Crop gene-environment-phenotype prediction method and system
By constructing a multimodal feature extraction and fusion framework and utilizing convolutional neural networks and Bi-LSTM/Transformer models, the problem of insufficient prediction accuracy in traditional breeding methods is solved, and efficient and accurate crop phenotypic prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QIONGTAI TEACHERS COLLEGE
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-01
AI Technical Summary
Traditional crop breeding methods have long breeding cycles and high costs, making it difficult to accurately predict the performance of new gene combinations under different environmental conditions. Furthermore, multi-source heterogeneous data are difficult to integrate efficiently, affecting the accuracy of phenotypic prediction.
A multimodal feature extraction and fusion framework is constructed. Representation learning is performed through convolutional neural networks (CNNs). Combined with genotype data and environmental data, features are extracted using Bi-LSTM and Transformer models, and cross-modal semantic alignment and feature fusion are performed.
It improves the accuracy and efficiency of crop phenotypic prediction and provides a scientific basis to support breeding and planting management.
Smart Images
Figure CN121963893A_ABST
Abstract
Description
A method and system for predicting crop genes, environment, and phenotype. Technical Field
[0001] This application relates to the field of bioinformatics technology, and more specifically, to a method and system for predicting crop gene-environment-phenotype. Background Technology
[0002] The phenotypic traits of crops, such as yield, quality, and resistance, are determined by both genotype and environmental factors. Traditional crop breeding methods mainly rely on extensive field trials and experience accumulation. This approach has many shortcomings, such as long breeding cycles, high costs, and difficulty in accurately predicting the performance of new gene combinations under different environmental conditions, making it difficult to meet the demands of modern agriculture for efficient and precise breeding.
[0003] In recent years, the rapid development of high-throughput gene sequencing technology, bioinformatics, and remote sensing technology has enabled the large-scale acquisition of crop genomic, environmental, and phenotypic data. However, these data sources are highly heterogeneous, have high dimensionality, and vary in quality. How to efficiently integrate multi-source heterogeneous data and explore their intrinsic relationships is the core challenge of current crop phenotypic prediction research. Summary of the Invention
[0004] In view of this, this application provides a crop gene-environment-phenotype prediction method and system, which improves the accuracy of phenotype prediction by constructing an efficient multimodal feature extraction and fusion framework to achieve collaborative analysis of genomic data, environmental data and phenotypic data.
[0005] The technical solution provided in this application is as follows: In a first aspect, this application provides a crop gene-environment-phenotype prediction method, comprising: acquiring genotype data associated with the phenotypic data of the crop to be predicted and environmental data of the target environment; wherein, the genotype data includes single nucleotide polymorphism (SNP) markers, and the target environment includes meteorological environment; extracting features from the preprocessed genotype data and the environmental data, and fusing the obtained genotype features and environmental features into multimodal features; inputting the fused multimodal features into a phenotypic prediction model pre-constructed based on a convolutional neural network (CNN) for representation learning, to obtain the target phenotypic value corresponding to the crop to be predicted.
[0006] One possible implementation involves acquiring genotype data and environmental data of the target environment associated with the phenotypic data of the crop to be predicted, including: filtering the phenotypic data of the crop to be predicted to remove records or outliers from the sowing or harvesting period; performing quality control on the genotype data to remove SNP markers with a MAF of less than 1% and missing values greater than 50%; processing the environmental data to remove environmental data that does not match the genotype and phenotypic data; and marking the filtered phenotypic data with germplasm identifiers (GIDs) and associating it with the corresponding genotype and environmental data.
[0007] In one possible implementation, feature extraction of the genotype data includes: converting the genotype data into a numerical format (0, 1, 2) based on format conversion; where 0 represents the homozygous major allele, 1 represents the heterozygous, and 2 represents the homozygous minor allele; using an embedding layer as a genotype feature extractor to map the genotype data into a high-dimensional continuous vector space; and constructing a genotype expression matrix to capture the semantic features of the SNP markers in the numerical format.
[0008] One possible implementation involves feature extraction of the environmental data, including: establishing an environmental feature extraction model based on Bi-LSTM and Transformer, the model comprising a Bi-LSTM layer and a Transformer encoding layer; wherein the Bi-LSTM layer is used to capture time-series information of multiple environmental variables throughout the entire growth cycle of the crop to be predicted; the Transformer encoding layer is used to convert the environmental data at each time step into a corresponding token vector through a token sequence, and to generate a complete token sequence by combining position embedding (PosEmbed); and the global dependencies of the environmental data are gradually aggregated using multi-head self-attention mechanism (MSA) and layer normalization (LN), and an environmental feature representation is generated through [reg] tokens.
[0009] In one possible implementation, the multimodal feature fusion of the obtained genotype features and environmental features includes: representing the genotype features and environmental features respectively through an embedding layer; performing cross-modal semantic alignment of the genotype features and environmental features based on a multimodal interaction module (MIM) to generate fused multimodal features; wherein the MIM module consists of a self-attention mechanism and a cross-attention mechanism (Cross-Attention), used to progressively update the representations of the genotype features and environmental features.
[0010] One possible implementation involves generating fused multimodal features based on a multimodal interaction module (MIM), including: generating intramodal context representations through a self-attention mechanism. in, Contextual representation of genotypic characteristics within a modality. A contextual representation of the characteristics of the modal environment; This represents the genotype characteristics. The environmental feature representation is represented; a fused genotype feature representation and environmental feature representation are generated through a cross-modal attention mechanism, and are represented as follows: in, This indicates that a fused genotype feature representation is generated through a cross-modal attention mechanism. This indicates that a fused environmental feature representation is generated through a cross-modal attention mechanism; the final genotype feature representation and environmental feature representation are generated through a feedforward neural network (FFN), as follows: in, This represents the final characteristic of the genotype. Represents the final environmental characteristic; positive sample pairs of the genotype final characteristic representation and the environmental final characteristic representation ( , ) as positive samples, with ( , )and( , As negative samples; construct a contrastive loss function to maximize the similarity between positive samples while minimizing the similarity between negative samples, expressed as: Where sim is the cosine similarity function, and τ is the temperature parameter. Let N be the set of positive samples and N be the set of negative samples.
[0011] In one possible implementation, the phenotypic prediction model pre-built based on a convolutional neural network (CNN) includes three convolutional layers, each containing 32 filters with a kernel size of 9 and a stride of 2; each convolutional layer is followed by a flattening layer and a regression head; the regression head contains three fully connected layers; the step of inputting the fused multimodal features into the pre-built phenotypic prediction model for representation learning includes: processing the fused multimodal features through a sliding window of the convolutional layers. Local feature extraction is performed to extract a high-order representation of the interaction between genotype and environment; the output of each convolutional layer undergoes a non-linear transformation of the input x using the Leaky ReLU activation function, as follows: The output of the convolutional layer is normalized by the batch normalization (BN) layer, and the calculation formula is as follows: Where, μ and Here, represents the mean and variance of the input, respectively, and γ and β are learnable scaling and offset parameters. The constant is small; the fully connected layer is used to perform linear transformation and nonlinear activation on the high-order representation of genotype-environment interaction, and the high-order features are gradually mapped to the target phenotypic value.
[0012] In one possible implementation, a Dropout operation is introduced after the first fully connected layer of the regression head, with a dropout rate of 0.3, to reduce the risk of overfitting. The Dropout operation randomly discards a portion of neurons during training, and the calculation formula is expressed as follows: in, Let p be a binary mask generated with probability p.
[0013] Secondly, this application provides a crop gene-environment-phenotype prediction system, including a data acquisition module, a feature extraction module, a multimodal fusion module, and a prediction module. The data acquisition module is used to acquire genotype data associated with the phenotypic data of the crop to be predicted and environmental data of the target environment. The feature extraction module uses an embedding layer as a genotype feature extractor to extract genotype features from the genotype data. The feature extraction module also includes an environmental feature extraction model based on Bi-LSTM and Transformer to extract environmental features from the environmental data. The multimodal fusion module represents the genotype features and environmental features respectively through the embedding layer. A multimodal interaction module (MIM) performs cross-modal semantic alignment of the genotype features and environmental features to generate fused multimodal features. The prediction module inputs the fused multimodal features into a pre-constructed phenotypic prediction model based on a convolutional neural network (CNN) for representation learning to obtain the target phenotypic value corresponding to the crop to be predicted.
[0014] Compared with existing technologies, the technical solution provided in this application has the following beneficial effects: By collecting genotype data, phenotypic data, and environmental data, this application comprehensively obtains information on multiple factors affecting crop phenotypes, providing a rich data foundation for accurate prediction. Genotype features are extracted using embedding layers, and environmental features are extracted using models based on Bi-LSTM and Transformer, mining key information from genotype and environmental data to accurately characterize their respective features. Furthermore, the multimodal interaction module (MIM) performs cross-modal semantic alignment of genotype and environmental features, effectively integrating information from both modalities to generate fused multimodal features, further enhancing the feature representation capability. Finally, the fused features are input into a phenotypic prediction model built on a convolutional neural network (CNN) for representation learning, accurately predicting the target phenotypic value of the crop to be predicted, thus providing a strong scientific basis for crop breeding, planting management, and other related fields. Attached Figure Description
[0015] Figure 1 is a flowchart of a crop gene-environment-phenotype prediction method provided in Embodiment 1 of this application.
[0016] Figure 2 is a flowchart of a crop gene-environment-phenotype prediction method based on contrastive learning provided in Embodiment 2 of this application.
[0017] Figure 3 is a flowchart of the method for deep fusion of genotype features and environmental features through a multimodal interaction module provided in Embodiment 2 of this application.
[0018] Figure 4 is a flowchart of the method for obtaining the target phenotypic value of the crop to be predicted by representation learning based on the fusion features obtained by multimodal fusion according to Embodiment 2 of this application.
[0019] Figure 5 is a schematic diagram of the structure of a crop gene-environment-phenotype prediction system provided in Embodiment 3 of this application. Detailed Implementation
[0020] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0021] Example 1, as shown in Figure 1, is a flowchart of a crop gene-environment-phenotype prediction method provided in Example 1 of this application. As shown in Figure 1, the specific implementation steps of the above method include:
[0022] Step 101: Obtain genotypic data and environmental data of the target environment that are associated with the phenotypic data of the crop to be predicted.
[0023] The aforementioned genotypic data includes, but is not limited to, molecular marker data, such as single nucleotide polymorphism (SNP) markers. The aforementioned phenotypic data includes, but is not limited to, crop trait data, such as crop yield, plant height, and other phenotypic records. The aforementioned environmental data of the target environment includes, but is not limited to, meteorological environmental data.
[0024] After obtaining the above data, this embodiment of the application preprocesses the data, including data cleaning. Specifically, quality control is performed on the genotype data, removing SNP markers with an MAF of less than 1% and missing values greater than 50%. Phenotype data is filtered to remove records from sowing or harvesting periods and outliers. Environmental data is preprocessed by averaging weekly climate data to improve computational efficiency and removing environmental data that does not match the genotype and phenotype records, ultimately obtaining an accurate and high-quality comprehensive dataset to provide a data foundation for subsequent genome prediction research.
[0025] Furthermore, based on format conversion, genotype data is converted into numerical format, phenotypic data is labeled using germplasm identifiers (GIDs), and the labeled phenotypic data is associated with the corresponding genotype data and environmental data.
[0026] As one feasible approach, this application embodiment uses the HapMap format to convert genotype data into a numerical format (0, 1, 2), where 0 represents the major allele of homozygotes, 1 represents heterozygotes, and 2 represents the minor alleles of homozygotes. Phenotypic data is labeled with germplasm identifiers (GIDs) and associated with the corresponding genotype data and environmental data to form a genotype-phenotype-environment dataset, facilitating subsequent phenotypic prediction.
[0027] Step 102: Extract features from the above genotype and environmental data.
[0028] In the first embodiment of this application, based on word embedding technology in the field of natural language processing, the EmbeddingLayer is used as a genotype feature extractor. The gene code is regarded as a word, and a high-dimensional continuous vector space is mapped. A genotype representation matrix is constructed to capture the semantic features of SNP numerical codes, so as to improve the representation capability of genotype data.
[0029] For environmental feature extraction, an environmental feature extraction model based on Bi-LSTM and Transformer is established, including a Bi-LSTM layer and a Transformer encoding layer. Specifically, the Bi-LSTM layer is used to capture the time-series information of multiple environmental variables throughout the entire growth cycle of the crop to be predicted;
[0030] The Transformer encoding layer is used to convert the environment data at each time step into a corresponding token vector through a token sequence, and combines it with position embedding (PosEmbed) to generate a complete token sequence. Multi-head self-attention (MSA) and layer normalization (LN) are used to progressively aggregate the global dependencies of the environment data, generating an environment feature representation through [reg] tokens.
[0031] Step 103: Perform multimodal feature fusion on the obtained genotype features and environmental features.
[0032] Specifically, this application introduces Multimodal Contrastive Learning (MCL) and a Multimodal Interaction Module (MIM). The MCL is used to deeply fuse genotype features and environmental features, and the MCL is used to optimize the fused multimodal feature representation.
[0033] As one possible approach, Embodiment 1 of this application first represents genotype features and environmental features separately through embedding layers. Subsequently, based on the multimodal interaction module (MIM), the genotype features and environmental features are semantically aligned across modalities to generate fused multimodal features.
[0034] Step 104: Input the fused multimodal features into the phenotypic prediction model pre-built based on convolutional neural network (CNN) for representation learning to obtain the target phenotypic value corresponding to the crop to be predicted.
[0035] Compared with the prior art, the technical solution provided in Embodiment 1 of this application has the following beneficial effects: This application uses the multimodal interaction module MIM to perform cross-modal semantic alignment, which effectively solves the problem of differences in dimensionality and semantic expression between genotype features and environmental features, making the fused multimodal features more representative. Then, through the phenotypic prediction model constructed by the convolutional neural network, representation learning can be performed, which can more accurately predict the target phenotypic value of the crop to be predicted, improve the accuracy and efficiency of genome prediction, and effectively solve the problems of complex data processing and insufficient prediction accuracy in traditional methods.
[0036] Example 2, as shown in Figure 2, is a flowchart of a crop gene-environment-phenotype prediction method based on contrastive learning provided in Example 2 of this application. As shown in Figure 2, the specific implementation steps of the above method include:
[0037] Step 201: Obtain the genotype data, phenotypic data, and environmental data of the target environment of the crop to be predicted, and preprocess the above data.
[0038] Example 2 of this application uses wheat as an example. Genotype and phenotypic data of crop wheat are obtained from the database of the International Maize and Wheat Improvement Center (CIMMYT), including single nucleotide polymorphism (SNP) markers and phenotypic records of wheat yield and plant height.
[0039] The aforementioned environmental data was obtained from the AgERA5 dataset, covering daily meteorological data from all locations of the International Wheat Improvement Network (IWIN) from 1979 to 2019, including nine climate variables: precipitation, maximum and minimum relative humidity, shortwave radiation, maximum and minimum temperature, maximum vapor pressure difference, and 2m and 10m wind speeds. In this embodiment, to improve computational efficiency, the weekly climate data is averaged. Climate data for each recorded planting location from two weeks before the sowing date to the end of the harvest date is extracted using the location number (Loc_no) and sowing date (SOWING_DATE) fields as its corresponding environmental data.
[0040] The above data underwent preprocessing, including data cleaning. Specifically, genotypic data underwent quality control, removing SNP markers with an MAF below 1% and missing values exceeding 50%. Phenotypic data was filtered to remove records from sowing or harvesting periods, as well as outliers. Environmental data was preprocessed by averaging weekly climate data to improve computational efficiency and removing environmental data that did not match the genotypic and phenotypic records. The final result was an accurate, high-quality comprehensive dataset, providing a data foundation for subsequent genome prediction research.
[0041] Step 202: Convert genotype data into numerical format based on format conversion, use germplasm identifiers (GIDs) to label phenotypic data, and associate the labeled phenotypic data with the corresponding genotype data and environmental data.
[0042] In this embodiment, genotype data is converted into a numerical format (0, 1, 2) using the HapMap format, where 0 represents the major allele of a homozygote, 1 represents a heterozygote, and 2 represents the minor allele of a homozygote. Phenotypic data is labeled using germplasm identifiers (GIDs) and associated with the corresponding genotype and environmental data to form a genotype-phenotype-environment dataset.
[0043] Step 203: Using an embedding layer as a genotype feature extractor, the above genotype data is mapped to a high-dimensional continuous vector space to construct a genotype expression matrix.
[0044] In this embodiment of the application, based on word embedding technology in the field of natural language processing, an embedding layer is used as a genotype feature extractor. Gene codes are treated as words, and a high-dimensional continuous vector space is mapped. A genotype representation matrix is constructed to capture the semantic features of SNP numerical codes, thereby improving the representation capability of genotype data.
[0045] Specifically, an Embedding Layer is used as the genotype feature extractor, which maintains a trainable lookup matrix. , where d is the embedding dimension. A lookup matrix maps the numerical code (0,1,2) of each SNP to a high-dimensional real-valued vector, generating a genotype representation matrix. , where S is the total number of SNP markers. The embedding layer improves the representation capability of genotype data by capturing the semantic features of SNP numerical codes.
[0046] The embedding layer design is inspired by word embedding techniques in natural language processing. It treats each SNP numerical code (0,1,2) as a specific word, capturing its useful semantic features through mapping to a high-dimensional continuous vector space. The embedding layer uses a trainable lookup matrix. The mapping is implemented, where d is the embedding dimension and 3 is the number of possible values for the SNP numerical code. The columns of the lookup matrix are indexed by each discrete SNP numerical code, and the output is a vector of all query columns. .
[0047] Step 204: Establish an environmental feature extraction model based on Bi-LSTM and Transformer to capture time series information of multiple climate variables throughout the entire growth cycle of the crop to be predicted and extract environmental features.
[0048] Furthermore, in this embodiment, the extracted environmental features are fused with genotype features to provide high-quality environmental information for subsequent multimodal fusion and phenotypic prediction.
[0049] The environmental feature extraction model based on Bi-LSTM and Transformer mainly consists of Bi-LSTM layers and Transformer encoding layers. The Bi-LSTM layers are used to capture time-series information of climate variables, while the Transformer encoding layers are used to learn a global environmental representation.
[0050] Specifically, the Transformer encoding layer converts the climate data at each time step into a corresponding token vector through the token sequence, and combines it with PosEmbed to generate a complete token sequence. Then, the Transformer encoder layer calculates and gradually aggregates the global dependencies of climate variables using Multi-Head Attention and Layer Normalization, and finally generates an environmental feature representation through the [reg] token.
[0051] The calculation method for generating the token sequence in the Transformer encoding layer is as follows: Where v ’ reg For a special [reg]token, v ’ N For token embedding, FC is a fully connected layer, e N PosEmbed represents the environmental data at the Nth time step, where PosEmbed is the location embedding.
[0052] The calculation method for the Transformer encoder layer is as follows: Among them, MSA stands for Multi-head Self-Attention, LN stands for Layer Normalization, and MLP stands for Multilayer Perceptron. Embed the token of the previous layer. This is the feature representation after MSA.
[0053] The calculation method for the multi-head self-attention mechanism in the Transformer coding layer is as follows: in, The weight matrix is a learnable matrix. For query, key, and value input, d ’ `head` is the scaling factor, and `Softmax` is the softmax operation along the row dimension. H For the number of attention heads, , , is a trainable parameter matrix.
[0054] The Transformer encoder layer generates an environment feature representation using the [reg] token. The representation of the [reg] token in the last Transformer encoder layer is extracted and processed through an MLP header to generate the final environment feature embedding. This is used for subsequent phenotypic prediction tasks.
[0055] Step 205: Deeply fuse genotype features and environmental features through the multimodal interaction module, and optimize the fused multimodal feature representation through multimodal contrastive learning (MCL).
[0056] Specifically, genotype features and environmental features are first represented by an embedding layer, then cross-modal semantic alignment is performed by a multimodal interaction module (MIM), and finally fused multimodal features are generated for phenotypic prediction.
[0057] The aforementioned MIM module consists of a self-attention mechanism and a cross-attention mechanism, progressively updating the representations of genotype features and environmental features. As shown in Figure 3, step 205 further includes:
[0058] Step 2051: In each layer, generate intramodal context representations using a self-attention mechanism: in, This represents the contextual representation of genotypic traits within a modality at layer l. This represents the contextual representation of the modal environmental features at layer l. Indicates the first Genotypic characteristics of the layer Indicates the first The environmental characteristics of the layer are represented.
[0059] Step 2052: Generate the fused genotype feature representation and environmental feature representation through a cross-modal attention mechanism, as follows: in, This indicates that a fused genotype feature representation is generated through a cross-modal attention mechanism. This indicates that a fused environmental feature representation is generated through a cross-modal attention mechanism.
[0060] Step 2053: Generate the final genotype feature representation and environmental feature representation using a feedforward neural network (FFN), as follows: in, This represents the final characteristic of the genotype. This represents the final characteristics of the environment.
[0061] Step 2054: Positive sample pairs represented by genotype final characteristics and environmental final characteristics ( , ) as positive samples, with ( , )and( , ) as negative samples.
[0062] in, Indicates and Semantically similar but with Gene features that are not related Indicates and Semantically similar but with An environmental feature representation where no association exists.
[0063] Step 2055: Construct a contrastive loss function to maximize the similarity between positive samples while minimizing the similarity between negative samples.
[0064] The contrastive loss function described above is expressed as follows: Where sim is the cosine similarity function, and τ is the temperature parameter. Let N be the set of positive samples and N be the set of negative samples. Represented by the final genotype characteristics of each layer Aggregation yields, Represented by the final environmental features of each layer Obtained by aggregation.
[0065] Step 206: Input the fused features obtained based on multimodal fusion and the phenotypic data into the convolutional neural network (CNN) for representation learning, thereby predicting the corresponding phenotypic values.
[0066] In this embodiment, multimodal fusion is used to obtain gene-environment fusion features, which are combined with a prediction network built based on convolutional neural network (CNN), and the training process is optimized and dynamically controlled to accurately predict the target trait value.
[0067] In this embodiment, the experimental environment used PyTorch library (version 1.12.1) on a workstation equipped with an Intel(R) Xeon(R) Gold 6133 @ 2.50 GHz CPU, two NVIDIA RTX 3090 GPUs, and Ubuntu 20.04.4 LTS operating system for model training. The Adam optimizer was used for training with an initial learning rate of 0.005 and a weight decay of 0.001 to control the update magnitude and regularization strength of model parameters. 150 epochs were run with a mini-batch size of 128 to balance training efficiency and memory usage. The learning rate was dynamically adjusted to optimize model performance; when the training loss value did not decrease for several consecutive epochs, the learning rate was multiplied by 0.5. The mean squared error (MSE) between the predicted phenotypic values and the true phenotypic records was used as the loss function to optimize the model during training.
[0068] Specifically, the prediction network built with a Convolutional Neural Network (CNN) consists of three 1D convolutional layers, each containing 32 filters with a kernel size of 9 and a stride of 2. Each convolutional layer is followed by a flattening layer and a regression head, which contains three fully connected layers.
[0069] Based on this, as shown in Figure 4, step 206 above specifically includes: step 2061, fusing features through a sliding window of the convolutional layer. Local feature extraction is performed to extract a high-order representation of the genotype-environment interaction. The output of each convolutional layer undergoes a non-linear transformation of the input x using the Leaky ReLU activation function, as expressed in the following formula:
[0070] Step 2062: The output of the convolutional layer is normalized by a batch normalization layer. The calculation formula is as follows: Where, μ and Here, represents the mean and variance of the input, respectively, and γ and β are learnable scaling and offset parameters. It is a small constant used for numerical stability.
[0071] Step 2063: Introduce a Dropout operation after the first fully connected layer of the regression head, with a dropout rate of 0.3, to reduce the risk of overfitting.
[0072] The Dropout operation randomly discards a portion of neurons during training. The calculation formula is as follows: in, Let p be a binary mask generated with probability p.
[0073] Step 2064: Use fully connected layers to perform linear transformation and nonlinear activation on the higher-order representation of genotype-environment interaction, and gradually map the higher-order features to the target phenotypic values.
[0074] Referring to Figure 5, which is a schematic diagram of the structure of a crop gene-environment-phenotype prediction system provided in Embodiment 3 of this application, the crop gene-phenotype prediction system includes a data acquisition module 51, a feature extraction module 52, a multimodal fusion module 53, and a prediction module 54.
[0075] The aforementioned data acquisition module 51 is specifically used to obtain wheat genotype and phenotypic data from the CIMMYT database, and daily meteorological data from 1979 to 2019 from the AgERA5 dataset. It also cleans the obtained genotype, phenotypic, and environmental data, removing missing and outlier values to obtain accurate and high-quality data, providing a data foundation for subsequent genome prediction research. After data processing and integration, a genotype-phenotype-environment dataset is formed, which serves as the input for the subsequent feature extractor.
[0076] The aforementioned feature extraction module 52 is specifically used to construct a genotype feature extractor and an environmental feature extractor based on word embedding technology in the field of natural language processing, extracting genotype features and environmental features respectively, and providing high-quality feature input for the multimodal fusion module 53.
[0077] The aforementioned multimodal fusion module 53 is specifically used to construct a multimodal fusion framework based on gene features and environmental features. It uses the Broadcast Addition operation to fuse environmental features and gene features, introduces MultimodalContrastive Learning and Multimodal Interaction Module to achieve deep feature fusion, and inputs the multimodal fused features into a CNN-based prediction model to regress the target phenotypic value, thereby achieving more accurate genome prediction.
[0078] The aforementioned prediction module 54 is used to load a prediction network based on a convolutional neural network, utilize the genotype-environment-phenotype fusion features provided by the multimodal fusion module 53 to predict the target trait value, and output the prediction results to provide data support for crop genome prediction.
[0079] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the 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, and relevant parts can be referred to the method section.
[0080] 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 crop gene-environment-phenotype prediction method, characterized in that, include: Acquire genotype data and environmental data of the target environment associated with the phenotypic data of the crop to be predicted; wherein the genotype data includes single nucleotide polymorphism (SNP) markers, and the target environment includes meteorological environment; extract features from the preprocessed genotype data and environmental data, and fuse the obtained genotype features and environmental features into multimodal features; input the fused multimodal features into a phenotypic prediction model pre-built based on a convolutional neural network (CNN) for representation learning, to obtain the target phenotypic value corresponding to the crop to be predicted.
2. The crop gene-environment-phenotype prediction method according to claim 1, characterized in that, Acquiring genotype data and environmental data of the target environment associated with the phenotypic data of the crop to be predicted includes: filtering the phenotypic data of the crop to be predicted to remove records or outliers at the sowing or harvesting dates; performing quality control on the genotype data to remove SNP markers with a MAF of less than 1% and missing values greater than 50%; processing the environmental data to remove environmental data that does not match the genotype and phenotypic data; and associating the filtered phenotypic data with the corresponding genotype and environmental data using germplasm identifiers (GIDs).
3. The crop gene-environment-phenotype prediction method according to claim 1, characterized in that, Feature extraction of the genotype data includes: converting the genotype data into a numerical format (0, 1, 2) based on format conversion; where 0 represents the homozygous major allele, 1 represents the heterozygous, and 2 represents the homozygous minor allele; using an embedding layer as a genotype feature extractor to map the genotype data into a high-dimensional continuous vector space; and constructing a genotype expression matrix to capture the semantic features of the SNP markers in numerical format.
4. The crop gene-environment-phenotype prediction method according to claim 1, characterized in that, Feature extraction of the environmental data includes: establishing an environmental feature extraction model based on Bi-LSTM and Transformer, wherein the environmental feature extraction model includes a Bi-LSTM layer and a Transformer encoding layer; wherein the Bi-LSTM layer is used to capture the time series information of multiple environmental variables throughout the entire growth cycle of the crop to be predicted; the Transformer encoding layer is used to convert the environmental data at each time step into a corresponding token vector through a token sequence, and combine it with position embedding (PosEmbed) to generate a complete token sequence; and using multi-head self-attention mechanism (MSA) and layer normalization (LN) to gradually aggregate the global dependencies of the environmental data, and generate an environmental feature representation through [reg] tokens.
5. The crop gene-environment-phenotype prediction method according to claim 1, characterized in that, The step of fusing the obtained genotype features and environmental features into multimodal features includes: representing the genotype features and environmental features respectively through an embedding layer; performing cross-modal semantic alignment of the genotype features and environmental features based on a multimodal interaction module (MIM) to generate fused multimodal features; wherein the MIM module consists of a self-attention mechanism and a cross-attention mechanism (Cross-Attention), used to progressively update the representations of genotype features and environmental features.
6. The crop gene-environment-phenotype prediction method according to claim 5, characterized in that, The multimodal features generated based on the Multimodal Interaction Module (MIM) include: generating intramodal context representations through a self-attention mechanism. in, This represents the contextual representation of genotypic traits within a modality at layer l. This represents the contextual representation of the modal environment features at layer l. Indicates the first Genotypic characteristics of the layer Indicates the first The environmental feature representation of the layer; the fused genotype feature representation and environmental feature representation are generated through a cross-modal attention mechanism, and are represented as follows: in, This indicates that a fused genotype feature representation is generated through a cross-modal attention mechanism. This indicates that a fused environmental feature representation is generated through a cross-modal attention mechanism; the final genotype feature representation and environmental feature representation are generated through a feedforward neural network (FFN), as follows: in, This represents the final characteristic of the genotype. Represents the final environmental characteristic; positive sample pairs of the genotype final characteristic representation and the environmental final characteristic representation ( , ) as positive samples, with ( , )and( , ) as negative samples; where, Indicates and Semantically similar but with Gene features that are not related Indicates and Semantically similar but with Environmental features without correlation are represented; a contrastive loss function is constructed to maximize the similarity between positive samples while minimizing the similarity between negative samples, expressed as: Where sim is the cosine similarity function, and τ is the temperature parameter. Let N be the set of positive samples and N be the set of negative samples. Represented by the final genotype characteristics of each layer Aggregation yields, Represented by the final environmental features of each layer Obtained by aggregation.
7. The crop gene-environment-phenotype prediction method according to claim 1, characterized in that, The pre-built phenotypic prediction model based on a convolutional neural network (CNN) includes three convolutional layers, each containing 32 filters with a kernel size of 9 and a stride of 2. Each convolutional layer is followed by a flattening layer and a regression head. The regression head contains three fully connected layers. The process of inputting the fused multimodal features into the pre-built phenotypic prediction model for representation learning includes: processing the fused multimodal features through a sliding window in the convolutional layers. Local feature extraction is performed to extract a high-order representation of the interaction between genotype and environment; the output of each convolutional layer undergoes a non-linear transformation of the input x using the Leaky ReLU activation function, as follows: The output of the convolutional layer is normalized by the batch normalization (BN) layer, and the calculation formula is as follows: Where, μ and Here, represents the mean and variance of the input, respectively, and γ and β are learnable scaling and offset parameters. The constant is small; the fully connected layer is used to perform linear transformation and nonlinear activation on the high-order representation of genotype-environment interaction, and the high-order features are gradually mapped to the target phenotypic value.
8. The crop gene-environment-phenotype prediction method according to claim 7, characterized in that, The method further includes: introducing a Dropout operation after the first fully connected layer of the regression head, with a dropout rate of 0.3, to reduce the risk of overfitting; the Dropout operation randomly discards a portion of neurons during training, and the calculation formula is expressed as: in, Let p be a binary mask generated with probability p.
9. A crop gene-environment-phenotype prediction system, characterized in that, The system includes a data acquisition module, a feature extraction module, a multimodal fusion module, and a prediction module. The data acquisition module acquires genotype data associated with the phenotypic data of the crop to be predicted and environmental data of the target environment. The genotype data includes single nucleotide polymorphism (SNP) markers, and the target environment includes meteorological conditions. The feature extraction module uses an embedding layer as a genotype feature extractor to extract genotype features from the genotype data. The feature extraction module also includes an environmental feature extraction model based on Bi-LSTM and Transformer to extract environmental features from the environmental data. The multimodal fusion module represents the genotype features and environmental features respectively through embedding layers. A multimodal interaction module (MIM) performs cross-modal semantic alignment of the genotype features and environmental features to generate fused multimodal features. The prediction module inputs the fused multimodal features into a pre-built phenotypic prediction model based on a convolutional neural network (CNN) for representation learning to obtain the target phenotypic value corresponding to the crop to be predicted.