A multi-omics plant phenotype prediction method fusing small rnas

Abstract

The application discloses a kind of small RNA fusion multiomics plant phenotype prediction method, it is related to bioinformatics and agricultural biology field, including the following steps: obtaining and pre-processing genotype, transcript expression and small RNA expression data, obtain standardized feature matrix;The standardized feature matrix is screened based on the feature of pearson correlation, and the corresponding genotype, transcript expression, small RNA expression feature subset is obtained;Based on the pre-trained MirGP prediction model, by multi-branch feature extraction, multi-level efficient channel attention fusion and regression prediction, output plant phenotype prediction result.The method introduces multiple small RNA regulation characteristics and constructs hierarchical deep learning fusion architecture, expands the multiomics feature dimension, reduces data redundancy, realizes the effective integration and deep mining of cross-modal features.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and agricultural biology, and more specifically to a multi-omics plant phenotypic prediction method incorporating small RNAs. Background Technology

[0002] Plant phenotypic prediction is a key step in crop genetics and breeding. It aims to accurately estimate the complex traits of plants using multidimensional molecular data such as genomics and transcriptomics, thereby accelerating the breeding process of superior varieties. For example, maize, as an important global food crop and genetic model species, has a highly complex genome, and the functions of many genes have not yet been elucidated, which places higher demands on the accuracy and generalization ability of phenotypic prediction methods.

[0003] Existing plant phenotypic prediction methods mainly fall into three categories: linear mixture models, traditional machine learning methods, and deep learning methods. Linear mixture models, represented by rrBLUP, GBLUP, and Bayesian series, are robust in predicting traits dominated by additive genetic effects. However, they are essentially linear models and cannot capture nonlinear interactions and complex regulatory relationships between genes. Traditional machine learning methods, such as random forests and support vector machines, can partially characterize nonlinear relationships, but they are prone to overfitting when dealing with high-dimensional, small-sample data. Furthermore, feature engineering relies heavily on human experience and is difficult to systematically integrate heterogeneous multi-omics data.

[0004] Deep learning-based prediction models have received widespread attention in recent years. Convolutional neural networks, recurrent neural networks, and their variants have been used to mine local haplotype patterns or sequence features, achieving prediction accuracy superior to traditional methods. However, existing deep learning models still have significant limitations in multi-omics fusion: most studies only use SNP markers or transcript expression levels as inputs, neglecting the important role of small RNA expression and its mediated posttranscriptional regulatory network in the formation of complex traits. Small RNAs such as miRNA, sRNA cluster, isomiR, and tRF are key regulators of gene expression, and their expression levels are closely related to various agronomic traits, but they have not yet been systematically integrated into the phenotypic prediction framework.

[0005] Furthermore, existing multi-omics fusion methods mostly employ simple feature splicing or shallow fusion strategies, lacking refined interactive modeling mechanisms tailored to the characteristics of different modalities. Genomic data is characterized by high dimensionality and sparsity, transcript expression data reflects gene expression activity, and small RNA expression data carries post-transcriptional regulatory signals. These three data have different statistical distributions and biological meanings, and simple splicing cannot fully explore their complementary information. At the same time, most models lack interpretability analysis of prediction results, making it difficult to analyze the contribution mechanism of key features to phenotype from a biological perspective, thus limiting their guiding value in breeding practice.

[0006] Therefore, designing a multi-omics plant phenotypic prediction method that integrates small RNAs to systematically integrate genotype, transcript expression, and expression characteristics of multiple small RNAs, and to achieve refined fusion of cross-modal information and interpretable analysis of multi-omics phenotypic prediction, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a multi-omics plant phenotypic prediction method that integrates small RNAs. It aims to achieve accurate prediction of complex plant traits by systematically integrating the expression features of various small RNAs such as miRNA, sRNA cluster, isomiR, and tRF with genotype features and transcript expression features, and designing a deep learning model composed of a multi-branch coding network, a two-level efficient channel attention mechanism, and a hybrid depth separable convolutional module.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multi-omics plant phenotypic prediction method incorporating small RNAs includes the following steps:

[0010] S1. Obtain and preprocess genotype, transcript expression, and small RNA expression data to obtain a standardized feature matrix;

[0011] S2. Perform feature screening based on Pearson correlation on the standardized feature matrix to obtain the corresponding genotype, transcript expression, and small RNA expression feature subsets.

[0012] S3. Based on the pre-trained MirGP prediction model, the plant phenotypic prediction results are output through multi-branch feature extraction, multi-level efficient channel attention fusion and regression prediction.

[0013] Preferably, in step S1, the preprocessing of genotype data includes:

[0014] Single nucleotide polymorphism (SNP) sites are encoded as numerical values ​​according to allele counts, where homozygotes are encoded as 0, heterozygotes as 1, and the other homozygotes as 2. Deletion values ​​are denoted as NaN.

[0015] High deletion sites are filtered out using a preset deletion rate threshold, and rare variants are removed using a preset minor allele frequency threshold.

[0016] Missing loci were filled using the mode of the column, and the correlation coefficient between loci was calculated within a sliding window. ,when When the value exceeds 0.2, redundant sites are randomly removed to obtain the genotype feature matrix.

[0017] Preferably, in S1, the preprocessing of transcript expression and small RNA expression data includes:

[0018] The transcript expression matrix and the small RNA expression matrix are read in and aligned; the small RNA expression matrix includes the miRNA expression matrix, sRNA cluster expression matrix, isomiR expression matrix, and tRF expression matrix.

[0019] All expression matrices were forcibly converted to numerical values. Unparseable entries were filled with 0.0 and missing values ​​were filled with column mode to obtain the corresponding transcript expression feature matrix and small RNA expression feature matrix.

[0020] Preferably, in step S2, the feature screening based on Pearson correlation includes:

[0021] Calculate the Pearson correlation coefficient r between each molecular feature vector x and the target trait vector y;

[0022] Calculate the t-statistic based on the degrees of freedom df=n The t-distribution statistic for 2 is used to calculate the two-tailed p-value, where n is the sample size.

[0023] Filtering p-values ​​less than a preset threshold Features are entered into subsets. Features within subsets are then sorted in ascending order of p-value. If p-values ​​are the same, they are sorted by the absolute value of their correlation coefficients. Determine priority by descending order.

[0024] Preferably, S3 includes:

[0025] Genotype, transcript expression, and small RNA expression feature subsets are respectively input into the linear embedding layer for linear transformation, and then concatenated along the feature dimension before being input into the first high-efficiency channel attention module for channel weighting to obtain the initial fusion features;

[0026] Genotype, transcript expression, and small RNA expression feature subsets are respectively input into the corresponding feature extraction networks for feature extraction. After being concatenated with the initial fusion feature along the feature dimension, they are input into the second high-efficiency channel attention module for channel weighting to obtain multi-dimensional fusion features. The feature extraction network is composed of mixed depthwise separable convolutional modules.

[0027] The multi-dimensional fused features are extracted into deep features through the top-level hybrid depth separable convolution module, and the plant phenotype prediction results are output through a multilayer perceptron.

[0028] Preferably, in the feature extraction network, the feature extraction network for processing genotype feature subsets consists of a hybrid depth separable convolutional module, the feature extraction network for processing small RNA expression feature subsets consists of two tandem hybrid depth separable convolutional modules, and the feature extraction network for processing transcript expression feature subsets consists of two tandem hybrid depth separable convolutional modules.

[0029] Preferably, the mixed depthwise separable convolutional module adopts a MixConv structure, which divides the input feature map into g sub-maps according to channels. The corresponding convolution kernels are divided into Where h and w represent the height and width of the feature map, respectively. , Let represent the number of channels and the kernel size of the t-th subgraph, respectively, and m be the output channel number multiplier;

[0030] Each sub-image is subjected to depthwise convolution and then stitched together along the channel dimension. The images are then passed through a batch normalization layer, a LeakyReLU activation function, and a max pooling layer in sequence before being output.

[0031] Preferably, the data processing procedures of both the first high-efficiency channel attention module and the second high-efficiency channel attention module include:

[0032] Adaptive average pooling is performed on the input feature map X to obtain the channel descriptor Y;

[0033] The one-dimensional convolution kernel size k is adaptively determined based on the number of channels C, and a one-dimensional convolution is performed on the channel descriptor Y to capture local cross-channel interactions, thus obtaining the convolution output. ;

[0034] The convolution output Channel weights ω are generated by activating the Sigmoid function;

[0035] The input feature map X is multiplied by the channel weight ω channel by channel to obtain the weighted output feature map. .

[0036] Preferably, the adaptive determination of the one-dimensional convolution kernel size k based on the number of channels C includes:

[0037]

[0038] Where γ and b are preset hyperparameters, This indicates taking the nearest odd number.

[0039] Preferably, the multilayer perceptron includes multiple fully connected layers, each of which performs a linear transformation, and the output dimension of the last fully connected layer is consistent with the target phenotype dimension to obtain a predicted value.

[0040] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0041] 1. This method integrates the expression features of multiple small RNAs, including miRNA, sRNA cluster, isomiR, and tRF, into a multi-omics analysis framework. Together with genotype features and transcript expression features, it forms the input of the prediction model. Compared with traditional methods that rely solely on SNPs or transcript expression levels, this design can capture biological signals at the post-transcriptional regulatory level, providing a more comprehensive molecular-level information basis for phenotypic prediction.

[0042] 2. In the genotype data preprocessing stage, it adopts missing rate filtering and minor allele frequency screening. In the omics expression data level, it adopts significance screening based on Pearson correlation to remove noisy features and redundant sites layer by layer, and obtains feature subsets that are more related to the target phenotype. This lays a high-quality data foundation for subsequent model training and reduces the risk of overfitting and computational burden caused by high-dimensional features.

[0043] 3. A deep learning architecture for multi-omics fusion was constructed, realizing the effective integration and deep mining of cross-modal features. The MirGP model adopts a hierarchical design with multi-branch linear embedding, multi-level ECA attention weighting, hybrid deep separable convolution, and multi-layer perceptron regression. The multi-branch structure retains the independent features of each modality, the multi-level attention modules realize inter-modal weighting and cross-scale channel filtering respectively, and the hybrid convolution module captures multi-scale local patterns. This architecture can achieve efficient fusion of cross-modal information while retaining omics specificity, providing a systematic deep modeling scheme for predicting complex traits in plants. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 A flowchart of a multi-omics plant phenotypic prediction method incorporating small RNAs is provided in this embodiment of the invention.

[0046] Figure 2 This is a schematic diagram of the MirGP prediction model and depthwise separable convolutional module structure provided in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the high-efficiency channel attention module structure provided in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram showing the comparison of prediction performance of the DNNGP model before and after feature selection in an embodiment of the present invention.

[0049] Figure 5 A schematic diagram showing the comparison of phenotypic prediction performance between the MirGP model and other classic models provided in this embodiment of the invention;

[0050] Figure 6 A schematic diagram illustrating the proportion of sample contribution values ​​for genotype (G) and miRNA (SE) features provided in embodiments of the present invention;

[0051] Figure 7 This is a schematic diagram showing the distribution of contribution values ​​of genotype (G) and miRNA (SE) for each sample provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] like Figure 1 As shown in the figure, this embodiment provides a multi-omics plant phenotypic prediction method fused with small RNAs, including the following steps:

[0054] S1. Obtain and preprocess genotype, transcript expression, and small RNA expression data to obtain a standardized feature matrix;

[0055] S2. Perform feature screening based on Pearson correlation on the standardized feature matrix to obtain the corresponding genotype, transcript expression, and small RNA expression feature subsets.

[0056] S3. Based on the pre-trained MirGP prediction model, the plant phenotypic prediction results are output through multi-branch feature extraction, multi-level efficient channel attention fusion and regression prediction.

[0057] By integrating the expression features of multiple small RNAs with genotype and transcript expression features, it overcomes the shortcomings of traditional methods in covering small RNA regulatory information. Through a hierarchical feature screening mechanism, it effectively reduces data redundancy and noise. By retaining modality-specific information through a multi-branch coding network and combining a two-level efficient channel attention mechanism with a hybrid depth separable convolution module, it achieves refined fusion of cross-modal features and extraction of multi-scale abstract features, thereby improving the predictive ability of complex plant traits.

[0058] The following provides a further detailed explanation of each step in the above method;

[0059] In this embodiment, S1, genotype, transcript expression, and small RNA expression data are acquired and preprocessed to obtain a standardized feature matrix;

[0060] Genotype data exists in the form of single nucleotide polymorphism (SNP) sites, usually recorded in character form, such as AA, AT, TT, or A / G. In this embodiment, SNP data is converted into numerical codes: at each site, the major and minor alleles are determined based on the allele count in the sample. The major allele homozygosity is encoded as 0, the heterozygosity as 1, and the minor allele homozygosity as 2. For sites that cannot be determined or missing data, they are uniformly marked as NaN. This encoding method preserves the dose-response information of genetic variation and facilitates subsequent numerical calculations.

[0061] Subsequently, a two-step site-level quality control screening was performed. The first step used a 5% missing rate threshold to filter high-missing sites, that is, to delete SNP sites missing in more than 5% of the samples. The second step used a minor allele frequency (MAF) threshold of 0.05 to remove rare variants with too low a variation frequency in the population, so as to avoid their interference with subsequent statistical analysis. For the retained sites, the column mode, that is, the coding value of the site with the highest frequency in all samples, was used for missing filling. This can maintain the data distribution characteristics while avoiding the introduction of systematic bias due to missing values.

[0062] Furthermore, to mitigate feature redundancy caused by linkage disequilibrium (LD), this embodiment performs correlation coefficient-based pruning within a sliding window; specifically, the window size is set to 500 sites, the step size is 50 sites, and the correlation coefficient between adjacent sites is calculated within each window. ;when When the value exceeds 0.2, one of the loci is randomly removed, thereby effectively reducing the feature dimensionality while preserving genetic information;

[0063] This step retained 250 genotypes from 368 genotype samples, and retained 48,228 genotype features from 532,501 genotype features, including PUT-163a-71311771-3117, PUT-163a-148943685-475, and PZE-106111890. The entire preprocessing process used a fixed random seed to ensure the reproducibility of the results. After the above processing, a genotype feature matrix G was obtained, where the rows correspond to samples and the columns correspond to the selected SNP loci.

[0064] Further, preprocessed transcript expression and small RNA expression data included:

[0065] This embodiment obtains four sets of small RNA expression matrices: miRNA expression matrix SE, sRNAcluster expression matrix CE, isomiR expression matrix IE, and tRF expression matrix TRE. These data are stored in the form of samples as rows and molecular features as columns. To ensure the reliability and consistency of the data in subsequent analysis, all expression matrices are first precisely aligned with the target trait data in terms of sample dimensions. Only entries that are common to both matrices and target trait data and have no missing samples are retained, while entries with missing sample information in any dataset are deleted. After completing the sample alignment, numerical standardization is performed on all expression matrices: all entries in the matrix are forcibly converted to numerical format. For abnormal entries that cannot be parsed, such as null values ​​and non-numerical characters, they are uniformly filled with 0.0 to complete data cleaning. Finally, a multi-omics feature matrix without missing data is obtained, specifically including the transcript expression feature matrix TE and the small RNA feature matrices SE, CE, IE, and TRE.

[0066] To achieve sample alignment, the feature dimensions remained unchanged after the above processing. From the 338 samples of transcript expression and the 338 samples of four sets of small RNA expression, 250 samples with the same sample ID were selected and retained. Through the above preprocessing process, the original high-dimensional and heterogeneous data was converted into a standardized feature matrix with uniform format and controllable quality, providing a reliable data foundation for subsequent feature selection and model construction.

[0067] In embodiment S2, the standardized feature matrix is ​​subjected to feature screening based on Pearson correlation to obtain the corresponding genotype, small RNA, and transcript expression feature subsets; wherein, the feature screening based on Pearson correlation includes:

[0068] Calculate the Pearson correlation coefficient r between each molecular feature vector x and the target trait vector y;

[0069]

[0070] Where n is the number of samples. and , respectively, are the mean values ​​of feature x and trait y. The correlation coefficient r ranges from [-1, 1], and the larger its absolute value, the stronger the linear association between the feature and the trait.

[0071] To assess the statistical significance of the correlation, the t-statistic is further calculated based on the degrees of freedom df=n. The t-distribution statistic for 2 is used to calculate the two-tailed p-value, where n is the sample size.

[0072]

[0073] Filtering p-values ​​less than a preset threshold Features are selected into a subset. In this embodiment, the threshold for feature selection is set to p < 0.05, meaning only features with statistically significant correlation to the target trait are retained. If no feature in a certain omics reaches this threshold, to avoid complete information loss, all computable features of that omics are retained for subsequent steps. Among the features that pass the threshold, the priority is determined by ranking the p-value first, then the effect size: first, they are sorted in ascending order of p-value, with smaller p-values ​​indicating more reliable correlation; when p-values ​​are the same or close, they are sorted by the absolute value of the correlation coefficient. Sort in descending order. The larger the value, the stronger the effect size.

[0074] In this embodiment, the dimensions after processing each phenotype are different. Taking the plant height phenotype (PH) as an example, after the above steps, chr2 was screened and retained from a total of 48,228 genotype characteristics, including PUT-163a-71311771-3117, PUT-163a-18179900-1492532501, and PZE-106111890.A total of 5993 genotype characteristics, including S_6984714, PZE-110023356, and PZE-106111890, were examined; among transcript expression characteristics such as GRMZM2G100965_T01, GRMZM5G825324_T0, and GRMZM2G346834_T01, transcript expression characteristics such as GRMZM5G825324_T0, GRMZM2G346834_T01, and GRMZM2G078174_T01 were screened and retained; and 322 other genotype characteristics, including zma-miR159a-5p, zma-miR171l-3p, and zma-miR169f-5p, were also examined. Among the 37,430 miRNA expression features, 28 miRNA expression features, including zma-miR159a-5p, zma-miR169f-5p, and zma-miR166m-3p, were screened and retained; among the 37,430 sRNA cluster expression features, including Cluster_6, Cluster_738482, and Cluster_469602, 2,133 sRNA cluster expression features, including Cluster_738482, Cluster_469602, and Cluster_462822, were screened and retained; “TTCGGACCAGGCTTCAT” From 1869 isomiR expression features such as “ACCC”, “TGCTCGCTTCTCTTTCTGTC”, and “AAAGATTATTGAGCTGGTGGTGGA”, 223 isomiR expression features such as “TTCGGACCAGGCTTCATACCC”, “TGCTCGCTTCTCTTTCTGTC”, and “AAGCTCAGGAGGGATAGCGCT” were selected and retained; tRF-2_tRNA-Pro-TGG-9-1_25, tRF-5_tRNA-Val-AAC-11-1_24, and tRF-2_tRNA- From tRF expression features such as Arg-ACG-5-1_18, tRF expression features including tRF-2_tRNA-Pro-TGG-9-1_25, tRF-5_tRNA-Val-AAC-11-1_24, and tRF-5_tRNA-Ile-TAT-1-1_20 were screened and retained. From 250 small RNA expression features such as zma-miR171l-3p, zma-miR159a-5p, and zma-miR444a, 28 features including zma-miR159a-5p, zma-miR444a, and zma-miR399i-5p were screened and retained.

[0075] like Figure 2As shown, in this embodiment, S3, based on the pre-trained MirGP prediction model, outputs plant phenotypic prediction results through multi-branch feature extraction, multi-level efficient channel attention fusion, and regression prediction. This step aims to deeply fuse and abstract the selected multi-omics features through a carefully designed deep learning architecture, ultimately outputting high-precision phenotypic prediction results; specifically including:

[0076] S31. Input the genotype, transcript expression, and small RNA expression feature subsets into the linear embedding layer for linear transformation, and then concatenate them along the feature dimension. Finally, input them into the first high-efficiency channel attention module for channel weighting to obtain the initial fusion features.

[0077] S32. Genotype, transcript expression, and small RNA expression feature subsets are respectively input into the corresponding feature extraction networks for feature extraction, and then concatenated with the initial fusion features along the feature dimension. After concatenation, the features are input into the second high-efficiency channel attention module for channel weighting to obtain multi-dimensional fusion features. The feature extraction network is composed of mixed depthwise separable convolutional modules.

[0078] S33. The multi-dimensional fused features are extracted into deep features through the top-level hybrid depth separable convolution module, and the plant phenotype prediction results are output through a multilayer perceptron.

[0079] Furthermore, in the feature extraction network, the feature extraction network for processing genotype feature subsets consists of a hybrid depthwise separable convolutional module, the feature extraction network for processing small RNA expression feature subsets consists of two tandem hybrid depthwise separable convolutional modules, and the feature extraction network for processing transcript expression feature subsets consists of two tandem hybrid depthwise separable convolutional modules. This differentiated design considers the inherent characteristics of different omics data: genotype data has high dimensionality but relatively regular structure, and shallow feature extraction can capture the main patterns; transcript expression and small RNA expression data contain more complex sequence structures and regulatory relationships, requiring deeper networks to extract higher-order abstract features.

[0080] Furthermore, the hybrid depthwise separable convolution module adopts the MixConv structure. Unlike traditional convolution which uses a single-size convolution kernel, MixConv uses multiple sizes of convolution kernels in parallel during the depthwise convolution stage, such as 1×1, 3×3, 5×5, and 7×7. The input feature maps are grouped by channel, and each group is subjected to depthwise convolution using a different size convolution kernel. Finally, the outputs of each group are concatenated along the channel dimension.

[0081] Specifically, a typical convolution operation is as follows:

[0082]

[0083] in Represents the input feature map, These represent the height, width, and number of channels of the feature map, respectively. Represents the depthwise convolution kernel. Represents the kernel size. The number of channels in the output feature map;

[0084] In MixConv, the input feature map is divided into multiple feature maps based on the channels, i.e. All feature subgraphs Maintain consistency The sum of them is The corresponding convolution kernel It was also split into Its mathematical expression is:

[0085]

[0086]

[0087] Subsequently, each sub-image is subjected to depthwise convolution and then stitched together along the channel dimension, and then output through batch normalization layer, LeakyReLU activation function and max pooling layer in sequence; the design of MixConv enables the network to capture small-scale local patterns and large-scale global patterns at the same time, enhancing the ability to characterize multi-scale biological signals.

[0088] like Figure 3 As shown, the data processing procedures for both the first and second high-efficiency channel attention modules include:

[0089] Adaptive average pooling is performed on the input feature map X to obtain the channel descriptor Y;

[0090] The one-dimensional convolution kernel size k is adaptively determined based on the number of channels C, and a one-dimensional convolution is performed on the channel descriptor Y to capture local cross-channel interactions, thus obtaining the convolution output. ;

[0091]

[0092] Where γ and b are preset hyperparameters, This indicates taking the nearest odd number.

[0093] The convolution output Channel weights ω are generated by activating the Sigmoid function;

[0094] The input feature map X is multiplied by the channel weight ω channel by channel to obtain the weighted output feature map. ;

[0095] In this embodiment, the design of the two-level high-efficiency channel attention module realizes a progressive feature enhancement that first weights the modalities and then filters the channels across scales: the first level strengthens the contribution features within each modality, while the second level filters the core features most closely related to the target traits from the fused features at different scales, effectively suppressing noise and redundant information.

[0096] Furthermore, the multilayer perceptron includes multiple fully connected layers, each of which performs a linear transformation, and the output dimension of the last fully connected layer is consistent with the target phenotypic dimension to obtain the predicted value. Through the hierarchical design of multi-branch encoding, two-level attention fusion, integrated convolution, and adaptive convergent MLP regression in this embodiment, the MirGP model achieves efficient fusion and accurate prediction of multi-omics data.

[0097] In this embodiment, model training is performed end-to-end. The selected feature subset is divided into training and test sets according to a certain ratio. During training, the SmoothL1 loss function is used to calculate the difference between the predicted and true values. This loss function combines the advantages of L1 and L2 loss, and the gradient is stable when the error is large and the transition is smooth when the error is small. The optimizer is Adam, which can adaptively adjust the learning rate and performs well in sparse gradient scenarios. The network parameters are updated through the backpropagation algorithm until the model converges on the validation set.

[0098] Furthermore, to enhance the biological interpretability of the model, this method also includes a model interpretation step: Specifically, the SHAP value is used to perform feature attribution analysis on the trained MirGP model. The SHAP value is based on the Shapley value in game theory, which can quantify the contribution of each input feature to the final prediction result. For genotype features, SNP sites are aggregated to gene IDs according to gene coordinates, and the SHAP values ​​of the sites within the gene are summed or averaged by sign or absolute value to obtain the gene-level SHAP attribution matrix. For small RNA expression features, small RNA expression is merged by family and arm to avoid redundancy within the family and statistical dilution caused by name variants, to obtain the family and arm-level SHAP attribution matrix.

[0099] Through the above aggregation, two comparable matrices can be obtained for each sample and each trait: gene, sample and family arm, sample, which provides a quantitative basis for analyzing key regulatory factors and verifying biological hypotheses.

[0100] The multi-omics plant phenotypic prediction method fused with small RNA proposed in this embodiment showed significant feature screening and optimization effects in the validation of 16 maize traits, including plant height (PH), ear position (EH), tall leaf width (ELW), tall leaf length (ELL), inflorescence main axis length (TMAL), inflorescence branch number (TBN), number of leaves above ear (LNAE), ear length (EL), ear diameter (ED), cob diameter (CD), number of kernels per row (KNR), weight per 100 kernels (HGW), ear weight (CW), silking time (ST), pollination time (PS), and tasseling time (HD). Figure 4 As shown, for all the maize traits mentioned above, the DNNGP model with genotype as the feature showed a significant improvement in the Pearson r coefficient of prediction performance after Pearson correlation feature screening. This fully demonstrates the effectiveness of the hierarchical feature screening mechanism in reducing data redundancy and improving the basic prediction ability of the model, laying a high-quality data foundation for subsequent multi-omics feature fusion and model construction.

[0101] Meanwhile, the MirGP model constructed in this implementation method achieved comprehensive performance superiority in phenotypic prediction of 16 maize traits, such as... Figure 5 As shown, compared to classic models such as PNNGS, SVR, RF, DNNGP, SoyDNGP, and Ridge, the MirGP model, with G and CE omics as input features, demonstrates leading predictive performance across all maize traits; and in terms of feature contribution, as... Figure 6 As shown, when G and SE are input features, both genotype (G) and miRNA (SE) features contribute to the prediction of 16 maize traits to varying degrees, demonstrating the value of fusing small RNA expression features with genotype features; and as... Figure 7 The results using plant height (PH) as the predicted phenotype show that the sample contribution values ​​of G and SE traits exhibit differentiated trait expression patterns. Combined with the trait attribution analysis of SHAP values, the specific roles of different traits in maize phenotypic regulation can be further analyzed. This method achieves high-precision phenotypic prediction while also possessing good biological interpretability, providing reliable technical support for the genetic analysis and molecular breeding of complex maize traits.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0103] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-omics plant phenotypic prediction method incorporating small RNAs, characterized in that, Includes the following steps: S1. Obtain and preprocess genotype, transcript expression, and small RNA expression data to obtain a standardized feature matrix; S2. Perform feature screening based on Pearson correlation on the standardized feature matrix to obtain the corresponding genotype, transcript expression, and small RNA expression feature subsets. S3. Based on the pre-trained MirGP prediction model, the plant phenotypic prediction results are output through multi-branch feature extraction, multi-level efficient channel attention fusion, and regression prediction; including: Genotype, transcript expression, and small RNA expression feature subsets are respectively input into the linear embedding layer for linear transformation, and then concatenated along the feature dimension before being input into the first high-efficiency channel attention module for channel weighting to obtain the initial fusion features; Genotype, transcript expression, and small RNA expression feature subsets are respectively input into the corresponding feature extraction networks for feature extraction. After being concatenated with the initial fusion feature along the feature dimension, they are input into the second high-efficiency channel attention module for channel weighting to obtain multi-dimensional fusion features. The feature extraction network is composed of mixed depthwise separable convolutional modules. The multi-dimensional fused features are extracted into deep features through the top-level hybrid depth separable convolution module, and the plant phenotype prediction results are output through a multilayer perceptron.

2. The multi-omics plant phenotypic prediction method fused with small RNAs according to claim 1, characterized in that, In S1, the preprocessed genotype data includes: Single nucleotide polymorphism (SNP) sites are encoded as numerical values ​​according to allele counts, where homozygotes are encoded as 0, heterozygotes as 1, and the other homozygotes as 2. Deletion values ​​are denoted as NaN. High deletion sites are filtered out using a preset deletion rate threshold, and rare variants are removed using a preset minor allele frequency threshold. Missing loci were filled using the mode of the column, and the correlation coefficient between loci was calculated within a sliding window. ,when When the value exceeds 0.2, redundant sites are randomly removed to obtain the genotype feature matrix.

3. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 1, characterized in that, In S1, the preprocessed transcript expression and small RNA expression data include: The transcript expression matrix and the small RNA expression matrix are read in and aligned; the small RNA expression matrix includes the miRNA expression matrix, sRNA cluster expression matrix, isomiR expression matrix, and tRF expression matrix. All expression matrices were forcibly converted to numerical values. Unparseable entries were filled with 0.0 and missing values ​​were filled with column mode to obtain the corresponding transcript expression feature matrix and small RNA expression feature matrix.

4. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 1, characterized in that, In step S2, feature selection based on Pearson correlation includes: Calculate the Pearson correlation coefficient r between each molecular feature vector x and the target trait vector y; Calculate the t-statistic based on the degrees of freedom df=n The t-distribution statistic for 2 is used to calculate the two-tailed p-value, where n is the sample size. Filtering p-values ​​less than a preset threshold Features are entered into subsets. Features within subsets are then sorted in ascending order of p-value. If p-values ​​are the same, they are sorted by the absolute value of their correlation coefficients. Determine priority by descending order.

5. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 1, characterized in that, In the feature extraction network, the feature extraction network for processing genotype feature subsets consists of a hybrid depth separable convolutional module, the feature extraction network for processing small RNA expression feature subsets consists of two tandem hybrid depth separable convolutional modules, and the feature extraction network for processing transcript expression feature subsets consists of two tandem hybrid depth separable convolutional modules.

6. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 5, characterized in that, The hybrid depthwise separable convolutional module adopts a MixConv structure, which divides the input feature map into g sub-maps according to channels. The corresponding convolution kernels are divided into Where h and w represent the height and width of the feature map, respectively. , Let represent the number of channels and the kernel size of the t-th subgraph, respectively, and m be the output channel number multiplier; Each sub-image is subjected to depthwise convolution and then stitched together along the channel dimension. The images are then passed through a batch normalization layer, a LeakyReLU activation function, and a max pooling layer in sequence before being output.

7. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 1, characterized in that, The data processing procedures of both the first and second high-efficiency channel attention modules include: Adaptive average pooling is performed on the input feature map X to obtain the channel descriptor Y; The one-dimensional convolution kernel size k is adaptively determined based on the number of channels C, and a one-dimensional convolution is performed on the channel descriptor Y to capture local cross-channel interactions, thus obtaining the convolution output. ; The convolution output Channel weights ω are generated by activating the Sigmoid function; The input feature map X is multiplied by the channel weight ω channel by channel to obtain the weighted output feature map. .

8. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 7, characterized in that, The adaptive determination of the one-dimensional convolution kernel size k based on the number of channels C includes: Where γ and b are preset hyperparameters, This indicates taking the nearest odd number.

9. The multi-omics plant phenotypic prediction method based on fused small RNAs according to claim 1, characterized in that, The multilayer perceptron includes multiple fully connected layers. Each fully connected layer performs a linear transformation, and the output dimension of the last fully connected layer is consistent with the target phenotype dimension to obtain the predicted value.

Images