A method for identifying key sRNA of rice against Magnaporthe oryzae based on machine learning

Abstract

The application discloses a rice anti-rice blast key SRNA identification method based on machine learning, comprising the following steps: in step one, in a rice infection data set, rice SRNA and MRNA sequence information is obtained; in step two, screening conditions are set, and differentially expressed rice SRNA and MRNA are screened; in step three, a heterogeneous interaction network of differentially expressed rice genes is constructed; in step four, in the heterogeneous interaction network, modeling explicit relationships and modeling implicit relationships are jointly optimized, and through the joint optimization result, rice anti-infection key SRNA in the process of rice blast infection of rice is screened out. By constructing the heterogeneous interaction network of differentially expressed rice genes and performing joint optimization, key rice SRNA in the process of rice resisting rice blast infection can be found out, the deficiency that traditional methods cannot consider the mutual influence between biological molecules of the same type is made up, and the prediction accuracy of the method is improved.

Description

Technical Field

[0001] This invention relates to a method for identifying key SRNAs of rice resistance to rice blast fungus based on machine learning, which belongs to the field of biological big data mining technology. Background Technology

[0002] Rice is a vital food crop worldwide, but it is susceptible to damage from insects and microorganisms, especially from rice blast fungus, a fungal disease that can infect rice throughout its entire growth cycle, severely impacting yield. However, the high variability of rice blast fungus makes chemical control costly, while also causing environmental pollution and affecting food safety. Therefore, identifying the key pathogenic factors of rice blast fungus at the cellular level to understand its infection mechanisms is crucial for the long-term control of rice blast disease.

[0003] In fungal-plant interactions, researchers have increasingly focused on the role of key biomolecules. A growing body of research indicates that small non-coding RNAs play crucial roles in many biological processes, such as RNA modification, regulating gene transcription and translation, and participating in protein translocation and degradation. However, with the rapid growth of biological data, traditional bioanalysis methods have faced significant challenges in many fields. The latest machine learning methods, especially deep learning, can greatly accelerate research in bioinformatics. More importantly, they can uncover hidden connections and patterns within vast amounts of biological data, providing crucial theoretical support for further research and exploration.

[0004] Currently, machine learning research on the rice blast fungus infection process mainly focuses on the biomolecules that play a key role in infection, while neglecting the key biomolecules that help rice resist infection. Furthermore, analysis of infection networks only considers the regulatory relationships between different biomolecules, ignoring the implicit relationships between similar biomolecules. Therefore, the analysis results for infection networks where similar biomolecules influence each other are somewhat lacking, and there is room for improvement. Summary of the Invention

[0005] This invention designs and develops a machine learning-based method for identifying key SRNAs in rice resistance to rice blast. By constructing a heterogeneous interaction network of differentially expressed rice genes and performing joint optimization, it can identify key rice SRNAs in the process of rice resistance to rice blast infection, making up for the shortcomings of traditional methods that cannot consider the interaction between similar biomolecules and improving the accuracy of the method's predictions.

[0006] The technical solution provided by this invention is as follows:

[0007] A machine learning-based method for identifying key SRNAs for rice blast resistance includes:

[0008] Step 1: Obtain rice sRNA and mRNA sequence information from the rice infection dataset through gene bank mapping and comparison;

[0009] Step 2: Set screening criteria to screen for differentially expressed sRNAs and mRNAs in rice;

[0010] Step 3: Construct a heterogeneous interaction network for differentially expressed rice genes;

[0011] Step 4: In the heterogeneous interaction network, jointly optimize and model explicit relationships and implicit relationships. Based on the joint optimization results, screen out key SRNAs for rice resistance to infection during the rice blast fungus infection process.

[0012] Preferably, step one includes:

[0013] Based on a dataset of rice infected with blast fungus with a given infection duration, gene sequence information was extracted and mapped into a rice RNA library. The extracted rice gene sequences were matched with gene sequences in the rice RNA library, and the successfully matched sequences were identified as rice RNA sequences.

[0014] The matched rice RNA sequence information is mapped into the rice sRNA library. Successful matches are rice sRNA sequences, and the rest are rice transcribed mRNAs.

[0015] Preferably, step two includes:

[0016] Expression levels of rice sRNA and rice mRNA were statistically analyzed in the infection dataset.

[0017] The expression data were standardized using the 3 / 4 quantile standardization method.

[0018] Differentially expressed genes were screened by setting a fold change threshold for expression levels, and these genes, along with newly emerging genes in resistant rice compared to susceptible rice, were used as differentially expressed genes.

[0019] Preferably, step three includes:

[0020] The differentially expressed genes were input into the database for targeted gene prediction, and the prediction results of rice SRNA targeting rice MRNA were obtained.

[0021] Based on the prediction results, a heterogeneous network node graph is constructed. Rice sRNA is set as a type I site and rice mRNA is set as a type II site. Edges are formed between the two types of sites that have a targeting relationship.

[0022] The edges are weighted based on the fold change in gene expression corresponding to each pair of fixed points, and the weight is the product of the fold changes in gene expression between the two fixed points.

[0023] Preferably, step four includes:

[0024] The constructed heterogeneous network is reconstructed, and explicit modeling tasks are performed from the observable edges, while implicit modeling tasks are performed from the unobservable edges. Fixed-point embeddings are learned by jointly optimizing the above two tasks, and key anti-infection SRNAs are screened out through similarity analysis.

[0025] Preferably, in step four, the display modeling task includes:

[0026] The explicit relationship is modeled by considering the local proximity between two dissimilar connection vertices in a heterogeneous network. Modeling is performed using both empirical and joint probability distributions, with KL divergence used to measure the difference between the two distributions. Learning is achieved by minimizing an objective function, which is:

[0027]

[0028] In the formula, For u i and v j The joint probability distribution, For u i and v j Local similarity; and u i and v j Embedded vector, e ij Let E be an edge consisting of nodes i and j in the edge set E.

[0029] Preferably, the implicit modeling task includes:

[0030] By considering the implicit relationship between two identical nodes without direct edge connection in a heterogeneous network, the second-order proximity between two nodes i and j, for a vertex u, is defined as:

[0031]

[0032] In the formula, w ik w represents the weight between node i and its neighboring node k in the second-class node v. jk Let U be the weight between node j and adjacent node k in the second type of node v, and U be the set of fixed points u.

[0033] For a binary vertex v, the second-order proximity between two nodes i and j is defined as:

[0034]

[0035] Generate two homogeneous network graphs, namely:

[0036]

[0037]

[0038] Preferably, it also includes:

[0039] The overall objective function is jointly optimized to obtain the embedding vector of each fixed point. Through joint similarity analysis between the two types of vertices, the sRNAs in the top-ranked relationships are selected as the rice sRNAs that play a key role in resisting infection.

[0040] The overall objective function is:

[0041] max imizeL=αlog O2+βlog O3-γO1,

[0042] In the formula, α, β, and γ are hyperparameters specified according to the actual situation, O1 is the difference function between the explicit relationships, and O2 and O3 are the node u i and v j The difference function of the implicit relationship between them.

[0043] The beneficial effects of this invention are as follows:

[0044] 1. This invention aims to identify key SRNAs in rice that resist rice blast fungus infection from the perspective of resistance to rice blast fungus infection;

[0045] 2. This invention proposes that implicit relationships exist between genes of the same type in interaction, thereby enabling better discovery and analysis of gene interactions;

[0046] 3. This invention proposes a joint optimization strategy that comprehensively considers both explicit and implicit relationships between genes, which can more effectively promote learning and improve generalization. Attached Figure Description

[0047] Figure 1 This is a flowchart of the rice blast resistance SRNA recognition method using machine learning described in this invention.

[0048] Figure 2 This invention describes the process for obtaining and processing rice blast fungus-infected datasets.

[0049] Figure 3 This is a schematic diagram of the weighted heterogeneous network construction of differentially expressed sRNA and mRNA in rice according to the present invention.

[0050] Figure 4 This is a training graph for the skip-gram model described in this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0052] like Figure 1-4 As shown, this invention provides a machine learning-based method for identifying key SRNAs in rice resistance to rice blast fungus. By constructing a heterogeneous interaction network of differentially expressed rice genes and performing joint optimization, it can identify key rice SRNAs in the process of rice resistance to rice blast fungus infection. This overcomes the shortcomings of traditional methods that cannot consider the interactions between similar biomolecules, thus improving the accuracy of the method's predictions. The method includes:

[0053] Step 1: Obtain rice sRNA and mRNA sequence information from the rice infection dataset;

[0054] The dataset includes rice infection data and rice blast infection datasets for a given infection duration, comprising both susceptible and resistant rice plants. The rice SRNA and mRNA sequence information was obtained by comparing these datasets with a gene bank mapping.

[0055] First, the gene sequence information is extracted and mapped into a rice RNA library. The rice RNA sequences that can be successfully matched are the rice RNA sequences.

[0056] The matched rice RNA sequence information is mapped to the rice sRNA library. Successful matches are rice sRNAs, and the rest are rice transcribed mRNAs.

[0057] Step 2: Set screening criteria to screen for differentially expressed sRNAs and mRNAs in rice;

[0058] The frequency of occurrence of each rice SRNA and rice MRNA was counted, which is the expression level. Differentially expressed genes were then screened based on the fold increase or decrease in expression level, including:

[0059] Expression levels of rice sRNA and rice mRNA were statistically analyzed in the infection dataset, i.e., the frequency of occurrence was counted.

[0060] To ensure comparability between different datasets, the expression level data needs to be standardized using the 3 / 4 quantile standardization method. The specific process involves sorting the obtained rice sRNA and rice mRNA expression level data in descending order, then using the expression level at the 3 / 4 position in the sort as the benchmark, and finally recording all other expression level values ​​as multiples of the benchmark value.

[0061] Set a fold change F for differential expression. If the expression level of a certain gene in a disease-resistant rice plant increases or decreases by a factor greater than N compared to the susceptibility level, it is initially considered to be a differentially expressed gene. In addition, genes that appear in the disease-resistant type but not in the susceptible type should be screened. All differentially expressed sRNAs and mRNAs are screened by comparison, and their fold changes are recorded.

[0062] Step 3: Construct a heterogeneous interaction network for differentially expressed rice genes;

[0063] A heterogeneous interaction network of rice genes was constructed using the obtained differentially expressed sRNAs and mRNAs, including:

[0064] The differentially expressed sRNA and mRNA obtained in step two were input into the bibiserv database to obtain the prediction results of rice sRNA targeting rice mRNA;

[0065] Based on the prediction results, a heterogeneous network node graph is constructed. Rice sRNA is denoted as a type 1 vertex, and rice mRNA is denoted as a type 2 vertex. An edge is formed between the two types of vertices that have a targeting relationship.

[0066] The edges are weighted according to the differential expression fold of the genes corresponding to each pair of vertices, and the weight is the product of the differential expression folds of the two vertices;

[0067] Step 4: In the heterogeneous interaction network, jointly optimize and model explicit relationships and implicit relationships. Based on the joint optimization results, screen out key SRNAs for rice resistance to infection during the rice blast fungus infection process.

[0068] In the constructed heterogeneous interaction network, explicit relationships are modeled for heterogeneous nodes using local similarity, while implicit relationships are modeled for homogeneous nodes using a walk method and vertex embeddings learned from the obtained corpus. Through joint optimization of these two methods, key SRNAs for rice resistance to blast fungus infection during rice growth were finally screened, including:

[0069] The explicit relationship is modeled by considering the local proximity between two dissimilar connection vertices in a heterogeneous network. Modeling is performed using both empirical and joint probability distributions, with KL divergence used to measure the difference between the two distributions. Learning is achieved by minimizing an objective function, which is:

[0070]

[0071] In the formula, For u i and v j The joint probability distribution, For u i and v j Local similarity; and u i and v j Embedded vectors.

[0072] By considering the implicit relationship between two identical nodes without direct edge connection in a heterogeneous network, the second-order proximity between two nodes i and j, for a vertex u, is defined as:

[0073]

[0074] In the formula, w ik w represents the weight between node i and its neighboring node k in the second-class node v. jk Let be the weight between node j and its adjacent node k in the second-class node v;

[0075] Similarly, for a vertex v of type II, the second-order proximity between two nodes i and j is defined as:

[0076]

[0077] Generate two homogeneous network graphs, namely:

[0078]

[0079]

[0080] Partial random walks are performed on graphs u and v respectively. The number of steps is determined based on the degree centrality of each node, allowing the generated vertex sequences to have variable lengths. The vertex sequences obtained by the partial random walks are then fed into the Skipgram model to learn vertex embeddings.

[0081] The overall objective function is jointly optimized:

[0082] max imizeL=αlog O2+βlog O3-γO1;

[0083] The joint was optimized using the gradient ascent (SGA) method, and the embedding vector of each vertex was finally obtained. Through joint similarity analysis between the two types of vertices, the SRNAs in the top few relationships were selected as rice SRNAs that play a key role in resisting infection.

[0084] Specifically, it includes:

[0085] In this invention, as a preferred embodiment, VMware 15 Pro software is used to install the CentOS 7.7 system and run it in a Linux environment.

[0086] Step 1) Collect mixed sequencing read data of rice blast fungus-infected resistant and susceptible rice from the NCBI database. Then, use the bowtie2 tool to compare and obtain mixed sequencing read data of rice sRNA and rice mRNA. Finally, based on the difference in expression levels between the resistant and susceptible data, screen out differentially expressed sRNA and mRNA.

[0087] like Figure 2 As shown, the specific implementation of the alignment process and screening of differentially expressed genes is as follows:

[0088] Step 1.1) Collect mixed sequencing reads of blast fungus infection of resistant rice and susceptible rice 36 hours after infection in the NCBI database and store them in the local database. First, use the trim galore tool to perform quality control on the raw data. Then, use the bowtie2 tool to build a library of rice genome data to obtain an index file. Match the mixed sequencing reads of blast fungus infection of resistant rice and susceptible rice 36 hours after infection with the index and output the matching information to the sam file. Then, use the samtool tool to remove redundancy from the sam file and retain the original matching information to finally obtain the infected rice RNA data.

[0089] Step 1.2) The data obtained in Step 1.1) are length-controlled, and those with a length of 19-25 nucleotides are selected as sRNAs with regulatory functions, while the rest are mRNAs responsible for translation.

[0090] Step 1.3) The length-controlled sRNA and mRNA data are matched with the rice genome structure annotation file, and the number of times each sequence appears in the mapping process is counted and recorded as the expression level of the sRNA or mRNA.

[0091] Step 2) Standardize gene expression levels using the 3 / 4 quantile standardization method; and set an expression level fold threshold to screen differentially expressed genes. At the same time, count newly emerging genes in the resistant and susceptible types as differentially expressed genes.

[0092] The process of screening differentially expressed genes is as follows:

[0093] Step 2.1) Arrange the calculated rice gene expression data in descending order from largest to smallest. Then, take the expression value at the 3 / 4 position in this sort as the benchmark. Finally, record all other expression values ​​as multiples of the benchmark value. Set a threshold F and filter out genes in the two datasets whose expression levels are upregulated or downregulated by more than F.

[0094] Step 2.2) Sort the newly emerging genes in resistant rice relative to susceptible rice in descending order after infection, screen genes with an increase of 12 or more, and standardize the increase as the differential expression fold.

[0095] Step 2.3) Extract the genes screened in Step 2.1) and Step 2.2) and merge them into differentially expressed genes.

[0096] Step 3) Using the differentially expressed sRNA and mRNA datasets from Step 2), target gene prediction is performed, and then a heterogeneous interaction network of rice genes is generated based on the predicted target relationships and the differential expression folds.

[0097] The process of constructing a heterogeneous interaction network is as follows:

[0098] Step 3.1) Input the fasta format data of rice sRNA and rice mRNA into the bibiserv website as parameters, select the RNAhybrid option to predict target genes, and screen the parts with minimum free energy less than -24.0 kcal / mol to obtain the target relationship.

[0099] Step 3.2) For each pair of sRNA and mRNA targeting relationships, the product of the fold change in sRNA and mRNA expression levels is used as the weight between them, and finally a weighted heterogeneous interaction network is constructed.

[0100] Step 4) Figure 3 As shown, the heterogeneous interaction network constructed in step 3) is reconstructed. Explicit modeling is performed on the observable edges (SRNA-MRNA), and implicit modeling is performed on the unobservable edges (SRNA-SRNA, MRNA-MRNA). Vertex embeddings are then learned by jointly optimizing these two tasks. Finally, key anti-infection SRNAs are screened out through similarity association analysis.

[0101] The specific implementation process of explicit modeling, implicit modeling, and joint optimization is as follows:

[0102] Step 4.1) Explicit relationships are modeled by considering the local proximity between two dissimilar connected vertices in a heterogeneous network. The joint probability (empirical distribution) between vertices u[i] and v[j] is defined as:

[0103]

[0104] In the formula, w ij It is edge e ij The weight.

[0105] Use u i and v jThe inner product of the two embeddings is used to estimate local similarity (reconstruct the distribution):

[0106]

[0107] In the formula, It is an embedding vector of a vertex u. It is the embedding vector of a second-class vertex v.

[0108] Using the KL divergence to measure the difference between two distributions, we learn the following by minimizing this objective function:

[0109]

[0110] In the formula, For u i and v j The joint probability distribution, For u i and v j Local similarity; and u i and v j Embedded vector, e ij Let E be an edge consisting of nodes i and j in the edge set E.

[0111] Step 4.2) By modeling an implicit relationship between two nodes of the same type that are not directly connected by edges in a heterogeneous network, the second-order proximity between two nodes i and j is defined as:

[0112]

[0113] In the formula, w ik w represents the weight between node i and its neighboring node k in the second-class node v. jk Let U be the weight between node j and adjacent node k in the second type of node v, and U be the set of fixed points u.

[0114] Similarly, for a vertex v of type II, the second-order proximity between two nodes i and j is defined as:

[0115]

[0116] Generate two homogeneous network graphs, namely:

[0117]

[0118]

[0119] Step 4.3) Perform biased random walks on graphs u and v respectively. The number of steps is determined based on the degree centrality of each node, allowing the generated vertex sequence to have a variable length. The formula for random walks in the network is:

[0120]

[0121] Where, π vx It is the transition probability between nodes, c i For the i-th walk, Z is a constant.

[0122] In a biased random walk, π vx The definition is as follows:

[0123] π vx =α pq (t, x)·w vx

[0124] α pq (t, x) is defined as:

[0125]

[0126] In the formula, d tx The shortest path from node t to x is defined by the value {0, 1, 2}. p and q are hyperparameters set according to the actual situation to determine whether to perform a width traversal or a depth traversal.

[0127] Step 4.4) Subsequently, the vertex sequence obtained from the biased random walk is fed into the skip-gram model to learn vertex embeddings, such as... Figure 4 As shown, the total number of nodes in the graph is set to L, and the dimension of the embedding matrix is ​​N. Each node is one-hot encoded with dimension L, and the embedding matrix is ​​initialized as follows: W L×N In one training iteration, a 128×L matrix dot product matrix W is formed using one-hot encodings of 128 nodes. L×N The output matrix has a shape of 128×N and is passed through a hidden layer with 128 computational units; a fully connected layer is used as the output layer, containing 128 neurons; finally, the classifier uses the Sigmoid activation function.

[0128] For input u, the output v' section result is compared with the actual v section. If they do not match, the backpropagation algorithm is used to correct the weight vectors W and W'.

[0129] Step 4.5) In a bipartite graph, there are two types of nodes, u and v. Taking u as an example, let D... u For the corpus consisting of u-node sequences, C s (u i ) represents node u iThe context node; maximize the following conditional probabilities:

[0130]

[0131] In the formula, C S (u i ) is node u i The context node, S is a certain prediction formed by the sequence of u nodes;

[0132] Similarly, for node v,

[0133]

[0134] Wherein, P(u c |u i ) and P(u c |u j The conditional probability is parameterized based on the inner product kernel with softmax output:

[0135]

[0136]

[0137] In the formula, V is the set of nodes v, and U is the set of nodes u. and They are nodes u i and node v j eigenvectors, and These are the preceding and following nodes in the corpus S, respectively, and exp() is an exponential function with base e.

[0138] Step 4.6) Employ joint optimization of the overall objective function:

[0139] The overall objective function is:

[0140] max imizeL=αlog O2+βlog O3-γO1,

[0141] In the formula, α, β, and γ are hyperparameters specified according to the actual situation, O1 is the difference function between the explicit relationships, and O2 and O3 are the node u i and v j The difference function of the implicit relationship between them;

[0142] The joint algorithm is optimized using the gradient ascent (SGA) method. Considering that the three parts of the above equation define a training instance differently, the following gradient update method is adopted:

[0143] Step 1: For an implicit relation, update the expression by maximizing the last part L2 = -γO1. and

[0144]

[0145]

[0146] In the formula, and They are nodes u i and node v j The feature vector, λ is the learning rate, γ is a user-defined hyperparameter, and ω is a function vector. ij Let σ be the weight between node i and node j, and let σ() represent the activation function sigmoid().

[0147] step2: change u i and v j As a central node, using SGA to maximize the objective functions L2 = αlogO2 and L2 = βlogO2 can maximize the preservation of implicit relations. For a given central node u i or v j and context node u c or v c Update their embeddings:

[0148]

[0149]

[0150] In the formula, λ is the learning rate, α and β are user-defined hyperparameters, and σ() indicates the use of the sigmoid() activation function. This means selecting node u from the set of all nodes. i In the corpus, the corpus of each vertex forms the complement of the set, and z is u. i Context node set {u c}and and centralized nodes

[0151] I(z,u i The indicator function is used to determine whether node z is node u. i Similarly, update the context vector based on the context:

[0152]

[0153]

[0154] Step 4.7) Perform cosine similarity analysis on the embedding vectors of nodes u and v, and select the rice sRNAs with the highest similarity in the 20 pairs of relationships. These are the key sRNAs in the process of rice resistance to rice blast fungus infection, including:

[0155] Embedding vectors of nodes u and v and Perform cosine similarity analysis; the cosine similarity value is:

[0156]

[0157] By constructing a heterogeneous interaction network of differentially expressed rice genes and performing joint optimization, key rice sRNAs in the process of rice resistance to rice blast infection can be identified. This overcomes the shortcomings of traditional methods that cannot consider the interaction between similar biomolecules and improves the accuracy of the method's predictions.

[0158] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for identifying key SRNAs of rice resistance to rice blast fungus based on machine learning, characterized in that, include: Step 1: Obtain rice sRNA and mRNA sequence information from the rice infection dataset through gene bank mapping and comparison; Step 2: Set screening criteria to screen for differentially expressed sRNAs and mRNAs in rice; Step 3: Construct a heterogeneous interaction network for differentially expressed rice genes, including: The differentially expressed genes were input into the database for targeted gene prediction, and the prediction results of rice SRNA targeting rice MRNA were obtained. Based on the prediction results, a heterogeneous network node graph is constructed. Rice sRNA is set as a type I site and rice mRNA is set as a type II site. Edges are formed between the two types of sites that have a targeting relationship. The edges are weighted according to the fold change in gene expression corresponding to each pair of fixed points, and the weight is the product of the fold change in gene expression of the two fixed points. Step 4: In the heterogeneous interaction network, jointly optimize and model explicit relationships and implicit relationships. Based on the joint optimization results, screen out key SRNAs for rice resistance to blast fungus infection during rice growth, including: The constructed heterogeneous network is reconstructed, and an explicit modeling task is performed from the observable edges, while an implicit modeling task is performed from the unobservable edges. The fixed-point embedding is learned by jointly optimizing the above two tasks, and key anti-infection SRNAs are screened out through similarity analysis. Display modeling tasks include: The explicit relationship is modeled by considering the local proximity between two dissimilar connection vertices in a heterogeneous network. Modeling is performed using both empirical and joint probability distributions, with KL divergence used to measure the difference between the two distributions. Learning is achieved by minimizing an objective function, which is: In the formula, For u i and v j The joint probability distribution, For u i and v j Local similarity; and u i and v j Embedded vector, e ij Let E be an edge consisting of nodes i and j in the edge set E. Implicit modeling tasks include: By considering the implicit relationship between two identical nodes without direct edge connection in a heterogeneous network, the second-order proximity between two nodes i and j, for a vertex u, is defined as: In the formula, w ik w represents the weight between node i and its neighboring node k in the second-class node v. jk Let U be the weight between node j and adjacent node k in the second type of node v, and U be the set of fixed points u. For a binary vertex v, the second-order proximity between two nodes i and j is defined as: Generate two homogeneous network graphs, namely:

2. The method for identifying key SRNAs of rice resistance to rice blast fungus based on machine learning according to claim 1, characterized in that, Step one includes: Based on a dataset of rice infected with blast fungus with a given infection duration, gene sequence information was extracted and mapped into a rice RNA library. The extracted rice gene sequences were matched with gene sequences in the rice RNA library, and the successfully matched sequences were identified as rice RNA sequences. The matched rice RNA sequence information is mapped into the rice sRNA library. Successful matches are rice sRNA sequences, and the rest are rice transcribed mRNAs.

3. The method for identifying key SRNAs of rice resistance to rice blast fungus based on machine learning according to claim 1 or 2, characterized in that, Step two includes: Expression levels of rice sRNA and rice mRNA were statistically analyzed in the infection dataset. The expression data were standardized using the 3 / 4 quantile standardization method. Differentially expressed genes were screened by setting a fold change threshold for expression levels, and these genes, along with newly emerging genes in resistant rice compared to susceptible rice, were used as differentially expressed genes.

4. The method for identifying key SRNAs of rice resistance to rice blast fungus based on machine learning according to claim 3, characterized in that, Also includes: The overall objective function is jointly optimized to obtain the embedding vector of each fixed point. Through joint similarity analysis between the two types of vertices, the sRNAs in the top-ranked relationships are selected as the rice sRNAs that play a key role in resisting infection. The overall objective function is: maximizeL=αlogO2+βlogO3-γO1, In the formula, α, β, and γ are hyperparameters specified according to the actual situation, O1 is the difference function between the explicit relationships, and O2 and O3 are the node u i and v j The difference function of the implicit relationship between them.

Images