Interpretable crop genome prediction deep learning model

Abstract

The invention discloses an interpretable crop genome prediction deep learning model, puts forward a deep learning framework Cropform, fuses a convolutional neural network (CNN) and a multi-head self-attention mechanism, constructs a technical scheme integrating phenotype prediction and gene mining, automatically extracts local features of genome data through the CNN, and provides an explainable crop genome prediction deep learning model. In combination with a multi-head self-attention mechanism, global association among features is captured to realize high-precision prediction of complex phenotypes, and the prediction accuracy is maximally improved by 7.5% compared with CropGBM, DEM and the like. Key SNPs and genes can be accurately positioned through attention weight and SHAP value analysis, a genetic variation mechanism is disclosed, and multi-modal data fusion of SNP, InDel, gene expression and the like is supported to further improve performance. In order to improve practicability and convenience, the Cropform provides a free online network server. According to the method, the black box limitation of a traditional deep learning model is broken through, analysis of gene-phenotype association is assisted, and an efficient tool is provided for crop genome design and breeding.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of agricultural biotechnology and deep learning, and specifically provides an interpretable deep learning model for crop genome prediction. It involves a crop phenotype prediction and gene mining technology based on genomic sequence data, which can achieve high-precision prediction of complex phenotypes and analysis of key genetic variations through a deep learning framework. Background Technology

[0002] The emergence of genomic selection (GS) has ushered in a paradigm shift from traditional breeding to precision breeding. By integrating molecular markers and phenotypic data through genome-wide prediction models, it shortens the early selection cycle from years to months and can incorporate small-effect QTLs missed by traditional methods. However, current GS technologies have significant limitations. Traditional statistical models struggle to handle the nonlinear interactions between genotype and phenotype, resulting in insufficient accuracy in predicting complex traits. While machine learning models can capture nonlinear relationships, they rely on manual feature engineering, limiting their ability to mine global genome interactions. Deep learning models, although capable of automatically extracting features, cannot analyze key genetic variations due to their "black box" nature, failing to meet the closed-loop requirements of "prediction-analysis-verification" in breeding. Furthermore, the shortcomings of existing technologies in cross-crop adaptability and robustness to data fluctuations further highlight the urgency of developing a deep learning framework that combines high accuracy, strong robustness, and interpretability. This is a key breakthrough point for driving the leap from "experience-driven" to "data-driven" breeding. Summary of the Invention

[0003] This invention provides an interpretable deep learning model for crop genome prediction, which aims to address the problem of insufficient protection against human intervention in existing technologies.

[0004] This invention proposes Cropformer, an interpretable deep learning framework for crop genome prediction. It innovatively integrates convolutional neural networks (CNNs) with multi-head self-attention mechanisms to construct a complete technical solution combining phenotypic prediction and gene mining. CNNs automatically extract local features from genomic data, while the multi-head self-attention mechanism captures global correlations between features, achieving high-precision prediction of complex phenotypes. It demonstrates outstanding performance in over 20 traits across five crops: maize, rice, wheat, millet, and tomato, with prediction accuracy up to 7.5% higher than existing methods such as CropGBM and DEM. Furthermore, attention weight and SHAP value analysis can precisely locate key SNPs and genes, revealing genetic variation mechanisms. It also supports the fusion of multimodal data such as SNPs, InDels, and gene expression to further enhance performance. To improve usability, Cropformer provides a free online web server. This invention overcomes the "black box" limitations of traditional deep learning models, facilitating the analysis of gene-phenotype associations and providing an efficient tool for crop genome design breeding.

[0005] This invention addresses the technical bottlenecks of insufficient interpretability and weak robustness in existing crop genome prediction models by proposing a deep learning framework that integrates convolutional neural networks (CNNs) and multi-head self-attention mechanisms. This framework achieves high-precision prediction of complex crop phenotypes and interpretable analysis of genetic mechanisms, providing a complete technical solution for intelligent breeding. A first aspect of this invention provides an interpretable deep learning model for crop genome prediction.

[0006] It is constructed through the following steps:

[0007] Obtain genotype, phenotype, and environmental data for the target crop. These three types of data are sourced from public databases or proprietary data. All data are uniformly organized into CSV files with column names according to the system's example template, ensuring that the sample identifiers of the three types of data correspond one-to-one. If the original data is in PLINK / PED format, use the provided PED→CSV conversion script and tutorial to complete the formatting without changing the allele information. After completion, upload the CSV file and run it to obtain the prediction results.

[0008] A whole-genome prediction model was constructed, comprising a genotype feature extractor, a feature fusion module, a convolutional neural network module, and a prediction and regression module. The input consisted of genotype sequences in the training set that were selected by the maximum information coefficient (MIC) and fixed to a length of 10,000 loci. The model skeleton adopted a hybrid structure of "1D convolution + multi-head self-attention". The training was randomly divided into 80 / 20 groups and early stopping was used to suppress overfitting. MSE was used as the loss function for the regression task.

[0009] In terms of constructing training and interpretable output, a supervised sample consisting of "fixed-length genotype sequences selected by maximum information coefficient (MIC) and aligned in site order" and "corresponding phenotypic labels" is used, and end-to-end training and validation are carried out within the framework of an interpretable crop genome prediction deep learning model. After training, site-level importance is generated based on multi-head self-attention weights and their aggregation coefficients and mapped to genes / segments to form a list of candidate genetic variations associated with phenotypes and visualization results. In the multimodal implementation, the Top-K features of InDel / SV / gene expression modalities and genotypes can be further used as input to improve prediction performance while maintaining the interpretability of the output results.

[0010] In the above embodiments of the present invention, optionally, for the genotype data file, linkage disequilibrium pruning is first performed using PLINK to obtain the core locus set. After removing samples with missing information, the data is randomly divided into training / testing at 80% / 20% and the random seed is fixed. The maximum information coefficient is calculated in the training set, the 10,000 loci with the highest weights are selected, and the corresponding features are extracted from the test set using the same locus index. The genotype is constructed as the model input feature using a 0-9 encoding scheme to maintain the complete expression of locus allele information. The phenotypic data is aligned with the same sample set, and after removing missing information, it is divided into training / testing at 80% / 20% in the same way as the genotype. Nested cross-validation is used in the training phase, combined with early stopping to suppress overfitting. When discrete trait evaluation is required, continuous traits are divided into three / two categories according to existing rules for classification validation.

[0011] In the above embodiments of the present invention, optionally, the genotype feature extractor is used to extract features from genotype data; the genotype sequence aligned according to the site order is taken as input, and after embedding / linear projection, it is input into a one-dimensional convolution and normalization / residual unit to extract local localization signals; then it is sent to a multi-head self-attention modeling long-range dependency, and learnable weights are introduced into each attention head for weighted aggregation to obtain the genotype representation vector g;

[0012] The feature fusion module is used to fuse features from different genetic modalities; it is used to integrate genetic features in multi-molecular modality embodiments: several high-weight features are selected from SNP, InDel, GE, and SV modalities respectively, and the fused features are formed by column-wise splicing and aligned with g before being sent to the subsequent network; in embodiments that only use genotypes, this module acts as a direct / identical mapping, without changing the dimension and order of g, thereby avoiding redundant processing;

[0013] The convolutional neural network module is used to extract features and learn representations from the fused features; to perform deep representation learning on the fused vector using one to two layers of one-dimensional convolution and residual blocks to enhance the local interactions after fusion; after normalization and activation, a compact representation h is obtained by global average pooling or attention pooling in the sequence dimension, providing a stable input representation for the output head.

[0014] The prediction and regression module is used to regress the features passed through the convolutional neural network module to predict the phenotypic value of the trait; a multilayer perceptron (MLP) is used as the output head to perform nonlinear mapping on the representation h and output the phenotypic value of the continuous trait; during training, MSE is used, and early stopping, Dropout and L2 regularization are combined to control overfitting and improve generalization ability; when the task is switched to classification, only the output head needs to be changed to a classifier and cross-entropy loss is used.

[0015] In the above embodiments of the present invention, optionally, the framework of the model includes a feature processing module, a prediction module, and a parsing module;

[0016] The feature processing module performs three tasks: numerically encoding genomic variations, selecting Top-L sites in the training set using MIC and aligning the indexes of the same sites to the validation / test set, and then performing standardization or linear projection to unify feature dimensions, providing stable input for subsequent modeling.

[0017] The prediction module uses a convolution-attention hybrid backbone for representation learning. On one side, a small-kernel one-dimensional convolution is used to extract neighborhood patterns on the site axis while maintaining the sequence length. On the other side, multi-head self-attention models dependencies in multiple subspaces in parallel, taking into account both short-range LD and long-range effects across segments. Both are normalized / residual and pooled to obtain a compact decision vector. Learnable head weights are configured for each attention head and weighted aggregation is performed to improve robustness and generalization.

[0018] The parsing module first maps the attention distribution and its head-weighted back to genomic coordinates to obtain the importance of sites / segments for screening candidate SNPs and genes; then it calculates the SHAP value in the high-weighted region for secondary sorting and robustness verification, and can perform haplotype analysis to compare phenotypic differences, and finally forms a list of key genetic variations associated with traits and visualization results.

[0019] In the above embodiments of the present invention, optionally, the multi-head self-attention mechanism employs 8-head parallel computation, using... This represents the query vector, indicating the features currently of interest. The key vector represents the features associated with the query vector; The Value vector represents the feature values ​​associated with the query and key; X is the input feature matrix, representing the data after CNN processing. It is the self-attention matrix, representing the relationship between the query vector and the key vector; It is the output of the attention layer, representing the weighted value obtained by calculating the attention mechanism; It is the weight matrix of the query vector; It is the weight matrix of the key vectors; It is the weight matrix of the value vector; It is a key vector The dimension is used to normalize the calculation of the attention matrix; T represents the transpose of the matrix.

[0020] Generate respectively using learnable weight matrices , , Calculate the attention matrix Attention layer output Then, through linear transformation, Dropout, and normalization, computational efficiency and model accuracy are improved.

[0021] In the above embodiments of the present invention, optionally, multimodal data fusion is supported, specifically by extracting the Top 1500 weighted features from SNP, InDel, SV, and GE, constructing fused features by column concatenation, and the prediction performance after fusion is better than that of single-type data.

[0022] In the above embodiments of the present invention, optionally, two forms of prediction results can be output, including phenotypic prediction values ​​of continuous traits and classification results of discrete traits, and key SNPs and associated genes can be labeled.

[0023] In the above embodiments of the present invention, optionally, the system software is deployed on a crop breeding platform to support researchers in uploading genomic data and obtaining prediction results and genetic analysis reports.

[0024] After adopting the above technical solution, the present invention mainly has the following effects:

[0025] The data engineering is robust and highly reproducible. Unified numerical coding, MAF / missing data quality control, MIC-based fixed-length point selection, and cross-dataset point index alignment significantly reduce noise and redundancy, ensuring that data from different crops and batches can be stably reused and experimental results reproduced under the same process.

[0026] It exhibits robust generalization and stable convergence. Under the framework of 80 / 20 training / test split and nested cross-validation, it combines early stopping, Dropout and L2 regularization, and introduces learnable head weight aggregation for multi-head attention, which effectively suppresses overfitting and head imbalance, and improves generalization performance and training stability on new data.

[0027] Its representation capabilities take into account both local and long-range dependencies. Convolutional layers extract neighborhood patterns, and self-attention models cross-segment dependencies in parallel subspaces, enabling it to capture both nearest-neighbor LDs and long-range effects simultaneously. Compared to backbones that only use convolution or attention, it is more suitable for modeling complex quantitative traits, with better prediction accuracy and robustness.

[0028] Interpretability forms an "evidence chain." First, it obtains site-level weights based on multi-head attention, then introduces SHAP / XGBoost within the candidate window for secondary verification, and supplements it with haplotype analysis to output gene / site-level evidence and visualization results, thus realizing a closed loop from prediction to mechanism analysis, which facilitates scientific research review and breeding decisions.

[0029] It boasts efficient inference capabilities and ease of deployment. Fixed-length input and a rule-based network structure enable controllable memory usage and latency. The interface is clear and can be quickly integrated into online services or local platforms. It can complete batch inference and interpretation report generation with a single click, resulting in low engineering implementation costs. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a diagram of an interpretable deep learning model architecture for crop genome prediction.

[0032] Figure 2 This is a horizontal comparison chart of prediction results from similar models.

[0033] Figure 3 This is a diagram showing the results of the genetic analysis.

[0034] Figure 4 This is a screenshot of the online server interface. Detailed Implementation

[0035] To better understand the technical solution of this invention, it should be understood that the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0036] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be described in further detail. However, the described embodiments are only some, not all, of the embodiments of the present invention. Other embodiments based on the embodiments of the present invention, without inventive effort by those skilled in the art, are all within the scope of protection of the present invention.

[0038] To make the technical problems, technical solutions and beneficial effects of the present invention clearer, the following specific embodiments are provided, but are not intended to limit the scope of the present invention.

[0039] This invention provides an interpretable deep learning model for crop genome prediction. Addressing the technical bottlenecks of insufficient interpretability and weak robustness in existing crop genome prediction models, this invention proposes a deep learning framework that integrates convolutional neural networks (CNNs) and multi-head self-attention mechanisms. This framework achieves high-precision prediction of complex crop phenotypes and interpretable analysis of genetic mechanisms, providing a complete technical solution for intelligent breeding. Detailed technical solutions are as follows:

[0040] Step 1: Multimodal data preprocessing and feature engineering

[0041] Data Acquisition and Encoding: Crop genome variation data were acquired, SNPs were processed using the 0-9 encoding rule, and InDel and SV were encoded as 0, 1, and 2 using the PLINK tool.

[0042] Baseline dataset construction: Low-quality SNPs with a suballele frequency of less than 0.05 were filtered out; feature importance was calculated based on the maximum information coefficient, and the top 10,000 features strongly associated with the phenotype were selected, while weakly related information was removed to reduce redundancy.

[0043] Data standardization: Standardize and scale the selected features to eliminate differences in units and ensure the stability of model training.

[0044] Step 2: Adaptive Feature Extraction and Attention Enhancement

[0045] CNN Local Feature Extraction: A CNN layer is constructed using 3×3 convolutional kernels to automatically extract local features of the genome sequence, transforming the original input into a high-dimensional feature representation.

[0046] Multi-head self-attention global association capture: Introducing an 8-head multi-head self-attention mechanism, through calculation Generate attention weights. After function normalization, the global correlation between features is dynamically captured, generating a decision vector for phenotypic prediction.

[0047] Step 3: Multi-task prediction and genetic analysis

[0048] Multi-type trait prediction: Supports regression prediction of continuous traits and classification prediction of discrete traits, directly outputting phenotypic prediction values ​​or classification results.

[0049] Genetic variation analysis: Weights of attention mechanisms are extracted, and key SNPs are annotated to associated genes using SHAP value analysis and the XGBoost extension module. Candidate causal loci are located through haplotype analysis to reveal the association mechanism between genetic variation and phenotype.

[0050] Step 4: Model Optimization and Deployment

[0051] Overfitting prevention: Nested cross-validation is used to optimize hyperparameters, combined with an early stopping strategy to improve the robustness and generalization ability of the model.

[0052] Multimodal fusion optimization: Supports fusion of multiple data types such as SNP, InDel, SV, and GE: Extracts the top 1500 weighted features from each data type, and constructs a fusion feature matrix by column concatenation to further improve the prediction accuracy of complex traits; The model can be deployed on online servers or local breeding platforms, and supports users to upload data and obtain prediction results and analysis reports.

[0053] Specifically Figure 1 This paper illustrates the architecture of an interpretable deep learning model for crop genome prediction based on the present invention. It includes the architecture flow of the Cropformer model. This model takes genomic variation data from crops such as maize, rice, wheat, millet, and tomato as input. First, it processes SNPs using a 0-9 encoding rule, and then uses PLINK to encode InDel and SV as 0, 1, and 2. Next, it selects the Top 10,000 features associated with the phenotype based on the Maximum Information Coefficient (MIC), and then standardizes the training and test sets after splitting them in an 8:2 ratio. Figure 1 As shown in A and B, the core of the model consists of a CNN layer and an 8-head multi-head self-attention mechanism. The CNN layer uses 3×3 convolutional kernels to extract local features of the genome. The output features are processed by the self-attention mechanism, and attention weights are generated by calculation. After being normalized by softmax, global associations are captured, and finally, a decision vector is generated for phenotypic prediction, while also supporting subsequent genetic analysis.

[0054] Specifically, an interpretable deep learning model for crop genome prediction is constructed through the following steps:

[0055] Genotype, phenotype, and environmental data for the target crop can be obtained from public databases or proprietary data. All data should be uniformly organized into CSV files with column names according to the system's example template, ensuring that the sample identifiers of the three types of data correspond one-to-one. If the original data is in PLINK / PED or other formats, it is preferable to use the provided PED→CSV conversion script and tutorial to complete the formatting without changing the allele information. After completion, upload the CSV file and run it to obtain the prediction results.

[0056] For genotype files, linkage disequilibrium pruning is first performed using PLINK to obtain the core locus set. After removing samples with missing information, the training / testing sets are randomly divided at 80% / 20% with a fixed random seed. The maximum information coefficient is calculated in the training set, and the 10,000 loci with the highest weights are selected. Corresponding features are extracted from the test set using the same locus index. The genotypes are constructed as model input features using a 0-9 encoding scheme to maintain the complete expression of locus allele information. Phenotypic data are aligned with the same sample set, and after removing missing information, the training / testing sets are divided at 80% / 20% consistent with the genotypes. Nested cross-validation combined with early stopping is used during the training phase to suppress overfitting. When discrete trait evaluation is required, continuous traits are divided into three / two categories according to existing rules for classification validation.

[0057] A whole-genome prediction model was constructed, comprising a genotype feature extractor, a feature fusion module, a convolutional neural network module, and a prediction and regression module. The input consisted of genotype sequences in the training set that were selected by the maximum information coefficient (MIC) and fixed to a length of 10,000 loci. The model skeleton adopted a hybrid structure of "1D convolution + multi-head self-attention". The training was randomly divided into 80 / 20 groups and early stopping was used to suppress overfitting. MSE was used as the loss function for the regression task.

[0058] The genotype feature extractor is used to extract features from genotype data. The genotype sequence aligned according to the site order is taken as input, and after embedding / linear projection, it is input into one-dimensional convolution and normalization / residual units to extract local localization signals. Then, it is fed into multi-head self-attention modeling long-range dependencies, and learnable weights are introduced into each attention head for weighted aggregation to obtain the genotype representation vector g.

[0059] The feature fusion module is used to fuse features from different genetic modalities; it is used to integrate genetic features in multi-molecular modality embodiments: several high-weight features are selected from each of the SNP, InDel, GE, SV and other modalities, and the fused features are formed by column-wise splicing and aligned with g before being sent to the subsequent network; in embodiments that only use genotypes, this module acts as a direct / identical mapping, without changing the dimension and order of g, thereby avoiding redundant processing;

[0060] The convolutional neural network module is used to extract features and learn representations from the fused features; to perform deep representation learning on the fused vector using one to two layers of one-dimensional convolution and residual blocks to enhance the local interactions after fusion; after normalization and activation, a compact representation h is obtained by global average pooling or attention pooling in the sequence dimension, providing a stable input representation for the output head.

[0061] The prediction and regression module is used to regress the features passed through the convolutional neural network module to predict the phenotypic value of the trait; a multilayer perceptron (MLP) is used as the output head to perform nonlinear mapping on the representation h and output the phenotypic value of the continuous trait; during training, MSE is used, and early stopping, Dropout and L2 regularization are combined to control overfitting and improve generalization ability; when the task is switched to classification, only the output head needs to be changed to a classifier and cross-entropy loss is used.

[0062] In terms of training and interpretable output, supervised samples consisting of "fixed-length genotype sequences selected by MIC and aligned in order of loci + corresponding phenotypic labels" are trained and validated end-to-end within the above framework. After training, locus-level importance is generated based on multi-head attention weights and their aggregation coefficients and mapped to genes / segments to form a list of candidate genetic variations associated with phenotypes and visualization results. In the multimodal implementation, Top-K features of modalities such as InDel / SV / gene expression can be further used together with genotypes as input to improve prediction performance while maintaining the interpretability of the output results.

[0063] The framework of the model includes a feature processing module, a prediction module, and a parsing module;

[0064] The feature processing module performs three tasks: numerically encoding genomic variations, selecting Top-L sites in the training set using MIC and aligning the indexes of the same sites to the validation / test set, and then performing standardization or linear projection to unify feature dimensions, providing stable input for subsequent modeling.

[0065] The prediction module uses a convolution-attention hybrid backbone for representation learning. On one side, a small-kernel one-dimensional convolution is used to extract neighborhood patterns on the site axis while maintaining the sequence length. On the other side, multi-head self-attention models dependencies in multiple subspaces in parallel, taking into account both short-range LD and long-range effects across segments. Both are normalized / residual and pooled to obtain a compact decision vector. Learnable head weights are configured for each attention head and weighted aggregation is performed to improve robustness and generalization.

[0066] The parsing module first maps the attention distribution and its head-weighted back to genomic coordinates to obtain the importance of sites / segments for screening candidate SNPs and genes; then it calculates the SHAP value in the high-weighted region for secondary sorting and robustness verification, and can perform haplotype analysis to compare phenotypic differences, and finally forms a list of key genetic variations associated with traits and visualization results.

[0067] The multi-head self-attention mechanism employs 8-head parallel computation, using... This represents the query vector, indicating the feature currently of interest. A key vector represents the features associated with the query vector; The value vector represents the feature values ​​associated with the query and key; X is the input feature matrix, representing the data after CNN processing. It is the self-attention matrix, representing the relationship between the query vector and the key vector; It is the output of the attention layer, representing the weighted value obtained by calculating the attention mechanism; It is the weight matrix of the query vector; It is the weight matrix of the key vectors; It is the weight matrix of the value vector; It is a key vector The dimension is used to normalize the calculation of the attention matrix; T represents the transpose of the matrix.

[0068] Generate respectively using learnable weight matrices , , Calculate the attention matrix Attention layer output Then, through linear transformation, Dropout, and normalization, computational efficiency and model accuracy are improved.

[0069] It supports multimodal data fusion, specifically by extracting the top 1500 weighted features from SNP, InDel, SV, and GE, and constructing fused features by concatenating columns. The prediction performance after fusion is better than that of single-type data.

[0070] It can output two prediction result formats, including phenotypic prediction values ​​for continuous traits and classification results for discrete traits, and can annotate key SNPs and associated genes.

[0071] It can be deployed as system software on crop breeding platforms, supporting researchers to upload genomic data and obtain prediction results and genetic analysis reports.

[0072] Figure 2 The prediction performance of Cropformer is compared with that of existing models. Figure 2 Figures A through F show that on the maize dataset, the model's Pearson correlation coefficients (PCCs) for predicting heading date (DTT), plant height (PH), and ear weight (EW) reached 92.2%, 91.8%, and 76.3%, respectively, significantly outperforming models such as DEM (DTT=89.5%, PH=88.7%) and CropGBM (EW=70.8%). Through nested cross-validation and an early termination strategy, the model's robustness on the test set was effectively ensured, further validating its predictive reliability.

[0073] Figure 3The genetic analysis results of Cropformer are presented. By extracting attention weights, in the Manhattan plot of SNPs related to tasseling time (DTT) in maize, known flowering genes such as Zm00001d008941 (ATX3) and Zm00001d011956 (SDG118) have significant weights. Haplotype analysis shows that among the five haplotypes of ATX3, haplotype IV has the shortest DTT, and haplotype IV of SDG118 is also associated with early flowering. Combined with SHAP value analysis, the chr8:26,168,415 loci of ATX3 and the chr8:165,145,056 loci of SDG118 are key variants, validating the model's ability to capture and analyze quantitative trait loci.

[0074] Figure 4 The Cropformer online server interface is shown. Users can access this tool, select the target crop module, download sample data in CSV format or upload their own data, and click "Run" to obtain phenotypic predictions or classification results. It also outputs an analytical report containing attention weight heatmaps and haplotype networks, providing breeders with a convenient and practical tool.

[0075] Key inventive points and beneficial effects of this application:

[0076] Standardize data links, ensure reproducible results, establish a unified data processing workflow, effectively reduce noise and dimensionality, support reuse of different crops and batches under the same pipeline, and ensure consistent and comparable conclusions.

[0077] The training is more stable and the generalization is better. Under the framework of 80 / 20 splitting and nested cross-validation, combined with early stopping, Dropout, L2 regularization, and learnable aggregation for multi-head output, overfitting and head imbalance are suppressed, resulting in more stable convergence and more reliable performance on new data.

[0078] Local-long-range relationship collaborative modeling combines local pattern extraction with global dependency characterization of parallel subspaces, covering both nearest-neighbor LD and cross-segment effects; compared to single-type backbones, it is more suitable for complex quantitative traits, and both prediction accuracy and robustness are improved.

[0079] The interpretable closed loop from prediction to evidence first obtains the importance of loci / segments from attention distribution, then verifies it with SHAP / XGBoost within the candidate window, and supplements it with haplotype comparison to output loci / gene-level evidence and visualization reports, forming an auditable and verifiable chain of interpretations.

[0080] With controllable resources and easy integration and deployment, fixed-length input and rule-based structure make memory and latency controllable. The interface is compatible with common tools and can be quickly connected to online services or local platforms to achieve one-click batch inference and report generation, significantly reducing deployment and maintenance costs.

[0081] The above descriptions are merely embodiments of the present invention. Common knowledge such as specific structures and characteristics known in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the technical field to which the invention pertains before the application date or priority date, are able to obtain all prior art in the field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can improve and implement the solution based on the inspiration given in this application and their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application.

[0082] It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of this invention, and these should also be considered within the scope of protection of this invention. These modifications and improvements will not affect the effectiveness of the invention or the practicality of the patent. The scope of protection claimed in this application shall be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An interpretable crop genome prediction deep learning model, characterized in that, include: It is constructed through the following steps: Obtain genotype, phenotype, and environmental data for the target crop. These three types of data are sourced from public databases or proprietary data. All data are uniformly organized into CSV files with column names according to the system's example template, ensuring that the sample identifiers of the three types of data correspond one-to-one. If the original data is in PLINK / PED format, use the provided PED→CSV conversion script and tutorial to complete the formatting without changing the allele information. After completion, upload the CSV file and run it to obtain the prediction results. A whole-genome prediction model was constructed, comprising a genotype feature extractor, a feature fusion module, a convolutional neural network module, and a prediction and regression module. The input consisted of genotype sequences in the training set that were selected by the maximum information coefficient (MIC) and fixed to a length of 10,000 loci. The model skeleton adopted a hybrid structure of "1D convolution + multi-head self-attention". The training was randomly divided into 80 / 20 groups and early stopping was used to suppress overfitting. MSE was used as the loss function for the regression task. In terms of constructing training and interpretable output, a supervised sample consisting of "fixed-length genotype sequences selected by maximum information coefficient (MIC) and aligned in site order" and "corresponding phenotypic labels" is used, and end-to-end training and validation are carried out within the framework of an interpretable crop genome prediction deep learning model. After training, site-level importance is generated based on multi-head self-attention weights and their aggregation coefficients and mapped to genes / segments to form a list of candidate genetic variations associated with phenotypes and visualization results. In the multimodal implementation, the Top-K features of InDel / SV / gene expression modalities and genotypes can be used together as input to improve prediction performance while maintaining the interpretability of the output results.

2. The interpretable deep learning model for crop genome prediction according to claim 1, characterized in that, For genotype data files, linkage disequilibrium pruning is first performed using PLINK to obtain the core locus set. After removing samples with missing information, the data is randomly divided into training and testing sets at 80% / 20% with a fixed random seed. The maximum information coefficient is calculated in the training set, and the 10,000 loci with the highest weights are selected. Corresponding features are extracted from the test set using the same locus index. The genotypes are constructed as model input features using a 0-9 encoding scheme to maintain the complete expression of locus allele information. Phenotypic data are aligned with the same sample set, and after removing missing information, the data is divided into training and testing sets at 80% / 20% in the same manner as the genotypes. Nested cross-validation is used in the training phase, combined with early stopping to suppress overfitting. When discrete trait evaluation is required, continuous traits are divided into three / two categories according to existing rules for classification validation.

3. The interpretable deep learning model for crop genome prediction according to claim 1, characterized in that, The genotype feature extractor is used to extract features from genotype data. The genotype sequence aligned according to the site order is taken as input, and after embedding / linear projection, it is input into one-dimensional convolution and normalization / residual units to extract local localization signals. Then, it is fed into multi-head self-attention modeling long-range dependencies, and learnable weights are introduced into each attention head for weighted aggregation to obtain the genotype representation vector g. The feature fusion module is used to fuse features from different genetic modalities; it is used to integrate genetic features in multi-molecular modality embodiments: several high-weight features are selected from SNP, InDel, GE, and SV modalities respectively, and the fused features are formed by column-wise splicing and aligned with g before being sent to the subsequent network; in embodiments that only use genotypes, this module acts as a direct / identical mapping, without changing the dimension and order of g, thereby avoiding redundant processing; The convolutional neural network module is used to extract features and learn representations from the fused features; The fused vector is subjected to one to two layers of one-dimensional convolution and residual block deep representation learning to enhance the local interaction after fusion. After normalization and activation, a compact representation h is obtained by global average pooling or attention pooling in the sequence dimension, which provides a stable input representation for the output head. The prediction and regression module is used to regress the features passed through the convolutional neural network module to predict the phenotypic value of the trait; a multilayer perceptron (MLP) is used as the output head to perform nonlinear mapping on the representation h and output the phenotypic value of the continuous trait; during training, MSE is used, and early stopping, Dropout and L2 regularization are combined to control overfitting and improve generalization ability; when the task is switched to classification, only the output head needs to be changed to a classifier and cross-entropy loss is used.

4. The interpretable crop genome prediction deep learning model according to claim 1, characterized in that, The framework of the model includes a feature processing module, a prediction module, and a parsing module; The feature processing module performs three tasks: numerically encoding genomic variations, selecting Top-L sites in the training set using MIC and aligning the indexes of the same sites to the validation / test set, and then performing standardization or linear projection to unify feature dimensions, providing stable input for subsequent modeling. The prediction module uses a convolution-attention hybrid backbone for representation learning. On one side, a small-kernel one-dimensional convolution is used to extract neighborhood patterns on the site axis while maintaining the sequence length. On the other side, multi-head self-attention models dependencies in multiple subspaces in parallel, taking into account both short-range LD and long-range effects across segments. Both are normalized / residual and pooled to obtain a compact decision vector. Learnable head weights are configured for each attention head and weighted aggregation is performed to improve robustness and generalization. The parsing module first maps the attention distribution and its head-weighted back to genomic coordinates to obtain the importance of loci / segments for screening candidate SNPs and genes. Then, it calculates SHAP values ​​in high-weighted regions for secondary sorting and robustness verification, and can perform haplotype analysis to compare phenotypic differences, ultimately forming a list of key genetic variations associated with traits and visualization results.

5. An interpretable deep learning model for crop genome prediction according to claim 2 or 3, characterized in that, The multi-head self-attention mechanism employs 8-head parallel computation, using... This represents the query vector, indicating the features currently of interest. The key vector represents the features associated with the query vector; The Value vector represents the feature values ​​associated with the query and key; X is the input feature matrix, representing the data after CNN processing. It is the self-attention matrix, representing the relationship between the query vector and the key vector; It is the output of the attention layer, representing the weighted value obtained by calculating the attention mechanism; It is the weight matrix of the query vector; It is the weight matrix of the key vectors; It is the weight matrix of the value vector; It is a key vector The dimension is used to normalize the calculation of the attention matrix; T denotes the transpose of the matrix; Generate respectively using learnable weight matrices , , Calculate the attention matrix Attention layer output Then, through linear transformation, Dropout, and normalization, computational efficiency and model accuracy are improved.

6. The interpretable deep learning model for crop genome prediction according to claim 1, characterized in that, It supports multimodal data fusion, specifically by extracting the top 1500 weighted features from SNP, InDel, SV, and GE, and constructing fused features by concatenating columns. The prediction performance after fusion is better than that of single-type data.

7. The interpretable deep learning model for crop genome prediction according to claim 1, characterized in that, It can output two prediction result formats, including phenotypic prediction values ​​for continuous traits and classification results for discrete traits, and can annotate key SNPs and associated genes.

8. The interpretable deep learning model for crop genome prediction according to claim 1, characterized in that, Deployed as system software on crop breeding platforms, it supports researchers in uploading genomic data and obtaining prediction results and genetic analysis reports.

Images