Tomato salt tolerance prediction method and system based on deep learning

Abstract

The present application relates to the technical field of artificial intelligence and deep learning, in particular to a tomato salt tolerance prediction method and system based on deep learning, specifically as follows: obtaining a tomato salt tolerance genome, inputting it into a tomato salt tolerance prediction system based on deep learning for detection, sequentially passing through a salt-tolerant small sample gene recombination module, an epigenetic gene multi-granularity mining module, a heterogeneous salt-tolerant gene feature integration module, a salt-tolerant feature point fusion module and a tomato salt tolerance gene prediction module, and calculating the probability of having relevant tomato salt tolerance genes in the input tomato salt tolerance genome. The present application improves the prediction accuracy and generalization ability of salt tolerance genes under complex genetic background, and provides an intelligent prediction method for tomato salt-tolerant variety breeding.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and deep learning, and in particular to a tomato salt tolerance prediction method and system based on deep learning. BACKGROUND

[0002] In view of the problem of shortage of arable land resources in coastal areas, and the salinization of part of the soil due to seawater intrusion and human activities, with the rapid development of genomics and high-throughput phenotype technology, using deep learning method to predict tomato salt tolerance traits has become an important direction for breeding of salt-tolerant tomato new varieties. Traditional tomato salt tolerance screening mainly relies on field trials and physiological and biochemical index determination, and evaluates salt tolerance grade by measuring parameters such as plant height inhibition rate, chlorophyll content and ion balance. Such methods not only consume a lot of manpower and material resources, but also are easily disturbed by the environment, leading to errors in phenotype identification, and it is difficult to realize rapid screening of large-scale germplasm resources. In recent years, the phenotype-genome joint analysis technology based on deep learning has broken through the traditional limitations, and through neural network analysis of multi-modal data such as hyperspectral imaging and root dynamic growth video, the salt tolerance of tomato is predicted. This technology significantly shortens the breeding cycle and provides precise decision support for the design and breeding of salt-tolerant tomatoes.

[0003] Deep learning technology has been applied in the field of plant and animal gene expression prediction, and the prediction of tomato salt tolerance is also a key research direction. However, the existing prediction methods still have many defects: first, high-quality labeled data of tomato salt tolerance genes are scarce, and the expression of salt tolerance genes of plant samples grown in different environments is different, and the existing prediction methods cannot effectively utilize gene samples under the condition of few samples and many differences; second, the genome contains a large amount of data, and the feature dimension of epigenome data is huge, how to capture the required epigenes in high-dimensional data is also a hot research topic at present; finally, related salt tolerance genes are jointly regulated by multiple factors such as expressed genome sequence, epigenetic state and genome spatial position information, and the existing prediction methods cannot effectively integrate the above heterogeneous data.

[0004] Therefore, the present application proposes a tomato salt tolerance prediction method and system based on deep learning to solve the above problems. SUMMARY

[0005] The present application is aimed at the deficiencies of the prior art, and a tomato salt tolerance prediction method and system based on deep learning are developed, which improves the prediction accuracy and generalization ability of salt tolerance genes in complex genetic background, and provides an intelligent prediction method for tomato salt-tolerant variety breeding.

[0006] In one aspect, the technical scheme for solving the technical problem of the present application is a tomato salt tolerance prediction method based on deep learning. A tomato salt tolerance prediction model is constructed, and a tomato salt tolerance genome to be detected is input into the model for salt tolerance prediction. The operation in the model is as follows:

[0007] S1, a salt-tolerant small sample gene recombination module is constructed, and a tomato salt-tolerance genome to be detected is input After multi-level feature capture and attention-driven feature recombination operation, small sample gene recombination features are output ;

[0008] S2, an epigenetic multi-granularity mining module is constructed, and small sample gene recombination features are input After multi-level epigenetic feature positioning and multi-granularity feature mining operation, epigenetic mining features are output ;

[0009] S3, a heterogeneous salt-tolerant gene feature integration module is constructed, and epigenetic mining features are input After the context integration of heterogeneous feature information and the capture of long-distance spatial feature dependence operation of the module, heterogeneous features are output ;

[0010] S4, a salt-tolerant feature point fusion module is constructed, and small sample gene recombination features , epigenetic mining features and heterogeneous features are input, and after salt-tolerant gene multi-element feature fusion operation, salt-tolerant multi-element features are output ;

[0011] S5, a tomato salt-tolerance gene prediction module is constructed, and salt-tolerant multi-element features are input , and the probability of having related tomato salt-tolerance genes is predicted ;

[0012] S6, the parameters in the tomato salt-tolerance prediction model are optimized by a loss function, and a trained tomato salt-tolerance prediction model is obtained.

[0013] The operation in the salt-tolerant small sample gene recombination module is as follows:

[0014] (1) Multi-level feature capture operation:

[0015] A salt-tolerant small sample gene recombination module is established, and a tomato salt-tolerance genome to be detected is input into the salt-tolerant small sample gene recombination module Multi-level feature capture operation is performed on the genome , which is obtained through the deep separable convolution layer, batch normalization layer and non-linear activation function layer in the module, to obtain features ;

[0016] Construct a three-layer feature capture layer, each consisting of a convolutional layer, a depthwise separable convolutional layer, and... The activation function is composed of different kernel sizes in each depth-separable convolutional layer, and features at different scales are captured by setting different kernel sizes.

[0017] Features The input is fed into a three-layer feature capture layer, and the output features are obtained after passing through the first feature capture layer. Then the features Input to the second feature capture layer outputs features Then the features Input to the third feature capture layer outputs features ;

[0018] (2) Attention-driven feature reorganization operation:

[0019] Features The input is fed into the feature reconstruction network and undergoes an attention-driven feature reorganization operation, specifically used... Feature decomposition operation will decompose the features Cut into Feature blocks of varying sizes, focusing on local features, and then... Features are obtained by compressing spatial information through a global average pooling layer. ;

[0020] Features Feature transformation is performed after a fully connected layer, and then... Activation function generates attention weight features ;

[0021] Attention weight features use The feature folding operation reconstructs the features to generate reconstructed features. ;

[0022] (3) Feature fusion operation:

[0023] Features ,feature and reconstruction features After feature concatenation, convolutional layers, and... Nonlinear activation function layers are used to obtain small sample gene recombination features. .

[0024] The specific operations in the epigene multi-granularity mining module are as follows:

[0025] (1) Feature preprocessing operations:

[0026] The epigenetic multi-granularity mining module is established, and the small sample gene recombination features are input into the epigenetic multi-granularity mining module, and the convolution layer, batch normalization layer and non-linear activation function in the module are used to output the optimized intermediate features

[0027] (2) Double-branch multi-scale feature capturing operation:

[0028] The features are input into two parallel multi-scale salt-tolerant gene feature capturing networks;

[0029] The features are input into the first branch of the network, and the features are input into the deep separable convolution layer, activation function and full connection layer, and high-dimensional features are output

[0030] Meanwhile, the features are input into the second branch of the network, which is composed of a multi-level gene capturing module and a multi-scale gene mining module;

[0031] 1) In the multi-level gene capturing module, the features are input into the average pooling layer, convolution layer and activation function, and the features are output ; The features are input into two deep separable convolution layers with different convolution kernel sizes and activation function, and the features are output ; The features are input into the convolution layer and activation function, and the features are output ;

[0032] 2) In the multi-scale gene mining module, multi-granularity feature mining operation is performed;

[0033] Specifically, the features are input into the deep separable convolution layer, and the features are output

[0034] The multi-granularity feature mining operation performs multi-level feature extraction on the features , and each layer structure is a deep separable convolution layer and activation function, and the convolution kernel size of each deep separable convolution layer is different; After feature extraction on the features , the features are output , and then based on the features , the features are extracted and output​​​​​​​​​​​​​ , based on the feature , feature extraction is performed, and a feature is output , based on the feature , feature extraction is performed, and a feature is output ;

[0035] Then, the features extracted in each layer are spliced and output through a convolutional layer , , , , and a feature is output ;

[0036] (3) Feature integration operation:

[0037] The feature is fused with the feature using an element-wise multiplication operation, and a feature is output through a convolutional layer ;

[0038] The feature is spliced with the feature , and a feature is output through a convolutional layer and an activation function . .

[0039] The operations in the heterogeneous salt-tolerant gene feature integration module are as follows:

[0040] (1) Establish a heterogeneous salt-tolerant gene feature integration module, and input the epigenetic gene mining feature into the heterogeneous salt-tolerant gene feature integration module, which includes a global salt-tolerance feature mining module and a salt-tolerance feature correlation construction module;

[0041] 1) Global salt-tolerance feature mining module:

[0042] The feature passes through the global salt-tolerance feature mining module of the heterogeneous salt-tolerant gene feature integration module, which captures long-distance dependent features of heterogeneous data through a long-distance spatial feature dependency capturing operation, and the feature passes through a depth separable convolutional layer, a convolutional layer, a batch normalization layer, and an activation function, and a feature is output ;

[0043] The feature is input into a depth separable convolutional layer and a convolutional layer, and the resulting feature is added to the feature element by element, and a feature is output ;

[0044] 2) Salt tolerance feature correlation construction module:

[0045] The output features of the global salt tolerance feature mining module are input into the salt tolerance feature correlation construction module, which integrates the context information of heterogeneous features through a heterogeneous feature information context integration operation;

[0046] features After passing through the deep separable convolution layer, batch normalization layer and activation function of the module, the obtained features and features are spliced to output features , features After passing through the global average pooling layer, the fully connected layer, non-linear activation function, fully connected layer and activation function, the output feature ;

[0047] features and features are added element by element to output features ;

[0048] features After passing through the fully connected layer, the obtained features and features are spliced to output features ;

[0049] (2) Feature fusion operation:

[0050] features After passing through the convolution layer, the deep separable convolution layer and activation function, the output feature ;

[0051] features and features are added element by element to output heterogeneous features .

[0052] The operations in the tomato salt tolerance gene prediction module are as follows:

[0053] The tomato salt tolerance gene prediction module is established, and the salt tolerance multi-feature is input into the tomato salt tolerance gene prediction module. The salt tolerance multi-feature After passing through the convolution kernel, feature flattening operation, fully connected layer and activation function, the output predicted related tomato salt tolerance gene probability , according to the probability The size of the sample determines whether the tomato being tested has a salt-tolerant gene, and the final prediction result is output.

[0054] The specific operations in the salt-tolerant feature point fusion module are as follows:

[0055] (1) Initial integration:

[0056] Establish a salt tolerance feature fusion module to integrate small sample gene recombination features. Epigene mining features and heterogeneous features The input is fed into the salt-tolerant feature point fusion module, where it undergoes feature concatenation, convolutional layers, and... A nonlinear activation function is used for initial fusion, and the output is the initially fused features. ;

[0057] (2) Step-by-step operation of bi-branch micro-features:

[0058] feature The features are captured by a parallel salt tolerance gene fusion feature capture network, which captures small salt tolerance gene features, namely global small features and local small features.

[0059] 1) Features The first branch of the salt tolerance gene fusion feature capture network is input, and the features are processed through a global average pooling layer. Compression is performed, and local details are captured through depthwise separable convolution to obtain global features;

[0060] Specifically, the features After global average pooling layer, convolutional layer, batch normalization layer and Nonlinear activation function, output features Then the features After passing through depthwise separable convolutional layers, convolutional layers, and fully connected layers, global minute features are output. ;

[0061] 2) Features The second branch of the input salt tolerance gene fusion feature capture network retains salient features through max pooling layers and captures mesoscale details through depthwise separable convolutions.

[0062] Specifically, the features After max pooling layer, convolutional layer and Activation function, output features ; Features After two depth-separable convolutional layers, Activation functions and fully connected layers output localized minute features. ;

[0063] characteristics characteristics after using an element-by-element addition operation, followed by a convolution layer, an activation function, a fully connected layer, and an activation function, outputting salt-tolerant multi-element characteristics .

[0064] S6 is as follows:

[0065] According to the prediction result and the real label of the tomato salt-tolerant genome to be detected, a binary cross-entropy loss function is used to train the tomato salt-tolerance prediction model, and the parameters of the model in the training process are optimized and adjusted according to the loss function result, to obtain the trained tomato salt-tolerance prediction model.

[0066] On the other hand, the application also provides a tomato salt-tolerance prediction system based on deep learning, which comprises modules for executing processing instructions for each step in a tomato salt-tolerance prediction method based on deep learning, specifically comprising the following modules:

[0067] Genome data input module: receiving tomato salt-tolerant genome data to be detected, and pre-processing the data;

[0068] Salt-tolerant small sample gene recombination module: performing multi-level feature capture and attention-driven feature recombination on the input genome data, and outputting small sample gene recombination features;

[0069] Epigenetic multi-granularity mining module: performing multi-level epigenetic feature positioning and multi-granularity feature mining on the small sample gene recombination features, and outputting epigenetic mining features;

[0070] Heterogeneous salt-tolerant gene feature integration module: performing heterogeneous feature information context integration and long-distance spatial feature dependence capture on the epigenetic mining features, and outputting heterogeneous features;

[0071] Salt-tolerant feature point fusion module: fusing small sample gene recombination features, epigenetic mining features, and heterogeneous features, and outputting salt-tolerant multi-element features;

[0072] Salt-tolerance prediction and model optimization module: predicting tomato salt-tolerance gene probability based on salt-tolerant multi-element fusion features, and optimizing model parameters through a loss function, and outputting a final prediction result.

[0073] The effects provided in the summary are only the effects of the embodiments, not all the effects of the application. The above technical solutions have the following advantages or beneficial effects:

[0074] The present application aims at the problems of the prior art that the salt tolerance gene prediction method cannot effectively utilize the salt tolerance gene sample characteristics under the condition of few samples and many differences, accurately locate the salt tolerance related gene characteristics of high-dimensional gene data, and integrate heterogeneous data characteristics, and proposes a tomato salt tolerance prediction method and system based on deep learning;

[0075] Specifically, for the problem that the prediction model cannot effectively utilize the salt tolerance gene sample characteristics under the condition of few samples and many differences, the present application constructs a salt tolerance small sample gene recombination module, and through multi-level feature capture and attention-driven feature recombination operation, the diversity of salt tolerance gene characteristics in few samples and the cross-sample recombination salt tolerance gene characteristics can be effectively enhanced.

[0076] For the problem of how to locate the tomato related salt tolerance epigenetic gene characteristics of high-dimensional gene data, an epigenetic gene multi-granularity mining module is constructed, and through multi-level epigenetic feature positioning and multi-granularity feature mining operation, the related salt tolerance gene positioning problem under high-dimensional massive gene characteristics is effectively solved.

[0077] For the problem that the existing prediction model method cannot effectively integrate heterogeneous data characteristics because the salt tolerance gene in the gene is jointly regulated by multiple factors such as expression genome sequence, epigenetic state and genome spatial position information, the present application constructs a heterogeneous salt tolerance gene feature integration module, and through heterogeneous feature information context integration and long-distance spatial feature dependence capture operation, the problem of integrating heterogeneous gene data characteristics is solved.

[0078] The present application significantly improves the prediction accuracy and generalization ability of salt tolerance genes under complex gene background, provides an intelligent prediction method for tomato salt tolerance variety breeding, and provides reliable technical support for saline-alkali soil ecological restoration and efficient agricultural breeding. BRIEF DESCRIPTION OF DRAWINGS

[0079] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application.

[0080] Figure 1 The present application is a method flowchart.

[0081] Figure 2 The present application is a structure diagram of the salt tolerance small sample gene recombination module.

[0082] Figure 3 The present application is a structure diagram of the epigenetic gene multi-granularity mining module.

[0083] Figure 4 The present application is a structure diagram of the heterogeneous salt tolerance gene feature integration module.

[0084] Figure 5A structural schematic diagram of a salt-tolerant feature point fusion module.

[0085] Figure 6 A structural schematic diagram of a tomato salt-tolerance gene prediction module. DETAILED DESCRIPTION

[0086] In order to clearly illustrate the technical features of the scheme, the present application will be described in detail below through specific embodiments, and in conjunction with the accompanying drawings.

[0087] Embodiment 1

[0088] As shown in the figure, a tomato salt-tolerance prediction method based on deep learning constructs a tomato salt-tolerance prediction model, inputs the tomato salt-tolerance genome to be detected into the model for salt-tolerance prediction, including the following steps: Figure 1 S1, a salt-tolerant small sample gene recombination module is constructed, and a tomato salt-tolerance genome to be detected is input

[0089] , after multi-level feature capture and attention-driven feature recombination operation, small sample gene recombination features are output ;

[0090] S2, an epigenetic gene multi-granularity mining module is constructed, and small sample gene recombination features are input , after multi-level epigenetic feature positioning and multi-granularity feature mining operation, epigenetic gene mining features are output ;

[0091] S3, a heterogeneous salt-tolerant gene feature integration module is constructed, and epigenetic gene mining features are input , after the module's heterogeneous feature information context integration and long-distance spatial feature dependence capture operation, heterogeneous features are output ;

[0092] S4, a salt-tolerant feature point fusion module is constructed, and small sample gene recombination features , epigenetic gene mining features and heterogeneous features are input, after salt-tolerant gene multi-element feature fusion operation, salt-tolerant multi-element features are output ;

[0093] S5, a tomato salt-tolerance gene prediction module is constructed, and salt-tolerant multi-element features are input , and the probability of having relevant tomato salt-tolerance genes is predicted ;

[0094] S6, the parameters in the tomato salt-tolerance prediction model are optimized through a loss function, and a trained tomato salt-tolerance prediction model is obtained.

[0095] ​In a specific implementation, to address the issues of insufficient feature extraction and redundant information interference in small sample gene data, a salt-tolerant small sample gene recombination module is constructed. This module enhances the expression of key salt-tolerant gene features through multi-level feature capture and attention-driven feature recombination, such as... Figure 2 As shown, S1 is as follows:

[0096] (1) Multi-level feature capture operation:

[0097] Establish a small-sample salt-tolerant gene recombination module to incorporate the tomato salt-tolerant genome. Input into the salt-tolerant small sample gene recombination module for genome processing Multi-level feature capture operations are performed, and these operations are processed by convolutional kernels of size [size missing] in this module. Depth-separable convolutional layers, batch normalization layers, and Nonlinear activation function layers yield features ;

[0098] feature After passing through a convolution kernel with a size of The convolutional layer and kernel size are Depth-separable convolutional layers and Activation function to obtain features ;

[0099] feature After passing through a convolution kernel with a size of The convolutional layer and kernel size are Depth-separable convolutional layers and Activation function to obtain features ;

[0100] feature After passing through a convolution kernel with a size of The convolutional layer and kernel size are Depth-separable convolutional layers and Activation function to obtain features ;

[0101] The calculation formulas involved in the above process are as follows:

[0102] ,

[0103] ,

[0104] ,

[0105] ,

[0106] in, Indicates the batch normalization layer. represents a convolutional layer with a convolution kernel size of represents a convolutional layer with a convolution kernel size of represents a depth separable convolutional layer with a convolution kernel size of represents a depth separable convolutional layer with a convolution kernel size of

[0107] The depth separable convolution with different convolution kernels captures gene features of different sizes, splits the standard convolution into channel-wise convolution and point convolution, reduces the amount of calculation, is suitable for small sample scenarios, and can avoid overfitting caused by too many parameters.

[0108] (2) Attention-driven feature reconstruction operation:

[0109] The feature is input to the feature reconstruction network and subjected to the attention-driven feature reconstruction operation to suppress noise interference of irrelevant gene fragments, solve the problem of too high proportion of redundant feature information, and clearly mark the core gene region related to salt tolerance. The network includes the following operations:

[0110] The feature is cut into a feature block with a size of using the feature disassembly operation, focusing on local features, and the feature is compressed for spatial information through a global average pooling layer to obtain feature .

[0111] The feature is subjected to a feature transformation operation through a fully connected layer, and an attention weight feature is generated using an activation function, thereby highlighting the key region of the salt-tolerant gene.

[0112] The attention weight feature is subjected to a feature folding operation to reconstruct the feature , thereby achieving weighted enhancement of the important region.

[0113] The calculation formula involved in the above process is:

[0114] ,

[0115] ,

[0116] ,

[0117] wherein represents a global average pooling layer,​​​​​​​ denotes a full connection layer, denotes using The feature disassembling operation disassembles the feature into feature blocks of a certain size, denotes an activation function;

[0118] (3) Feature fusion operation:

[0119] The feature , the feature and the reconstructed feature are subjected to a feature splicing operation, a convolution layer with a convolution kernel size of and a non-linear activation function layer to obtain small sample gene recombination features The three features fused are the basic feature , the large-scale feature and the attention weighted feature The small sample gene recombination features after fusion integrate multi-scale basic features and attention focused key features, providing “focused and rich” initial features for subsequent modules, which can alleviate the feature sparsity problem of small sample data.

[0120] The calculation formula involved in the process is:

[0121] ,

[0122] wherein, denotes a feature splicing operation.

[0123] In the specific implementation, a gene multi-granularity mining module is constructed to solve the problem that multi-scale and multi-level characteristics of epigenetic gene features are difficult to be captured simultaneously. Multi-granularity features are mined through a multi-branch network, and key information is integrated through an attention mechanism. As shown in FIG. 8, S2 is specifically as follows: Figure 3

[0124] (1) Feature preprocessing operation:

[0125] An epigenetic gene multi-granularity mining module is established. The small sample gene recombination features are input into the epigenetic gene multi-granularity mining module. The module is subjected to a convolution layer with a convolution kernel size of , a batch normalization layer and a non-linear activation function to output optimized intermediate features , which are used to provide more stable and efficient inputs for the multi-branch structure.

[0126] The calculation formula involved in the process is:​

[0127] ,

[0128] in, Indicates the batch normalization layer. Indicates the kernel size as Convolutional layers;

[0129] (2) Two-branch multi-scale feature capture operation:

[0130] Features The input is fed into two parallel multi-scale salt tolerance gene feature capture networks;

[0131] First, features First, through the first branch of the network, the features After convolution kernel size is Depth separable convolutional layers Activation functions and fully connected layers output high-dimensional features. ;

[0132] The calculation formulas involved in this process are as follows:

[0133] ,

[0134] in, Indicates a fully connected layer. Indicates the kernel size as The depth of the separable convolutional layer;

[0135] At the same time, features The input is fed into the second branch of the network, which consists of a multi-level gene capture module and a multi-scale gene mining module;

[0136] 1) In the multi-level gene capture module, features are located through multi-level epigenetic feature localization operations. After average pooling layers, the kernel size is... Convolutional layers and Activation function, output features ;feature After convolution kernel size is The depth of the separable convolutional layer and the kernel size are Depth-separable convolutional layers and Activation function, output features ;feature After convolution kernel size is Convolutional layers and Activation function, output features This can solve the problem that a single layer of features cannot cover the "local-global" appearance association;

[0137] The calculation formulas involved in the above process are as follows:

[0138] ,

[0139] ,

[0140] ,

[0141] in, Indicates the average pooling layer. Indicates the kernel size as The depth of the separable convolutional layer, kernel size is The depth of the separable convolutional layer, express Activation function;

[0142] 2) In the multi-scale gene mining module, through multi-granularity feature mining operations, features... After convolution kernel size is Depth-separable convolutional layers output features ;

[0143] Features After convolution kernel size is Depth separable convolutional layers and Activation function, output features ;

[0144] Features After convolution kernel size is Depth separable convolutional layers and Activation function, output features ;

[0145] Features After convolution kernel size is Depth separable convolutional layers and Activation function, output features ;

[0146] Features After convolution kernel size is Depth separable convolutional layers and Activation function, output features ;

[0147] Features , , as well as Feature concatenation is performed, followed by a convolution kernel with a size of [missing value]. convolutional layer with a kernel size of ;

[0148] By capturing different granularities of apparent features from fine to coarse through different convolution kernels, the fusion features can solve the problem of missing important apparent features by single scale features;

[0149] The calculation formula involved in the above process is:

[0150] ,

[0151] ,

[0152] ,

[0153] ,

[0154] ,

[0155] ,

[0156] Among them, represents the feature splicing operation;

[0157] (3) Feature integration operation:

[0158] The feature is fused with the feature using element-wise multiplication operation, and the output feature is obtained through a convolutional layer with a kernel size of ;

[0159] The feature is spliced with the feature , and the output apparent gene mining feature is obtained through a convolutional layer with a kernel size of and an activation function . The feature integrates local, multi-level and multi-granularity apparent gene features, and strengthens the key area through attention mechanism, solving the problem of complex and weak correlation of apparent gene features, and providing high-quality apparent features for subsequent heterogeneous integration;

[0160] The calculation formula involved in the above process is:

[0161] ,

[0162] ,

[0163] Among them, represents element-wise multiplication operation, representing feature concatenation operation, representing activation function.

[0164] In the detailed description, in view of the problem that there is heterogeneity for salt-tolerant gene features from different sources and global correlation is difficult to capture, a heterogeneous salt-tolerant gene feature integration module is constructed, and heterogeneous feature fusion is realized through long-distance dependence capture and context integration, as shown in Figure 4 S3 is specifically as follows:

[0165] (1) Establishing a heterogeneous salt-tolerant gene feature integration module, inputting the apparent gene mining features to the heterogeneous salt-tolerant gene feature integration module, which includes a global salt-tolerant feature mining module and a salt-tolerant feature correlation construction module;

[0166] 1) Global salt-tolerant feature mining module:

[0167] feature After the global salt-tolerant feature mining module of the heterogeneous salt-tolerant gene feature integration module, the module captures the long-distance dependence features of the heterogeneous data through the long-distance spatial feature dependence capture operation, and the feature After the module, the convolution kernel size of the depth separable convolution layer is , the convolution kernel size of the convolution layer is , the batch normalization layer is , and the activation function is , and the output feature is ;

[0168] The feature is input into the depth separable convolution layer with a convolution kernel size of , and the convolution layer with a convolution kernel size of , and the obtained feature is added to the feature element by element, and the output feature is , and the feature captures the long-distance spatial dependence of the heterogeneous feature, which can solve the problem that the local feature cannot reflect the overall salt-tolerant rule;

[0169] The calculation formula involved in the above process is:

[0170] ,

[0171] ,

[0172] wherein, represents an element-by-element addition operation, represents a batch normalization layer.

[0173] 2) Salt tolerance feature correlation construction module:

[0174] The output features of the global salt tolerance feature mining module are input into the salt tolerance feature correlation construction module, which integrates the context information of heterogeneous features through a heterogeneous feature information context integration operation;

[0175] features The convolution kernel size of this module is a depth separable convolution layer, a batch normalization layer, and an activation function, resulting in features and features are concatenated, outputting feature , which is input into a global average pooling layer, a fully connected layer, a non-linear activation function, a fully connected layer, and an activation function, outputting feature ;

[0176] features and features are added element by element, outputting feature ;

[0177] features are input into a fully connected layer, resulting in features and features are concatenated, outputting feature , which highlights important feature dimensions and enriches feature diversity by integrating the context information of heterogeneous features;

[0178]

[0179] ,

[0180] ,

[0181] ,

[0182] ,

[0183] wherein represents a batch normalization layer, a global average pooling layer, an element-wise addition operation;

[0184] (2) Feature fusion operation:

[0185] features are input into a convolution kernel with a size of a convolutional layer, a convolution kernel size of a depth separable convolutional layer, and an activation function, outputting a feature ;

[0186] The feature is added to the feature using an element-wise addition operation, and an isomerism feature is outputted, which integrates long-distance spatial dependence and context association of isomerism features, realizes full-dimensional feature expression of “local-global-isomerism association”, and provides globally consistent isomerism features for final fusion.

[0187] The calculation formula involved in the above is as follows:

[0188] ,

[0189] ,

[0190] Among them, represents an element-wise addition operation.

[0191] In the embodiment, to solve the problem that features from different sources cannot be fully utilized, a salt-tolerant feature point fusion module is constructed, and small key features are captured through multi-branch fusion to realize deep integration of multi-element features, as shown in FIG. 4. Figure 5 S4 is specifically as follows:

[0192] (1) Preliminary fusion:

[0193] A salt-tolerant feature point fusion module is established, and small sample gene recombination features , epigenetic gene mining features and isomerism features are input into the salt-tolerant feature point fusion module, features , features and features are subjected to feature splicing operation, a convolutional layer with a convolution kernel size of and a non-linear activation function , and features are outputted.

[0194] The calculation formula involved in the process is as follows:

[0195] ;

[0196] (2) Double-branch small feature step operation:

[0197] Features are subjected to parallel salt gene fusion feature capturing networks to capture small salt gene features in the features.

[0198] 1) Feature The first branch network of the input salt-tolerant gene fusion feature capture network compresses the features Get global features and capture local details; Specifically, the features are passed through a global average pooling layer, a convolutional layer with a kernel size of , a batch normalization layer, and a non-linear activation function, and the output features ; The features are passed through a depth separable convolutional layer with a kernel size of , a convolutional layer with a kernel size of , and a fully connected layer, and the output global micro features ;

[0199] The calculation formula involved in this process is:

[0200] ,

[0201] ,

[0202] wherein represents a global average pooling layer, represents a batch normalization layer, represents a fully connected layer;

[0203] 2) The features are input into the second branch network of the salt-tolerant gene fusion feature capture network, which retains significant features and captures mesoscale details; Specifically, the features are passed through a max pooling layer, a convolutional layer with a kernel size of , and an activation function, and the output features ; The features are passed through a depth separable convolutional layer with a kernel size of , an activation function, a depth separable convolutional layer with a kernel size of , and a fully connected layer, and the output local micro features ;

[0204] The features and the features are added element by element to obtain features that are passed through a convolutional layer with a kernel size of , an activation function, a fully connected layer, and a activation function, and the output salt-tolerant multi-element features , which integrates the micro key patterns in the multi-featured characteristics (such as the easily ignored salt-tolerant core gene fragments), solves the problem of missing fine-grained information in the previous characteristics, and provides the most comprehensive fusion features for prediction;

[0205] The calculation formula involved in the above process is:

[0206] ,

[0207] ,

[0208] ,

[0209] Among them, represents the max pooling layer, represents the element-wise addition operation, represents the activation function.

[0210] In the specific implementation, as shown in S5, the following is specific: Figure 6

[0211] The tomato salt-tolerance gene prediction module is established, the mapping from the fusion features to the salt-tolerance prediction is realized, the non-linear conversion problem from the features to the prediction results is solved, and specifically the salt-tolerance multi-featured characteristics are input into the tomato salt-tolerance gene prediction module, the salt-tolerance multi-featured characteristics are subjected to the convolution kernel with the convolution kernel size of , the feature flattening operation, the full connection layer and the activation function, and the predicted related tomato salt-tolerance gene probability is output; ;

[0212] The predicted probability is subjected to the max index value operation, and is taken as the final predicted salt-tolerance gene of the proposed method.

[0213] The calculation formula involved in the above process is:

[0214] ,

[0215] Among them, represents the feature flattening operation.

[0216] In the specific implementation, S6 is specific as follows:

[0217] ​​In view of the problem that the parameter optimization direction is not clear in model training, the tomato salt tolerance prediction model is learned through a loss function, the tomato salt tolerance prediction model is trained using a binary cross-entropy loss function according to the prediction result and the real label of the tomato salt tolerance genome to be detected, the parameters of the model in the training process are optimized and adjusted according to the loss function result, and the trained tomato salt tolerance prediction model is obtained.

[0218] The real label is obtained by labeling the tomato salt tolerance genome before the tomato salt tolerance genome to be detected is input into the tomato salt tolerance prediction model, and the SISALT1 gene is marked in the tomato salt tolerance genome, which is used to optimize and train the model with the prediction result of the model.

[0219] Embodiment 2

[0220] A tomato salt tolerance prediction system based on deep learning includes modules for executing processing instructions for each step in a tomato salt tolerance prediction method based on deep learning, specifically including the following modules:

[0221] A genome data input module receives tomato salt tolerance genome data to be detected and pre-processes the data.

[0222] A salt-tolerant small sample gene recombination module captures multi-level features and attention-driven feature recombination of input genome data, and outputs small sample gene recombination features.

[0223] An epigenetic multi-granularity mining module performs multi-level epigenetic feature positioning and multi-granularity feature mining on small sample gene recombination features, and outputs epigenetic mining features.

[0224] A heterogeneous salt-tolerant gene feature integration module integrates heterogeneous feature information context and long-distance spatial feature dependency capture on epigenetic mining features, and outputs heterogeneous features.

[0225] A salt-tolerant feature point fusion module fuses small sample gene recombination features, epigenetic mining features, and heterogeneous features, and outputs salt-tolerant multi-element features.

[0226] A salt tolerance prediction and model optimization module predicts tomato salt tolerance gene probability based on salt-tolerant multi-element fusion features, and optimizes model parameters through a loss function, and outputs a final prediction result.

[0227] Embodiment 3

[0228] In order to verify the effectiveness of the tomato salt tolerance prediction method and system based on deep learning in the field of tomato salt tolerance prediction technology, the tomato salt tolerance prediction method based on deep learning is verified under the same experimental conditions as the existing related prediction method, and the obtained comparative prediction results are verified and analyzed to illustrate the effectiveness of the method in the field of tomato salt tolerance prediction technology.

[0229] In the comparative test of the tomato salt tolerance prediction method based on deep learning and the existing prediction method, three prediction methods are selected as benchmark models for comparison, and the obtained experimental and prediction results are compared and analyzed to verify the effectiveness of the method in the field of tomato salt tolerance prediction technology.

[0230] The three benchmark models used in the comparative experiment are:

[0231] 1.ST-Net: Based on DenseNet121 transfer learning framework, the advantages are simple model structure and high computational efficiency; the disadvantages are ignoring local neighborhood and global organization information, and the prediction accuracy is limited.

[0232] 2.EGN: Enhanced prediction by dynamic retrieval of similar examples, the advantages are that it can use reference data sets to improve the representation of target points; the disadvantages are that it depends on the quality of examples, and it does not integrate multi-scale biological context, and the generalization ability of the model is insufficient.

[0233] 3.HisToGene: Based on ViT for global relationship modeling, the advantages are that it can capture long-range spatial dependencies of features; the disadvantages are high computational complexity and ignoring local features.

[0234] In the comparative experiment of the tomato salt tolerance prediction method based on deep learning proposed in the application, three experimental indicators are selected to measure the effectiveness of the proposed method: Mean Squared Error (MSE) is used to measure the mean square deviation between the predicted value and the true value, which is sensitive to outliers; Mean Absolute Error (MAE) reflects the average absolute deviation of the predicted value and the true value, which is not sensitive to outliers; and Pearson Correlation Coefficient Mean (PCC(M)) is used to evaluate the prediction consistency of the model for all gene expression space patterns, reflecting the ability of the model to capture the overall gene expression distribution rule.

[0235] In the experiment of the proposed tomato salt tolerance prediction method based on deep learning, 368 tomato germplasm resources were evaluated and identified for salt tolerance, and new salt-tolerant germplasm was obtained. The F6 recombinant inbred lines were constructed using salt-tolerant varieties (LA0722, LA1374), and the key genes for salt tolerance in tomato were fine-mapped. The data set constructed on this basis was used for model training and performance verification. The data set included whole genome transcriptome sequencing data of tomato root and leaf tissues under different salt stress conditions, covering multiple time point samples under multiple salt concentration treatments, and recording the gene expression profile changes closely related to plant stress response, such as the expression levels of ion channel proteins, antioxidant enzymes and osmotic adjustment substances related genes. In the experiment, the data set was classified according to the sample processing conditions and tissue types, and further divided into training set, validation set and test set in the ratio of 7:1:2 to ensure the effectiveness and generalization ability of the model training, and to test the prediction accuracy and stability of the proposed method under different salt stress levels.

[0236] In the comparative experiment of the proposed tomato salt tolerance prediction method based on deep learning, the experimental results are shown in Table 1. The experimental results show that the proposed tomato salt tolerance prediction method achieves the best results in the comparative experiment with three benchmark models, with MSE (Mean Square Error) reduced to 0.241, 2.5% lower than the optimal baseline HisToGene (0.266), indicating that the deviation of the prediction value and the true value of the proposed method is significantly reduced. MAE (Mean Absolute Error) is 0.395, which is 3.2% lower than HisToGene model 0.427, reflecting the improvement of the stability of the prediction results of the proposed method. PCC(M) (gene space pattern correlation mean) is improved to 0.118, which is 1.4% higher than HisToGene model 0.104, proving that the proposed method has stronger ability to capture the distribution of gene expression.

[0237] The above experimental results prove the effectiveness of the proposed method in the field of tomato salt tolerance prediction technology using deep learning, which can overcome the defects of existing benchmark methods. The method of the present application can focus on local and global organizational information, thereby improving the prediction accuracy and generalization ability of the model.

[0238] Table 1 Comparison results of the proposed method and benchmark models in the comparative experiment

[0239]

[0240] Although the specific embodiments of the application have been described above with reference to the accompanying drawings, the description is not a limitation on the scope of protection of the application. Various modifications or variations made by those skilled in the art without creative labor based on the technical solutions of the present application are still within the scope of protection of the present application.

Claims

1. A deep learning-based tomato salt tolerance prediction method, a tomato salt tolerance prediction model is constructed, and a tomato salt tolerance genome to be detected is input into the model for salt tolerance prediction, characterized in that the model The operation in the epigenetic multi-granularity mining module is specifically as follows: S1, construct a salt-tolerant small sample gene recombination module, input the tomato salt-tolerant genome to be detected , after multi-level feature capture and attention-driven feature recombination operation, output small sample gene recombination feature ; S2, construct an epigenetic multi-granularity mining module, input small sample gene recombination features , after multi-level epigenetic feature positioning and multi-granularity feature mining operation, output epigenetic mining features ; The operation in the epigenetic multi-granularity mining module is specifically as follows: (1) Feature preprocessing operation: Establish an epigenetic multi-granularity mining module, recombine the characteristics of small sample genes to the epigenetic multi-granularity mining module, through the convolution layer, batch normalization layer and non-linear activation function in the module, output the optimized intermediate features ; (2) Double-branch multi-scale feature capturing operation: characteristics input to two parallel multi-scale salt-tolerant gene feature capturing networks; Features Through the first branch of the network, features Through deep separable convolutional layers, Activation functions and fully connected layers, output high-dimensional features ; Meanwhile, the features are input into a second branch of the network, which is composed of a multi-level gene capturing module and a multi-scale gene mining module; 1) In the multi-level gene capture module, through the multi-level feature positioning operation, the features After the average pooling layer, the convolution layer and the activation function, the feature is output; the feature passes through two depth separable convolution layers with different convolution kernel sizes and the activation function, and the feature is output; the feature passes through the convolution layer and the activation function, and the feature is output; 2) In the multi-scale gene mining module, multi-granularity feature mining operation is performed; Specifically, the features The output features pass through a depth separable convolutional layer ; The granularity feature mining operation is performed on the feature Multi-level feature extraction is performed, each layer structure being a deep separable convolution layer and an activation function, the convolution kernel size of each deep separable convolution layer being different; after feature extraction is performed on the feature , the feature is output , and based on the feature , feature extraction is performed, and the feature is output , and based on the feature , feature extraction is performed, and the feature is output , and based on the feature , feature extraction is performed, and the feature is output . Then, the extracted features of each layer are spliced and output features after a convolution layer , , , are output ; (3) Feature integration operation: features with features feature fusion is performed using an element-wise multiplication operation, and is output through a convolution layer ; The features are combined The features are combined The features are combined, and the convolution layer and the activation function are used to output the apparent gene mining features The features are combined, and the convolution layer and the activation function are used to output the apparent gene mining features The features are combined, and the convolution layer and the activation function are used to output the apparent gene mining features S3, construct a heterogeneous salt-tolerant gene feature integration module, input the epigenetic gene mining features , after the module's heterogeneous feature information context integration, long-distance spatial feature dependence capture operation, output heterogeneous features ; S4, construct a salt-tolerant feature point fusion module, input small sample gene recombination features , epigenetic mining features , and heterogeneous features , after salt-tolerant gene multi-feature fusion operation, output salt-tolerant multi-feature ; S5, constructing a tomato salt-tolerance gene prediction module, inputting salt-tolerance multi-element features , predicting a tomato salt-tolerance gene probability ; S6, The parameters in the tomato salt tolerance prediction model are optimized by a loss function, and a trained tomato salt tolerance prediction model is obtained.

2. The tomato salt tolerance prediction method based on deep learning according to claim 1, characterized in that, The operation in the salt tolerance feature point fusion module is specifically as follows: (1) Preliminary fusion: The salt-tolerant feature point fusion module is established, and the small sample gene recombination features , epigenetic mining features and heterogeneous features are input into the salt-tolerant feature point fusion module, and the features are preliminarily fused through feature splicing operation, convolution layer and non-linear activation function of the module, and the preliminarily fused features are output ; (2) Double-branch micro-feature step operation: Features The salt-tolerant gene fusion features are captured by the parallel salt-tolerant gene fusion feature capture networks, and the salt-tolerant gene features are captured by the parallel salt-tolerant gene feature capture networks, wherein the salt-tolerant gene features are global salt-tolerant gene features and local salt-tolerant gene features, respectively. 1) Feature The first branch network of the input salt-tolerant gene fusion feature capture network compresses the features through a global average pooling layer, and captures local local details through a depth separable convolution to obtain global features; Specifically, the features pass through a global average pooling layer, a convolutional layer, a batch normalization layer, and a non-linear activation function, output features ; the features pass through a depthwise separable convolutional layer, a convolutional layer, and a fully connected layer, output global micro features ; 2) the feature The second branch network of the input salt-tolerant gene fusion feature capture network retains significant features through a max-pooling layer and captures mesoscale details through a depthwise separable convolution; Specifically, the features through a max-pooling layer, a convolutional layer, and an activation function, output features ; characteristics through two deep separable convolutional layers, activation functions and a fully connected layer, output local micro features ; characteristics with characteristics after using an element-wise addition operation, followed by a convolutional layer, an activation function, a fully connected layer, and an activation function, outputting salt-tolerant multi-element characteristics . 3.The tomato salt tolerance prediction method based on deep learning according to claim 1, characterized in that, The operation in the salt tolerance small sample gene recombination module is specifically as follows: (1) Multi-level feature capturing operation: A salt-tolerant small sample gene recombination module is established, and a tomato salt-tolerant genome to be detected is input into the salt-tolerant small sample gene recombination module A multi-level feature capturing operation is performed on the genome in the module, which passes through a deep separable convolution layer, a batch normalization layer and a non-linear activation function layer in the module to obtain features ​ The three-layer feature capturing layer is constructed, and each layer structure is composed of a convolution layer, a depth separable convolution layer and An activation function is formed, the convolution kernel size of the depth separable convolution layer in each layer is different, and different convolution kernel sizes are set to capture features of different scales. features input to a three-layer feature capture layer, output features from a first layer of the feature capture layer , again inputting the features to a second layer of the feature capture layer, output features , again inputting the features to a third layer of the feature capture layer, output features ; (2) Attention-driven feature recombination operation: features input to the feature reconstruction network undergoes an attention-driven feature reorganization operation, specifically using feature disassembling operation disassembles the features into feature blocks of a certain size, focusing on local features, and features are compressed by a global average pooling layer to obtain features ; Features Feature transformation is performed after a fully connected layer, and then... Activation function generates attention weight features ; Attention weight features using feature folding operation, which reconstructs its features to generate reconstructed features ; (3) Feature fusion operation: features , features , and reconstructed features After feature concatenation operation, convolution layer and non-linear activation function layer, small sample gene recombination features are obtained. 4.The tomato salt tolerance prediction method based on deep learning according to claim 1, characterized in that, The operation in the heterogeneous salt tolerance gene feature integration module is specifically as follows: (1) Establishing a heterogeneous salt-tolerant gene feature integration module, which inputs the apparent gene mining features to the heterogeneous salt-tolerant gene feature integration module, which includes a global salt-tolerance feature mining module and a salt-tolerance feature correlation construction module; 1) Global salt tolerance feature mining module: Features The global salt-tolerant feature mining module of the isomorphic salt-tolerant gene feature integration module captures the long-distance dependent features of the isomorphic data through a long-distance spatial feature dependent capture operation, and the features After the deep separable convolution layer, the convolution layer, The batch normalization layer and The activation function of the module, the output feature ; Features The input is fed into depthwise separable convolutional layers and convolutional layers to obtain features and features Element-wise addition operation, output characteristics ; 2) Salt tolerance feature correlation construction module: Output features of the global salt tolerance feature mining module are input to a salt tolerance feature correlation construction module that integrates the context information of the heterogeneous features through a heterogeneous feature information context integration operation. Features a deep separable convolutional layer, a batch normalization layer, and an activation function, the features and the features are spliced, and the features are outputted. a global average pooling layer, a fully connected layer, a non-linear activation function, a fully connected layer, and an activation function, and the features are outputted. Features With features Output features using an element-wise addition operation Features After the full connection layer, the obtained features and the features The feature splicing operation is performed, and the features ; (2) Feature fusion operation: Will through the convolutional layer, the depth separable convolutional layer and activation function, output feature ; characteristics with characteristics outputting heterogeneous features using an element-wise addition operation .

5. The tomato salt tolerance prediction method based on deep learning according to claim 1, characterized in that, The operation in the tomato salt tolerance gene prediction module is specifically as follows: A tomato salt tolerance gene prediction module is established, and specifically, salt tolerance multi-features are input into the tomato salt tolerance gene prediction module After convolution kernel, feature flattening operation, full connection layer and activation function, the predicted probability of related tomato having salt tolerance gene is output According to the probability size, it is judged whether the detected tomato has salt tolerance gene, and the final prediction result is output.

6. The tomato salt tolerance prediction method based on deep learning according to claim 1, wherein S6 Specifically as follows: According to the prediction result and the true label of the tomato salt tolerance genome to be detected, a binary cross-entropy loss function is used to train the tomato salt tolerance prediction model, and the parameters of the model in the training process are optimized and adjusted according to the loss function result, to obtain a trained tomato salt tolerance prediction model. 7.A deep learning-based tomato salt tolerance prediction system, characterized by, The method comprises the following steps: The genome data input module receives the tomato salt tolerance genome data to be detected and pre-processes the data; The salt tolerance small sample gene recombination module captures multi-level features and attention-driven feature recombination from the input genome data, and outputs small sample gene recombination features; The epigenetic multi-granularity mining module performs multi-level epigenetic feature positioning and multi-granularity feature mining on the small sample gene recombination features, and outputs epigenetic mining features; The heterogeneous salt tolerance gene feature integration module integrates heterogeneous feature information context and long-distance spatial feature dependence capture on the epigenetic mining features, and outputs heterogeneous features; The salt tolerance feature point fusion module fuses small sample gene recombination features, epigenetic mining features, and heterogeneous features, and outputs salt tolerance multi-element features; The salt tolerance prediction and model optimization module predicts tomato salt tolerance gene probability based on salt tolerance multi-element fusion features, optimizes model parameters through a loss function, and outputs the final prediction result.

Images