Wheat abiotic stress character SNP prediction method based on automatic CNN model

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a wheat abiotic stress character SNP (Single Nucleotide Polymorphism) prediction method based on an automatic CNN (Convolutional Neural Network) model, which comprises the following steps of: collecting GWAS data related to wheat from a GWAS Atlas database, and constructing a high-quality mutation site data set; performing feature extraction on the data set by using a DNA original sequence, text-based features and structure-based features, sorting the importance of input features by using an SHAP method, assisting feature selection, and explaining the result of the prediction model; then, an automatic architecture search framework is constructed, automatic neural architecture search is adopted, the model is evolved in a specified search space, the optimal performance is achieved on a verification set, and the accuracy, the AUPR and the AUC of the model are evaluated on a test set. The invention provides a deep learning framework which is efficient, accurate and explainable, and is used for predicting the SNP related to the wheat biological stress, so that the development of wheat stress resistance breeding and genomics research is promoted.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for predicting wheat abiotic stress trait SNP based on an automated CNN model. Background Art

[0002] Single nucleotide polymorphisms (SNPs) are the most common type of variation in the genome and are associated with a variety of crop traits, including disease resistance and stress tolerance. Genome-wide association studies (GWAS) are one of the main methods for identifying SNPs associated with specific traits. However, although GWAS has made significant progress in identifying major quantitative trait loci (QTLs) associated with biotic stress resistance, many SNPs controlled by minor genes remain unidentified. In addition, the lack of high-quality genotype data limits the further application of GWAS.

[0003] In recent years, deep learning (DL) technology has made remarkable progress in the fields of bioinformatics and genomics. DL methods, especially convolutional neural networks (CNNs), have performed well in extracting structural and local features from DNA sequences and have been used to predict functional sites, identify genetic variants, and predict complex phenotypes. For example, Pau Bellot et al. used DL to predict functional sites in human genomic regions, SWAT-CNN used DL to identify SNPs associated with phenotypes, and TLvar used a deep transfer learning model to predict functional non-coding variants in human genome sequences. Despite the success of these methods in human disease prediction, their application in wheat genotype-phenotype studies is still limited.

[0004] At present, some methods have tried to use deep learning technology to predict the relationship between SNP and biotic stress resistance in wheat. For example, DL-GWAS and DNNGP use multi-omics data to predict complex plant traits. However, these methods have the following disadvantages: First, existing methods rely heavily on high-quality genotype data. Due to the large number of repetitive sequences and complex genome structure in the wheat genome, the acquisition and processing of high-quality genotype data is still a challenge; second, existing methods mainly rely on traditional k-mer encoding or simple sequence features in feature extraction, and fail to fully utilize the structure and context information of DNA sequences; third, deep learning models are usually regarded as "black boxes" and it is difficult to explain their prediction results, which limits researchers' trust in the model prediction results and further biological interpretation; fourth, existing methods are mainly trained and verified for specific species, lacking cross-species generalization capabilities, which limits their application in different crops. Summary of the invention

[0005] The purpose of the embodiments of the present invention is to provide a method for predicting wheat abiotic stress traits SNP based on an automated CNN model, aiming to solve the technical problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: The SNP prediction method for wheat abiotic stress traits based on the automated CNN model includes the following steps: Construct a high-quality dataset of SNPs associated with biotic stress traits; Multi-dimensional feature extraction is performed on the original DNA sequence of the high-quality data set through text-based features and structure-based features, wherein the text-based features include one-hot encoding, DNA2vec, and DNABERT; and the structure-based features include DNAShape; The extracted multidimensional features of DNA sequences are merged into a high-dimensional feature vector in the order of one-hot, DNA2vec, DNABERT, and DNAshape. The SHAP value of each feature is calculated using the random forest model. The importance of the input features is ranked using the SHAP value, and the key features for model training are determined based on the ranking results. Build a deep learning model with a search space. In the automatic architecture search framework of the deep learning model, each layer of the network structure contains a variety of different computing operations, including one-dimensional convolution operations of different sizes, maximum pooling operations of different pool sizes, average pooling operations, and identity mapping; During the model training process, the search space and parameters are initialized, and the sub-model architecture is generated based on the controller. The sub-model architecture is trained and evaluated. As the number of iterations increases, when the performance indicators of the validation set no longer improve significantly or reach the preset iteration limit, the search converges and the optimal architecture is obtained. The final architecture is fine-tuned with a smaller learning rate. The fine-tuned model is used to predict SNPs related to wheat biological stress.

[0007] Furthermore, the construction of the high-quality data set includes: Based on the collection of wheat-related genome-wide association study data in the GWAS Atlas database, including single nucleotide polymorphisms (SNPs) associated with biotic stress traits; The collected data were preprocessed to construct a high-quality mutation site dataset, which included: excluding SNPs with minor allele frequencies greater than 0.01; screening data based on the standard of P value less than 1e-5; generating negative samples based on sequence homology differences; constructing negative samples using functional annotation information; generating negative samples in combination with evolutionary conservation analysis; and using the CD-HIT tool to identify and merge highly similar sequences.

[0008] Furthermore, in the process of calculating the SHAP value of each feature using the random forest model, the SHAP value is calculated using the following formula: ; Where: represents the SHAP value of feature i; M represents the number of trees in the random forest; t represents the index of the tree in the random forest; ! represents the factorial symbol; |S| represents the number of features in subset S, |M| represents the total number of trees; C m represents the number of feature subsets in tree m; f represents the random forest prediction model; F represents all feature sets; S is the feature subset that does not contain feature i; f x (S) represents the prediction value of model f when only feature subset S is considered on input x; f x (S∪{i}) is the prediction of model f on input x considering feature subset S and feature i.

[0009] Furthermore, the step of initializing the search space and parameters and generating a sub-model architecture based on the controller includes: Determine the optional computing operations and hyperparameter ranges for each layer of the network; set the probability distribution of computing operation selection to π(a;θ), where a represents the operation combination and θ represents the controller parameter; use RNN as the controller model and initialize the weight W of the controller model rnn , bias b rnn , determine the learning rate α in the REINFORCE algorithm; In each round of search iteration t, the controller follows the current architecture state s t And input information generates each layer of network calculation operation probability distribution π t (a; θ), based on this probability sampling, each layer of operation is determined to build a sub-model architecture a t .

[0010] Furthermore, the step of training and evaluating the sub-model architecture includes: Sub-model architecturea t In the training set D train Training, select the operation forward propagation according to the input feature vector x, with cross entropy loss Calculating predictions With the true label y loss, back propagation updates the parameter ω, and the loss function is expressed as: ; Among them, L(ω) is the average cross entropy loss on the entire training set, which is The sum and average of After training, on the validation set D val Evaluation, in which the ability of the model to distinguish biotic stress-related SNPs was comprehensively evaluated by calculating the AUPR and AUC indicators; According to the sub-model architecture t Calculate the reward R on the validation set performance indicator t , expressed as: ; Among them, w1 and w2 represent weight coefficients, w1+w2=1; a t represents the sub-model architecture selected at time step t; The policy gradient is calculated using the REINFORCE algorithm, expressed as: ; In the formula, the initial b0=0.1, b t =γb t-1 +(1-γ)R t , γ is the smoothing factor; n is the total number of time steps; t is the time step; π t (a t ;θ) is the policy function, which means that under the parameter θ, the sub-model architecture a is selected t The probability distribution of Update the controller parameters θ so that the controller adjusts the architecture generation strategy based on reward feedback.

[0011] Furthermore, it also includes the step of using a multi-granularity causal structure learning method to identify the causal relationship between DNA sequences and functional structures, including: using a sparse autoencoder SAE to separately construct an encoder for each input feature, coarse-graining the microscopic DNA sequence information into macroscopic variables, and using a decoder to reconstruct the data; constructing a multi-layer perceptron MLP for the microscopic variables, using the original microscopic variables and the macroscopic variables obtained by SAE coarse-graining in series as input, selecting the mean square error as the error function, continuously updating the weights with the help of the back-propagation algorithm, collecting multi-granularity weighted adjacency matrices from the relevant parameter matrix, deeply analyzing the matrix elements, and understanding the causal structure of DNA features.

[0012] Furthermore, the step of coarse-graining the microscopic DNA sequence information into macroscopic variables comprises: The input data X with dimension d is The coded data Z is obtained by adding the operation and activation function σ. The coded data Z is expressed as: ; In the formula, d represents the dimension of the input data X, and σ represents the activation function; Decoder based on Z , activation function σ, And reactivate the function σ to reconstruct the data Y, which is expressed as: ; During training, adjust the parameter Wencoder and W decoder , adding sparsity regularization terms makes the encoding sparse, filters key information, and achieves effective coarse-graining and feature extraction.

[0013] Furthermore, the step of constructing a multi-layer perceptron MLP for micro variables includes: In MLP training, by adjusting the weight parameter W mlp Learning the mapping relationship between variables, where the mean square error is the error function, back propagation calculation Update; among them, Represents the predicted value of the multilayer perceptron MLP for the kth sample; represents the observed value of the kth sample; n mlp represents the number of samples in the MLP training set; L mse represents the mean square error, which is used to measure the difference between the MLP predicted value and the true value; From the MLP first parameter matrix Collect the multi-granularity weighted adjacency matrix S, whose elements s ij Indicates the direction of the strength of the causal relationship between variables; analyze the matrix S to understand the causal structure of DNA characteristics.

[0014] Compared with the prior art, the beneficial effects of the wheat abiotic stress trait SNP prediction method based on the automated CNN model of the present invention are: first, in data processing, the present invention constructs a multi-benchmark data set, generates negative samples through sequence homology differences, functional annotation information, and evolutionary conservation analysis, combines a variety of sequence lengths, and uses CD-HIT to remove redundancy, solves the problem of data shortage, improves diversity and representativeness, and enhances the generalization ability of the model; second, in feature extraction, the present invention integrates multidimensional features such as unique hot encoding, DNA2vec, DNABERT and DNAShape, extracts from multiple levels, makes full use of sequence structure and context information, and more accurately identifies related SNPs; third, in model interpretability, the present invention uses the SHAP method to rank the importance of input features and explain the prediction results of the model, thereby enhancing the interpretability of the model.

[0015] In summary, the present invention aims to provide an efficient, accurate and interpretable deep learning framework for predicting SNPs related to wheat biotic stress, thereby promoting the development of wheat stress resistance breeding and genomics research. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0017] Figure 1 The present invention is a flowchart of the wheat abiotic stress trait SNP prediction method based on the automated CNN model; Figure 2 It is a schematic diagram of SHAP feature selection in the SNP prediction method for wheat abiotic stress traits based on the automated CNN model of the present invention; DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0020] Please refer to Figure 1 and Figure 2 In one embodiment of the present invention, a method for predicting wheat abiotic stress traits SNP based on an automated CNN model is provided. The prediction method of the present invention comprises the following steps: S101. Build a high-quality dataset of SNPs associated with biotic stress traits; S102, extracting multidimensional features from the original DNA sequence of the high-quality data set through text-based features and structure-based features, wherein the text-based features include one-hot encoding, DNA2vec, and DNABERT; and the structure-based features include DNAShape; S103, the extracted multidimensional features of the DNA sequence are merged into a high-dimensional feature vector in the order of one-hot, DNA2vec, DNABERT, and DNAshape, the SHAP value of each feature is calculated using a random forest model, the importance of the input features is ranked using the SHAP value, and the key features for model training are determined based on the ranking results; S104, constructing a deep learning model with a search space, wherein in an automatic architecture search framework of the deep learning model, each layer of the network structure includes a plurality of different computing operations, the computing operations including one-dimensional convolution operations of different sizes, maximum pooling operations of different pool sizes, average pooling operations, and identity mapping; S105. During the model training process, the search space and parameters are initialized, a sub-model architecture is generated based on the controller, and the sub-model architecture is trained and evaluated. As the number of iterations increases, when the performance index of the validation set no longer improves significantly or reaches the preset iteration limit, the search converges and the optimal architecture is obtained. The final architecture is fine-tuned with a smaller learning rate. The fine-tuned model is used to predict SNPs related to wheat biological stress.

[0021] Furthermore, the construction of the high-quality data set includes: Based on the collection of wheat-related genome-wide association study data in the GWAS Atlas database, including single nucleotide polymorphisms (SNPs) associated with biotic stress traits; The collected data were preprocessed to construct a high-quality mutation site dataset, which included: excluding SNPs with minor allele frequencies greater than 0.01; screening data based on the standard of P value less than 1e-5; generating negative samples based on sequence homology differences; constructing negative samples using functional annotation information; generating negative samples in combination with evolutionary conservation analysis; and using the CD-HIT tool to identify and merge highly similar sequences.

[0022] Specifically, the present invention collects genome-wide association study (GWAS) data related to wheat from the GWAS Atlas database; these data include 35,749 single nucleotide polymorphisms (SNPs) associated with biotic stress traits, and these SNPs are associated with various stress factors such as bacteria, fungi, pests and viruses; The present invention constructs a high-quality mutation site dataset by subjecting the collected data to a series of processing steps: first, SNPs with minor allele frequencies greater than 0.01 are excluded, and then a standard of P value less than 1e-5 is used to ensure the quality of the dataset, ultimately obtaining a dataset containing 8,797 high-quality SNPs.

[0023] In addition, the present invention also collected 85 SNPs associated with wheat blast as an independent test set.

[0024] In order to generate enough negative samples, we proposed three solutions: Solution 1: Generate negative samples based on sequence homology differences. For positive sample SNPs that have been identified as being associated with biotic stress traits, search for regions with low sequence homology in the entire wheat genome database, and select SNPs in these regions as negative samples; specifically, set a homology threshold, such as a sequence similarity of less than 30%, and use bioinformatics alignment tools to find SNPs that meet this condition and are not annotated as being directly associated with biotic stress. This ensures that negative samples and positive samples have significant differences in sequence characteristics, which helps the model learn the key features to distinguish between biotic stress-related and unrelated SNPs.

[0025] Solution 2: Use functional annotation information to construct negative samples. Refer to the existing wheat genome functional annotation database and select SNPs that are annotated as having no relation to biotic stress functions such as abiotic stress response or basic metabolic processes as negative samples. For example, some SNPs are located in genes that are only involved in basic physiological processes such as photosynthesis and transport of common nutrients in wheat, and there is no literature reporting that they are related to the defense mechanisms of bacteria, fungi, pests, and viruses. These SNPs are included in the negative sample set. This approach ensures the independence of negative samples at the functional level, allowing the model to better focus on characteristic patterns related to biotic stress.

[0026] Option three: Generate negative samples in combination with evolutionary conservation analysis. Analyze the conserved sequence regions of wheat during evolution, and select SNPs that are relatively non-conserved in evolution and far away from known biotic stress-related loci as negative samples. Since key genes and SNPs related to biotic stress usually have a certain degree of conservatism during evolution to maintain the survival adaptability of the species, in contrast, SNPs in these non-conserved regions are likely to be weakly associated with biotic stress. Through evolutionary analysis tools such as multiple sequence alignment, non-conserved sequence segments in the wheat genome are identified, and SNPs are extracted from them to construct negative samples, which further enriches the diversity of negative samples and improves the generalization ability of the model.

[0027] Based on the above three schemes and combined with two sequence lengths (41 and 101nt), multiple benchmark datasets were constructed.

[0028] Finally, the present invention uses the CD-HIT tool to reduce redundancy and bias between samples. CD-HIT helps us remove samples with more than 70% pairwise sequence similarity by identifying and merging highly similar sequences, ensuring the diversity and representativeness of the data set.

[0029] Therefore, in data processing, the present invention constructs multiple benchmark data sets, generates negative samples through sequence homology differences, functional annotation information, and evolutionary conservation analysis, combines a variety of sequence lengths, uses CD-HIT to remove redundancy, solves the problem of data shortage, improves diversity and representativeness, and enhances the generalization ability of the model.

[0030] Furthermore, in step S102 of the present invention, multi-dimensional information processing refers to extracting features from the original DNA sequence through four encodings: text-based features (such as one-hot encoding, DNA2vec and DNABERT) and structure-based features (such as DNAShape). Among them: About one-hot encoding: DNA sequences are composed of four nucleotides: "A" (adenine), "T" (thymine), "G" (guanine) and "C" (cytosine), which can be encoded using a binary method to convert text information into numerical values. In this invention, we use one-hot encoding for DNA sequence fragments, mapping "A" to [1,0,0,0], "T" to [0,1,0,0], "C" to [0,0,1,0], and "G" to [0,0,0,1].

[0031] For example, a DNA sequence segment of length l can be encoded into an l×4 matrix.

[0032] About DNA2vec: The DNA2vec model encodes variable-length k-mers in DNA sequences into consistent vector representations. Traditional k-mer representations, usually encoded as one-hot vectors, suffer from the curse of dimensionality and isometry problems, making them less effective for advanced machine learning applications; the present DNA2vec uses a shallow two-layer neural network, pre-trained with wheat whole genome data, to generate continuous vector embeddings for k-mers of varying lengths (three to eight). This approach effectively addresses the limitations of one-hot encoding by embedding k-mers into a low-dimensional space where the distances between vectors reflect their biological similarities.

[0033] About DNABERT: DNABERT-2 is able to effectively decode genome sequences of multiple species. Traditional k-mer tags are replaced by byte pair encoding. This change solves the problems of information leakage and computational inefficiency. The model of the present invention combines attention with linear bias (ALiBi) and FlashAttention to handle long sequences and improve performance. Pre-trained on a diverse dataset of 135 species, DNABERT-2 demonstrates excellent accuracy and efficiency in genome analysis tasks, achieving excellent results with a significant reduction in the number of parameters and computing resources compared to previous models. Based on this representation, for a DNA sequence fragment of length l, the fragment can be represented as a 768-dimensional vector.

[0034] About DNAShape: It outlines the importance of DNA structure in protein binding and gene regulation by considering the extended flanking regions of DNA sequences without the need for extensive molecular simulations, focusing on base-based reading and shape reading mechanisms; DNA shape features are crucial for understanding protein-DNA interactions; the present invention extracts 13 structural features from DNA fragments, and the conformational flexibility of DNA is crucial; what is often overlooked in research is the fluctuation of shape features; experimental results show that including these values ​​significantly improves model performance; Therefore, a DNA sequence segment of length 41 can be encoded into a 527-dimensional vector.

[0035] Therefore, in terms of feature extraction, the present invention integrates multi-dimensional features such as one-hot encoding, DNA2vec, DNABERT and DNAShape, extracts from multiple levels, makes full use of sequence structure and context information, and more accurately identifies related SNPs.

[0036] In step S103 of the present invention, the importance of the input features is ranked using the SHAP method, and the results of the prediction model are explained. The data extracted by the four encoding methods in step S102 are merged into a row vector with 1623 dimensions, and the SHAP value is calculated using the random forest model; this will improve the interpretability of the model and provide a basis for feature selection; Therefore, in terms of model interpretability, the present invention uses the SHAP method to rank the importance of input features and explain the prediction results of the model, thereby enhancing the interpretability of the model.

[0037] Specifically, the present invention combines the multidimensional features of the DNA sequence extracted in step S102 into a 1623 high-dimensional feature vector in the order of one-hot, DNA2vec, DNABERT, and DNAshape, providing a comprehensive DNA sequence feature representation for the model; The SHAP (SHapley Additive exPlanations) method is a tool based on game theory to explain model predictions. It can assign a contribution value to each feature, which represents the average contribution of each feature to the model prediction. The core idea of ​​the SHAP method is to attribute changes in model predictions to the contribution of each feature, similar to the distribution to the total beneficiary in a cooperative game.

[0038] Specifically, in the process of calculating the SHAP value of each feature using the random forest model, the present invention uses the following formula to calculate the SHAP value: ; Where: represents the SHAP value of feature i; M represents the number of trees in the random forest; t represents the index of the tree in the random forest; ! represents the factorial symbol; |S| represents the number of features in subset S, |M| represents the total number of trees; C m represents the number of feature subsets in tree m; f represents the random forest prediction model; F represents all feature sets; S is the feature subset that does not contain feature i; f x (S) represents the prediction value of model f when only feature subset S is considered on input x; f x (S∪{i}) is the prediction of model f on input x considering feature subset S and feature i.

[0039] In the random forest model, each decision tree contributes to the final prediction result. The SHAP value calculates the average impact of each feature on the model prediction by considering its contribution in all decision trees. Specifically, for each tree in the random forest, the present invention calculates the impact of the presence or absence of feature i on the model prediction value. Then, the average of these impacts is taken, which is the SHAP value of feature i. A positive value indicates that the feature increases the prediction value, and a negative value indicates that the feature decreases the prediction value. The higher the SHAP value, the greater the contribution of the feature to the model prediction result, and the more valuable the information provided when distinguishing SNPs associated with biotic stress.

[0040] The present invention ranks the 1623 features by importance according to the calculated SHAP values. After ranking, it can be clearly seen which features have a greater impact on the model prediction of BST-associated SNPs and which features have a relatively smaller impact.

[0041] The top-ranked features are selected as key features for subsequent model training.

[0042] In the present invention, 100 key features are selected, which play an important role in improving the prediction ability of the model. For example, in the Random_41_100 data set, the model trained using these key features performs well in various evaluation indicators (such as F1 score, AUPR and precision, etc.), and has obvious advantages over other data sets and feature combinations. By selecting key features, the influence of redundant and irrelevant features can be reduced, so that the model can pay more attention to the information that is really important for predicting BST-associated SNPs during training, thereby improving the efficiency of model training, enhancing the accuracy and robustness of the model, avoiding problems such as overfitting, and finally obtaining an optimized feature set that is more suitable for predicting wheat biological stress-related SNPs for subsequent model training and evaluation.

[0043] Furthermore, the step of initializing the search space and parameters and generating a sub-model architecture based on the controller includes: Determine the optional computing operations for each layer of the network (such as Conv1D with a kernel size of k∈{3, 4, 6, 8}, Maxpool, Avgpool, and Identity with a pool size of p∈{3, 4}) and the range of hyperparameters; set the probability distribution of computing operation selection to be π(a; θ), where a represents the operation combination and θ represents the controller parameter; use RNN as the controller model and initialize the weight W of the controller model rnn , bias b rnn , determine the learning rate α in the REINFORCE algorithm, which affects the learning speed and stability of the controller; In each round of search iteration t, the controller follows the current architecture state st And input information generates each layer of network calculation operation probability distribution π t (a; θ), based on this probability sampling, each layer of operation is determined to build a sub-model architecture a t ; For example, the probability of selecting different operations at a certain layer is p 1，t to p 4，t And the sum is 1, and the probability sampling is used to determine each layer of operation to build the sub-model architecture a t , achieving extensive exploration of the search space.

[0044] Furthermore, the step of training and evaluating the sub-model architecture includes: Sub-model architecturea t In the training set D train Training, select the operation forward propagation according to the input feature vector x, with cross entropy loss Calculating predictions With the true label y loss, back propagation updates the parameter ω, and the loss function is expressed as: ; Among them, L(ω) is the average cross entropy loss on the entire training set, which is The sum and average of After training, on the validation set D val Evaluation, in which the ability of the model to distinguish biotic stress-related SNPs was comprehensively evaluated by calculating the AUPR and AUC indicators; According to the sub-model architecture t Calculate the reward R on the validation set performance indicator t , expressed as: ; Among them, w1 and w2 represent weight coefficients, w1+w2=1; a t represents the sub-model architecture selected at time step t; The policy gradient is calculated using the REINFORCE algorithm, expressed as: ; In the formula, the initial b0=0.1, b t =γb t-1 +(1-γ)R t , γ is the smoothing factor; n is the total number of time steps; t is the time step; π t (a t ;θ) is the policy function, which means that under the parameter θ, the sub-model architecture a is selected t The probability distribution of Update the controller parameters θ so that the controller adjusts the architecture generation strategy based on reward feedback.

[0045] Furthermore, in the embodiment of the present invention, different sub-model architectures share some (such as convolutional layer) calculation graphs and parameters ω shared , such as the weight of the basic convolutional layer ω conv The initialization is the same and the intermediate calculation results can be shared, and the parameters are only adjusted in the part where the architecture changes; this reduces the consumption of computing resources, accelerates the architecture search, and increases the chance of finding the optimal architecture. At the same time, it reduces memory usage, improves training efficiency, and helps the model converge quickly.

[0046] In the present invention, the controller adjusts the strategy based on experience during iteration, the sub-model architecture performance is optimized, and the model explores a better architecture combination; as the iteration increases, the model explores the search space more fully and the performance is improved. When the performance index of the validation set is no longer significantly improved or reaches the preset iteration upper limit I, the search converges and the optimal architecture a* is obtained, which has good performance in processing wheat biological stress-related SNP data and can effectively extract feature prediction; the present invention uses a smaller learning rate α f In the complete training set D train Fine-tune the final architecture a* and fine-tune the convolutional layer weights , fully connected layer parameters etc., adapting to data characteristics to improve accuracy and stability; the fine-tuned model is used to predict wheat biological stress-related SNPs, the new data is input and processed through multidimensional feature encoding, and the prediction results (whether it is related and the probability) are output, which provides information for research and helps to understand the stress resistance mechanism and variety breeding.

[0047] Furthermore, the prediction method of the present invention also includes the step of using a multi-granularity causal structure learning method to identify the causal relationship between the DNA sequence and the functional structure, including: Sparse autoencoder (SAE) is used to construct an encoder for each input feature, coarse-graining the microscopic DNA sequence information into macroscopic variables, and the decoder is used to reconstruct the data. A multi-layer perceptron (MLP) is constructed for micro variables. The original micro variables and the macro variables obtained by SAE coarse-graining are connected in series as input. The mean square error is selected as the error function. The weights are continuously updated with the help of the back-propagation algorithm. Multi-granularity weighted adjacency matrices are collected from the relevant parameter matrix. The matrix elements are deeply analyzed to understand the causal structure of DNA characteristics.

[0048] The present invention applies the MgCSL method to identify the causal relationship between DNA sequence and functional structure, understand biological processes from a causal level, provide new insights into gene regulation and genetic mechanisms, and break through the limitations of traditional correlation research.

[0049] The present invention uses SAE to coarse-grain microscopic DNA sequence information into macroscopic variables. To ensure the independence of microscopic variables, an encoder is constructed separately for each input. The step of coarse-graining microscopic DNA sequence information into macroscopic variables includes: The input data X with dimension d is The coded data Z is obtained by adding the operation and activation function σ. The coded data Z is expressed as: ; In the formula, d represents the dimension of the input data X, and σ represents the activation function; Decoder based on Z , activation function σ, And reactivate the function σ to reconstruct the data Y, which is expressed as: ; During training, adjust the parameter W encoder and W decoder , adding sparsity regularization terms makes the encoding sparse, filters key information, and achieves effective coarse-graining and feature extraction.

[0050] Furthermore, the present invention constructs an MLP for each micro variable, connecting the SAE coarse-grained macro variable Z with the original micro variable X in series ( ) as input and use micro and macro information to comprehensively explore causal relationships.

[0051] The step of constructing a multi-layer perceptron MLP for micro variables includes: In MLP training, by adjusting the weight parameter W mlp Learning the mapping relationship between variables, where the mean square error is the error function, back propagation calculation Update; among them, Represents the predicted value of the multilayer perceptron MLP for the kth sample; represents the observed value of the kth sample; n mlp represents the number of samples in the MLP training set; L mse Represents the mean square error, which is used to measure the difference between the MLP predicted value and the true value; From the MLP first parameter matrix Collect the multi-granularity weighted adjacency matrix S, whose elements s ij Indicates the direction of the strength of the causal relationship of the variable; analyzes the matrix S to understand the causal structure of DNA characteristics, such as the macroscopic effects of certain structural characteristics (ProT and Buckle, etc.) on other characteristics and on gene expression and resistance to biotic stresses, providing clues for revealing gene regulatory mechanisms, helping to understand the genetic mechanisms related to biotic stress in wheat from a causal perspective, and laying the foundation for subsequent research and applications.

[0052] The present invention evaluates each model through an independent data set, and the results show that the A-CNN model of the present invention has outstanding performance. In terms of key indicators such as F1 score, AUPR, AUC and Accuracy, the A-CNN model shows certain advantages or competitiveness compared with models such as DL-GWAS, TLvar and DNNGP, which strongly proves its effectiveness and reliability in predicting SNPs related to wheat biotic stress traits.

[0053] In summary, the present invention aims to provide an efficient, accurate and interpretable deep learning framework for predicting SNPs related to wheat biotic stress, thereby promoting the development of wheat stress resistance breeding and genomics research.

[0054] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A method for predicting wheat abiotic stress traits SNP based on an automated CNN model, characterized in that: The following steps are involved: Construct a high-quality dataset of SNPs associated with biotic stress traits; Multi-dimensional feature extraction is performed on the original DNA sequence of the high-quality data set through text-based features and structure-based features, wherein the text-based features include one-hot encoding, DNA2vec, and DNABERT; and the structure-based features include DNAShape; The extracted multidimensional features of DNA sequences are merged into a high-dimensional feature vector in the order of one-hot, DNA2vec, DNABERT, and DNAshape. The SHAP value of each feature is calculated using the random forest model. The importance of the input features is ranked using the SHAP value, and the key features for model training are determined based on the ranking results. Build a deep learning model with a search space. In the automatic architecture search framework of the deep learning model, each layer of the network structure contains a variety of different computing operations, including one-dimensional convolution operations of different sizes, maximum pooling operations of different pool sizes, average pooling operations, and identity mapping; During the model training process, the search space and parameters are initialized, and the sub-model architecture is generated based on the controller. The sub-model architecture is trained and evaluated. As the number of iterations increases, when the performance indicators of the validation set no longer improve significantly or reach the preset iteration limit, the search converges and the optimal architecture is obtained. The final architecture is fine-tuned with a smaller learning rate. The fine-tuned model is used to predict SNPs related to wheat biological stress.

2. The method for predicting wheat abiotic stress traits SNP based on an automated CNN model according to claim 1, characterized in that: The construction of the high-quality dataset includes: Based on the collection of wheat-related genome-wide association study data in the GWAS Atlas database, including single nucleotide polymorphisms (SNPs) associated with biotic stress traits; The collected data were preprocessed to construct a high-quality mutation site dataset, which included: excluding SNPs with minor allele frequencies greater than 0.01; screening data based on the standard of P value less than 1e-5; generating negative samples based on sequence homology differences; constructing negative samples using functional annotation information; generating negative samples in combination with evolutionary conservation analysis; and using the CD-HIT tool to identify and merge highly similar sequences.

3. The method for predicting wheat abiotic stress traits SNP based on the automated CNN model according to claim 2, characterized in that: In the process of calculating the SHAP value of each feature using the random forest model, the SHAP value is calculated using the following formula: ； Where: represents the SHAP value of feature i; M represents the number of trees in the random forest; t represents the index of the tree in the random forest; ! represents the factorial symbol; |S| represents the number of features in subset S, |M| represents the total number of trees; C m represents the number of feature subsets in tree m; f represents the random forest prediction model; F represents all feature sets; S is the feature subset that does not contain feature i; f x (S) represents the prediction value of model f when only feature subset S is considered on input x; f x (S∪{i}) is the prediction of model f on input x considering feature subset S and feature i.

4. The method for predicting wheat abiotic stress traits SNP based on an automated CNN model according to claim 3, characterized in that: The step of initializing the search space and parameters and generating a sub-model architecture based on the controller includes: Determine the optional computational operations and hyperparameter ranges for each network layer; Set the probability distribution of the calculation operation selection to π(a;θ), where a represents the operation combination and θ represents the controller parameter; use RNN as the controller model and initialize the weight W of the controller model rnn , bias b rnn , determine the learning rate α in the REINFORCE algorithm; In each round of search iteration t, the controller follows the current architecture state s t And input information generates each layer of network calculation operation probability distribution π t (a; θ), based on this probability sampling, each layer of operation is determined to build a sub-model architecture a t .

5. The method for predicting wheat abiotic stress traits SNP based on the automated CNN model according to claim 4, characterized in that: The steps of training and evaluating the sub-model architecture include: Sub-model architecturea t In the training set D train Training, select the operation forward propagation according to the input feature vector x, with cross entropy loss Calculating predictions With the true label y loss, back propagation updates the parameter ω, and the loss function is expressed as: ； Among them, L(ω) is the average cross entropy loss on the entire training set, which is The sum and average of After training, on the validation set D val Evaluation, by calculating AUPR and AUC indicators, the ability of the model to distinguish biotic stress-related SNPs was comprehensively evaluated; According to the sub-model architecture t Calculate the reward R on the validation set performance indicator t , expressed as: ； Among them, w1 and w2 represent weight coefficients, w1+w2=1; a t represents the sub-model architecture selected at time step t; The policy gradient is calculated using the REINFORCE algorithm, expressed as: ； In the formula, b is the reward exponential moving average, initially b0=0.1, b t =γb t-1 +(1-γ)R t , γ is the smoothing factor; n is the total number of time steps; t is the time step; π t (a t ;θ) is the policy function, which means that under the parameter θ, the sub-model architecture a is selected t The probability distribution of Update the controller parameters θ so that the controller adjusts the architecture generation strategy based on reward feedback.

6. The method for predicting wheat abiotic stress traits SNP based on an automated CNN model according to any one of claims 2 to 5, characterized in that: The method also includes the steps of using a multi-granularity causal structure learning method to identify the causal relationship between DNA sequence and functional structure, including: A sparse autoencoder (SAE) is used to construct an encoder for each input feature, and the microscopic DNA sequence information is coarse-grained into macroscopic variables. The decoder is used to reconstruct the data. A multi-layer perceptron (MLP) is constructed for the microscopic variables, and the original microscopic variables and the macroscopic variables obtained by SAE coarse-graining are connected in series as input. The mean square error is selected as the error function, and the weights are continuously updated with the help of the back-propagation algorithm. Multi-granularity weighted adjacency matrices are collected from the relevant parameter matrices, and the matrix elements are deeply analyzed to understand the causal structure of DNA features.

7. The method for predicting wheat abiotic stress traits SNP based on an automated CNN model according to claim 6, characterized in that: The step of coarse-graining the microscopic DNA sequence information into macroscopic variables comprises: The input data X with dimension d is The coded data Z is obtained by adding the operation and activation function σ. The coded data Z is expressed as: ； In the formula, d represents the dimension of the input data X, and σ represents the activation function; Decoder based on Z , activation function σ, And reactivate the function σ to reconstruct the data Y, which is expressed as: ； During training, adjust the parameter W encoder and W decoder , adding sparsity regularization terms makes the encoding sparse, filters key information, and achieves coarse-graining and feature extraction.

8. The method for predicting wheat abiotic stress traits SNP based on the automated CNN model according to claim 7, characterized in that: The steps of constructing a multi-layer perceptron MLP for micro variables include: In MLP training, by adjusting the weight parameter W mlp Learning the mapping relationship between variables, where the mean square error is used as the error function and the back propagation calculation Update, the error function is expressed as: ； in, Represents the predicted value of the multilayer perceptron MLP for the kth sample; represents the observed value of the kth sample; n mlp represents the number of samples in the MLP training set; L mse represents the mean square error, which is used to measure the difference between the MLP predicted value and the true value; From the MLP first parameter matrix Collect the multi-granularity weighted adjacency matrix S, whose elements s ij Indicates the direction of the strength of the causal relationship between variables; analyze the matrix S to understand the causal structure of DNA characteristics.

Images