greatly shortening the time cycle from laboratory to field application and accelerating the research and development cycle of new disease-resistant varieties. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A flow chart of the method for identifying rice blast pathogenicity genes provided by the present invention;

[0038] Figure 2 It is a framework diagram of the CGAN-SAE deep learning model of the present invention;

[0039] Figure 3 It is a volcano diagram of the difference analysis results of the present invention;

[0040] Figure 4 Gene heat map of the top 20 genes whose expression levels were up-regulated or down-regulated in differential analysis;

[0041] Figure 5 Schematic diagram of the enrichment analysis results of the present invention; wherein, a is the top 10 key biological processes related to rice infection with rice blast in GO analysis; b is the cell composition significantly enriched in GO analysis; c is the molecular function significantly enriched in GO analysis; d is the KEGG analysis result;

[0042] Figure 6 The interaction network constructed for the present invention; wherein a is the interaction network constructed using differentially expressed genes (hiding nodes without connections), and b is the interaction network of key genes calculated by the cytohubba plug-in after beautification;

[0043] Figure 7 Box plot comparison of the expression levels of the 5 genes screened in the present invention in the normal group and the infected group; wherein a is Os01g0191800, b is Os01g0169800, c is Os03g0162200, d is Os11g0455500, and e is Os01g0927000;

[0044] Figure 8 The effect of the number of training times on the model effect. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the technical solution and advantages of the present invention, the present application is described in detail below in conjunction with the accompanying drawings, but it is not intended to limit the protection scope of the present invention.

[0046] like Figure 1 , Figure 2 As shown, the present invention provides a method for identifying rice blast pathogenicity genes applicable to small sample high-dimensional data, the method comprising the following steps:

[0047] Step 1. Data collection:

[0048] The rice gene chip expression data GSE39635 of rice hydroponically cultivated and 48 hours after infection with rice blast fungus were obtained from the GEO database. There are four different types of gene chip expression data. This example selects two types of all expression data for wild rice leaves and sets them as normal group (Normal) and infected group (Infected);

[0049] Step 2. Data processing;

[0050] The expression matrix obtained from GSE39635 was processed, all row names were converted into RAP IDs, and duplicate values ​​were removed, so that the expression level of each gene corresponding to the sample was obtained. In order to perform differential analysis, this example grouped and marked each sample to construct a grouped expression matrix.

[0051] In this embodiment, after the data processing in step 2, there are 29923 pieces of gene expression information in the expression matrix.

[0052] Step 3. Difference Analysis;

[0053] Differential analysis refers to a statistical method used in bioinformatics to identify whether there are significant differences in gene expression levels under different conditions or groups (such as healthy control group and disease group). Differentially expressed genes (DEGs) are genes whose expression levels have changed significantly under these different biological conditions. This type of analysis is of great significance for understanding gene regulatory networks, exploring disease mechanisms, and finding biomarkers.

[0054] For the microarray data constructed in step 2, this embodiment uses the "limma" package of the R programming language as a tool for differential analysis, and constructs a linear model to fit the expression level of the gene, thereby identifying the differentially expressed genes therein.

[0055] In this example, the gene expression profiles between the rice blast infection group and the normal group were carefully compared by using the "limma" package of the R programming language. After differential analysis, the LogFC (log fold change)>1 and adjustedP<0.05 were set, and 891 DEGs were finally obtained after screening, of which 551 genes were up-regulated and 340 genes were down-regulated.

[0056] In order to make these findings intuitively presented, the present invention draws a volcano diagram for the results of the difference analysis ( Figure 3 , and created a gene heat map for the top 20 genes with up-regulated and down-regulated expression ( Figure 4 ).

[0057] Enrichment Analysis:

[0058] Enrichment analysis is a statistical method to determine whether a group of differentially expressed genes are over-represented in a specific biological function, cellular component or biological process. Through enrichment analysis, we can find out the biological processes or functions in which these differentially expressed genes are involved, thereby explaining the possible biological impact of the observed expression changes.

[0059] This example uses bioinformatics tools and databases that are widely recognized by the scientific community to conduct a comprehensive analysis of differentially expressed genes (DEGs). Specifically, the bioinformatics tool DAVID (https: / / david.ncifcrf.gov) was used to perform Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis. Through rigorous GO analysis of differentially expressed genes, the goal is to reveal the basic biological processes and functions that these genes contribute to in the context of rice blast occurrence and progression. This detailed analysis can enable people to have a more comprehensive understanding of the molecular complexity behind rice blast infection.

[0060] Figure 5The enrichment analysis results of the present invention are illustrated; wherein, a illustrates the top 10 key biological processes (BPs) related to rice infection with rice blast in GO analysis; in these enriched BPs, some related processes can be found, such as hydrogen peroxide catabolic process, defense response, regulation of DNA-templated transcription, response to oxidative stress, protein phosphorylation, diterpene phytoalexin biosynthetic process, one-carbon metabolic process, response to other organisms, regulation of protein serine, threonine phosphatase activity, and transmembrane transport. Figure 5 b in the figure shows the significantly enriched cellular components (CC) in the GO analysis. In the analysis results, four significantly enriched CCs were found, including extracellular region, membrane, plasma membrane, and spindle midzone. These cellular structures may play an important role in the progression of rice blast disease. Figure 5The c in the figure indicates the significantly enriched molecular functions (MF) in the GO analysis. GO analysis shows that there are 10 significantly enriched MFs, covering DNA-binding transcription factor activity, heme binding, peroxidase activity, lactoperoxidase activity, monooxygenase activity, sequence-specific DNA binding, carbohydrate binding, oxidoreductase activity, acting on paired donors, withincorporation or reduction of molecular oxygen, protein serine / threonine kinase activity, and abscisic acid binding. These findings reveal the molecular interactions and signaling pathways related to rice blast, paving the way for further research and targeted therapeutic interventions.

[0061] KEGG enrichment analysis is a key technique in bioinformatics, which is used to identify whether a group of genes (such as differentially expressed genes) are significantly enriched in known metabolic pathways, signaling pathways or other biological processes. This analysis is indispensable for understanding the interaction between specific cellular processes and molecules during rice blast infection. It can serve as an important tool to explain the potential perturbations that occur during the pathogenesis of rice blast, and provide important clues for identifying potential pathogenic factors and therapeutic targets.

[0062] Figure 5The d in the figure is the result of KEGG analysis. Seven enriched KEGG pathways were identified through KEGG analysis, including diterpenoid biosynthesis, phenylpropanoid biosynthesis, biosynthesis of secondarymetabolites, metabolic pathways, tryptophan metabolism, MAPK signaling pathway – plant, and plant hormone signal transduction. Information from these pathways can provide a better understanding of how plants respond to pathogen invasion and which genes and metabolic processes are involved in this response. These findings not only help increase our understanding of plant-pathogen interactions, but also provide potential targets for the development of new disease resistance strategies, such as improving rice disease resistance by enhancing specific defense pathways. In addition, this information may also be used to guide crop breeding and select or create rice varieties with stronger disease resistance.

[0063] Step 4. Construction and analysis of protein-protein interaction network (PPI Network Analysis);

[0064] Constructing protein-protein interaction (PPI) networks through differentially expressed genes is an important bioinformatics task, which can help us better understand the functions and mechanisms behind changes in gene expression.

[0065] In order to build a reliable PPI network, this embodiment uses the STRING database (https: / / cn.string-db.org / ), and successfully constructs a PPI network after inputting the differentially expressed genes obtained after differential analysis into the STRING database. In order to more intuitively understand the interconnection between genes and visualize the complex PPI network, this embodiment uses Cytoscape_v3.10.2 software. Cytoscape software is mainly used to visualize complex networks and their interactions with tabular data. In addition, this embodiment integrates the CytoHubba plug-in in the Cytoscape software. CytoHubba is a tool specifically used to identify key nodes (hub nodes) in complex networks. These nodes usually play an important role in the network (may play the role of key regulatory factors or potential drug targets in disease mechanisms), which is crucial to understanding network functions.

[0066] In this example, the PPI network derived from the STRING database was successfully reconstructed using Cytoscape software. In addition to beautifying and enriching the network structure, key genes in the PPI network (some possible key genes in the process of rice being infected by rice blast fungus) were obtained, providing important clues for exploring key factors in the PPI network related to rice blast.

[0067] Specifically, the graphical interaction network (PPI network) successfully constructed in this embodiment contains 789 nodes and 67 edges. Since it contains many unconnected nodes, this embodiment sets an option to hide nodes without connections. The final network diagram is as follows: Figure 6 In addition, the present invention imports the node and edge files into Cytoscape, and uses the cytohubba plug-in in the Cytoscape software. Figure 6 Figure b shows the interaction network of key genes (red circles) calculated by the cytohubba plug-in after beautification. Through this network, 30 key genes related to rice blast were finally identified.

[0068] Step 5. Construct a CGAN-SAE deep learning model (Deep Learning Model), and use the key genes related to rice blast determined in step 4 to train the CGAN module, so that the trained CGAN module can generate simulated gene sample data (fake samples) that are difficult to distinguish from real data. Then, the newly generated simulated gene samples and real sample data are input into the SAE module for training, so that the trained SAE module can capture global features when performing feature extraction. The final extracted features are further processed through principal component analysis to reduce the redundancy of high-dimensional features. At the same time, the features are scored for importance to identify the core genes in the rice blast infection process.

[0069] In order to further study the characteristics of key factors of rice during rice blast infection, this example constructs a deep learning model (such as Figure 2 As shown in Figure 2, this model is designed to handle complex tasks with a large number of features and a limited number of samples.

[0070] CGAN is an extended version of Generative Adversarial Networks (GAN). GAN consists of two core components: the generator and the discriminator. On this basis, CGAN adds additional conditional variables, so that the generator and the discriminator can use these additional information during the training process, thereby improving the quality and diversity of the generated data. In this embodiment, the generator is used to generate fake data that is difficult to distinguish from real data, while the discriminator strives to distinguish between real data and generated data, forming a game between the two. The adversarial game process of the conditional generative adversarial network CGAN is used to generate forged samples, reducing the risk of overfitting of the machine learning model on a small sample high-dimensional data set.

[0071] In order to adapt the CGAN module to the data set of this embodiment, forged gene samples are generated (the existing CGAN module is mainly used to generate forged images), and the network layers and hyperparameters are adjusted; the network structure of the generator includes a first block layer, a second block layer, a third block layer, a fourth block layer, a Linear layer, and a tanh activation function, which are arranged in sequence. The first block layer is a linear layer without regularization; the second block layer, the third block layer, and the fourth block layer are all linear layers and regularized. The fourth block layer is followed by a Linear layer, and the tanh activation function is used after the Linear layer; the network structure of the discriminator includes a first Linear layer, a LeakyReLU activation function, a second Linear layer, a first Dropout layer, a LeakyReLU activation function, a third Linear layer, a second Dropout layer, a LeakyReLU activation function, and a fourth Linear layer, which are arranged in sequence; wherein the dropout rates of the first Dropout layer and the second Dropout layer are 0.4.

[0072] Specifically, the generator: self.model = nn.Sequential(

[0073] *block(latent_dim + n_classes, 128, normalize=False), / / noise dimension plus number of categories, no regularization, all following blocks are regularized

[0074] *block(128, 256),

[0075] *block(256, 512),

[0076] *block(512, 1024),

[0077] nn.Linear(1024, int(np.prod(img_shape))), / / Calculate the dimension product; for example, if img_shape is (3, 64, 64), then np.prod(img_shape) will be calculated as 3×64×64=12288

[0078] nn.Tanh() )

[0080] Discriminator

[0081] self.model = nn.Sequential(

[0082] nn.Linear(n_classes + int(np.prod(img_shape)), 512),

[0083] nn.LeakyReLU(0.2, inplace=True), / / Activation function selection.LeakyReLU

[0084] nn.Linear(512, 512),

[0085] nn.Dropout(0.4),

[0086] nn.LeakyReLU(0.2, inplace=True),

[0087] nn.Linear(512, 512),

[0088] nn.Dropout(0.4),

[0089] nn.LeakyReLU(0.2, inplace=True),

[0090] nn.Linear(512, 1), )

[0092] The training learning rate lr=0.0002 and the noise dimension latent_dim=50.

[0093] In this embodiment, the main formula (value function) of CGAN is shown in formula (1). In this embodiment, y refers to the conditional information, that is, the sample label (the uninfected label is recorded as 0 and the infected label is recorded as 1); x This is our data (30 genes from PPI network analysis); is the input noise; G and D are the generator and discriminator respectively, which are nonlinear mapping functions;

[0094] (1)

[0095] In the formula, It is to maximize V(G,D) for the discriminator D, that is, we hope that the discriminator can distinguish between real samples and forged samples as accurately as possible; is to minimize V(G,D) for the generator G, that is, it is hoped that the generator can generate sufficiently realistic fake samples so that the discriminator can hardly distinguish the difference between them and real data; D(G(z|y)) is the discriminator's response to the random noise vector generated by the generator G under the given condition y. zThe predicted value of the generated sample indicates the probability that the forged sample is mistaken for a real sample; Indicates that for the real data distribution The real data extracted from x Expected value; For a random noise distribution P z The expected value of the noise vector z extracted in (z).

[0096] The training steps of the CGAN module are as follows:

[0097] step1.1: The generator receives a random noise vector z and condition information y As input, generate gene samples ;

[0098] Step 1.2: The real sample (x real ,y real ) and the generated fake samples (x fake ,y real ) Input the discriminator to obtain the discriminator's predicted value D(x real ∣y real) And the predicted value D(x fake ∣y real ), and then calculate the discriminator loss L D (that is, the value function), and use the optimizer (adam optimizer) to update the parameters of the discriminator;

[0099] Step 1.3: Generate another batch of fake samples (x fake =G(z|y real ),y real ), input these fake samples into the discriminator, and obtain the discriminator's predicted value D(x fake ∣y real ), calculate the generator loss L G (that is, the value function), and use the optimizer to update the parameters of the generator;

[0100] step1.4: The generator and discriminator continue to compete and iteratively update until the specified number of training rounds is reached.

[0101] Sparse Autoencoder (SAE) is a variant of autoencoder that learns a concise representation of data by introducing sparsity constraints. Unlike traditional autoencoders, SAE limits the number of neuron activations in the hidden layer so that only a small number of neurons are activated, thereby extracting more effective and semantically meaningful features. This feature makes SAE particularly suitable for tasks such as feature extraction, dimensionality reduction, and pre-training of deep learning models.

[0102] Although SAE is so applicable in feature extraction, it is limited in capturing the global information of features. Genes do not work alone, but are expressed together through the interaction between multiple genes. Therefore, in order to better explore the pathogenic genes of rice blast and enhance the ability of SAE to capture global features, this embodiment adds a gated axial attention layer to both the encoder and decoder networks of SAE; specifically, the structure of the encoder of SAE includes a first linear layer, a relu activation function, a second linear layer, and a gated axial attention layer arranged in sequence; the structure of the decoder is the same as that of the encoder. The gated axial attention mechanism processes information sequentially along the height and width axes, achieving effective modeling of long-range dependencies while maintaining computational efficiency. In addition, due to its unique gating mechanism, it can adaptively adjust the attention weights, enabling the model to better capture the complex interactions between genes in the regulatory network.

[0103] The data processing steps for SAE and PCA are as follows:

[0104] Step 2.1: Map the gene sample X and the generated sample G to the encoder to obtain the low-dimensional representation D;

[0105] Step 2.2: Map the low-dimensional representation back to the original data space to obtain the reconstructed data X1;

[0106] step2.3: Use PCA to analyze and reconstruct the data X1 to obtain the gene score matrix Z.

[0107] In this embodiment, the loss function of the SAE module consists of reconstruction error and sparsity constraint, as shown in formula (2):

[0108]

[0109] in, Represents the weight of the sparsity constraint.

[0110] Reconstruction error As shown in formula (3):

[0111]

[0112] Where n represents the number of samples; Represents the sample of the i-th input; Table ith reconstructed sample.

[0113] Sparsity Constraint As shown in formula (4):

[0114]

[0115] Among them, m refers to the number of neurons in the hidden layer, that is, the dimension of the hidden layer in the sparse autoencoder; Refers to the preset sparsity target (set to 0.01 in this example), which indicates the average proportion of neurons that are expected to be activated in the hidden layer; Refers to the actual activation ratio of the jth neuron, calculated as the average value of the activation of the neuron in all input samples; The function represents the KL divergence.

[0116] In the model design, the present invention combines the advantages of CGAN and SAE, and the features finally extracted are further processed by principal component analysis (PCA) to evaluate the importance of the features.

[0117] PCA is a widely used linear dimensionality reduction method that reduces the dimensionality by projecting the original high-dimensional data onto a few principal components while retaining the original information of the data as much as possible. In this embodiment, PCA can not only effectively reduce the redundancy of high-dimensional features, but also score the importance of features to help identify the core genes in the rice blast infection process.

[0118] For the data set studied in this embodiment, the total number of samples is only 6. After protein-protein interaction (PPI) network screening, 30 related genes remain, that is, 30 features are included. In such a small sample and high dimensional situation, many traditional models are prone to overfitting problems. However, this embodiment generates new sample data through CGAN, and continuously introduces these new samples into the training process, which significantly reduces the risk of overfitting and enhances the generalization ability of the model. Finally, combining the feature extraction of SAE and the feature screening of PCA, the core genes of rice blast were successfully predicted.

[0119] Specifically, the present invention constructs a feature screening model based on a conditional generative adversarial network and a sparse autoencoder, takes the 30 genes analyzed by the PPI network as features, then inputs the samples (6 samples) into the model, and divides all positive and negative samples (infected as negative samples, normal as positive samples) into training sets and test sets at a ratio of 2:1. The model will be trained according to the input samples, and 100 new simulated samples are generated in the CGAN module, and are distributed to the training set and the test set at a ratio of 9:1. Then, the new samples (simulated samples) are combined with the real samples, and the sparse autoencoder is used to extract and reconstruct the input real data and the generated data. Finally, the PCA module performs feature importance analysis and outputs the importance scores of all features.

[0120] In this example, the top five features with the highest importance scores were selected as the final results, namely 'Os01g0191800', 'Os01g0169800', 'Os03g0162200', 'Os11g0455500', 'Os01g0927000'; and the expression level of each gene was compared with the control group. The box plot results showed that ( Figure 7 ), for which there were significant differences between the two groups, these features may be related to specific biological processes or disease mechanisms, providing a basis for further biological validation.

[0121] In order to verify the performance of the model, this embodiment uses AUC (Area Under the Curve) as the evaluation indicator. AUC is an indicator to measure the overall performance of the model, which can comprehensively reflect the performance of the model under all possible decision thresholds. Compared with a single accuracy rate, recall rate or F1 score, AUC has higher stability and reliability. Even if the threshold is changed, the change in AUC is relatively small. Therefore, it can more comprehensively reflect the pros and cons of the model when evaluating it.

[0122] In this study, the ROC (Receiver Operating Characteristic) curve and its AUC (Area Under the Curve) were used as key indicators to evaluate model performance. The ROC curve provides a comprehensive view of the performance of the classifier at different thresholds, not limited to the accuracy at a single threshold. This is particularly important in the field of bioinformatics, because gene expression data are often unbalanced, and using a single threshold may ignore the performance of the model on different categories. The ROC curve can intuitively show the relationship between the True Positive Rate (TPR) and the False Positive Rate (FPR) of the classifier at different thresholds, while the AUC quantifies the quality of this relationship. The AUC value is not affected by sample imbalance, so the performance of the model can be fairly evaluated even when the ratio of positive and negative samples is unbalanced.

[0123] After evaluating the samples generated by CGAN and extracting features through sparse autoencoders, this embodiment obtained the ROC-AUC score, and the results showed that the model of the present invention has a high discrimination ability. Specifically, the ROC-AUC score on the test set is 0.80.

[0124] The present invention compares the RandomForest, CGAN-RandomForest, CNN, CGAN-CNN and other models; RandomForest is one of the most popular machine learning methods for gene screening, and CNN is a deep learning model that also has very good results in feature extraction. As can be seen from Table 1, without the introduction of CGAN, all the scoring criteria of RandomForest and CNN reached 1, indicating that serious overfitting occurred, while the introduction of CGAN did reduce the risk of model overfitting.

[0125] Table 1 shows the scores of each model.

[0126] Accuracy F1 score Precision Recall AUC RandomForest 1.00 1.00 1.0 1.0 1.00 CGAN-RandomForest 0.45 0.45 0.47 0.45 0.53 CNN 1.0 1.0 1.0 1.0 1.0 CGAN-CNN 0.55 0.62 0.57 0.67 0.37 CGAN-SAE 0.75 0.77 0.83 0.71 0.80 CGAN-SVM 0.42 0.59 0.45 0.83 0.64 CGAN-XGBoost 0.58 0.62 0.67 0.57 0.67

[0127] From the above results, it can be seen that when CGAN is introduced and compared with other models, after comprehensive comparison of the above scores, the model of the present invention (CGAN-SAE) scores higher among these models and has the best effect.

[0128] In order to obtain a better model effect, the present invention conducted multiple experiments on the number of training rounds (epochs) of CGAN-SAE. Figure 8 It can be seen that when the epoch is 100 (both CGAN and SAE are 100), the model works best.

[0129] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art of the present invention, according to the idea of ​​the present invention, some simple deductions, deformations or substitutions can be made. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data, characterized in that: The method comprises the following steps: Step 1. Data collection; Step 2. Process the data collected in step 1 to obtain the expression level of each gene corresponding to the sample; Step 3. Perform differential analysis on the data processed in step 2 to identify differentially expressed genes in the data; Step 4. Construction and analysis of protein-protein interaction network; A protein-protein interaction network was constructed based on the differentially expressed genes determined in step 3, and then the protein-protein interaction network was reconstructed using Cytoscape software integrated with the CytoHubba plug-in. The reconstructed protein-protein interaction network was analyzed to determine the key genes related to rice blast disease; Step 5. Construct a CGAN-SAE deep learning model, and use the key genes related to rice blast determined in step 4 to train the CGAN module, so that the trained CGAN module can generate simulated gene sample data that is difficult to distinguish from real data; then input the newly generated simulated gene samples and the real samples obtained in step 1 into the SAE module for training, so that the trained SAE module can capture global features when performing feature extraction, and the extracted features are further processed by principal component analysis to reduce the redundancy of high-dimensional features, and at the same time, the features are scored for importance, so as to identify the core genes in the rice blast infection process; The CGAN module includes a generator and a discriminator; wherein the network structure of the generator includes a first block layer, a second block layer, a third block layer, a fourth block layer, a linear layer, and a tanh activation function, which are arranged in sequence. The first block layer is a linear layer and is not regularized; the second block layer, the third block layer, and the fourth block layer are all linear layers and regularized, the fourth block layer is followed by a linear layer, and the tanh activation function is used after the linear layer; the network structure of the discriminator includes a first linear layer, a leaky relu activation function, a second linear layer, a first dropout layer, a leaky relu activation function, a third linear layer, a second dropout layer, a leaky relu activation function, and a fourth linear layer, which are arranged in sequence; wherein the dropout rate of the first dropout layer and the second dropout layer is 0.4; The value function of the CGAN module is: ； In the formula, is to maximize V(G,D) for the discriminator D; is to minimize V(G,D) for the generator G; is the discriminator for the real data x The predicted value of , D(G(z|y)) is the given condition y In the case of z The predicted values ​​of the generated samples; Indicates that for the real data distribution The real data extracted from x Expected value; For a random noise distribution P z The noise vector extracted from (z) z Expected value; A gated axial attention layer is added to both the encoder and decoder networks of the SAE module. The network structure of the SAE encoder includes a first linear layer, a relu activation function, a second linear layer, and a gated axial attention layer arranged in sequence; the network structure of the decoder is the same as that of the encoder.

2. A method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data according to claim 1, characterized in that: In step 3, the "limma" package of the R programming language was used as a tool for differential analysis. A linear model was constructed to fit the expression levels of genes, thereby identifying differentially expressed genes in the data.

3. A method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data according to claim 1, characterized in that: In step 4, when constructing a protein-protein interaction network, the STRING database is used, and the differentially expressed genes obtained after differential analysis are input into the STRING database.

4. The method for identifying rice blast pathogenicity genes suitable for small sample high-dimensional data according to claim 1, characterized in that: Loss function of SAE module L By reconstruction error and sparsity constraints The composition is expressed as: ； ； ； in, represents the weight of adjusting the sparsity constraint; n represents the number of samples; Represents the sample of the i-th input; Table ith reconstructed sample; m refers to the number of neurons in the hidden layer, refers to the preset sparse target, refers to the actual activation ratio of the jth neuron, The function represents the KL divergence.

Images