RWKV-based multi-scale G*E interaction yield prediction method and system

Abstract

The invention provides an RWKV-based multi-scale G * E interaction yield prediction method and system, and the method comprises the steps: firstly extracting local features between genotype sites through employing a convolutional neural network, and capturing a global dependency relationship between the sites through employing a linear attention mechanism; then, respectively capturing features among all environment factors and multi-scale features in a whole growth period by using a time mixing module and a channel mixing module of the RWKV; and finally, fusing the cross-modal features to carry out linear fitting to obtain a predicted yield value. Constructing a multi-dimensional feature space by constructing an RWKVGP model and integrating a genotype feature vector, a light environment feature vector and a surface environment feature vector; the multi-level feature fusion mechanism significantly improves the characterization ability of the model to a genotype-environment interaction complex association mode, analyzes the interaction effect between different environment factors and genotypes, realizes the function of improving the accuracy of a genome prediction task, and accelerates the breeding selection efficiency of corn crops.

Description

Technical Field

[0001] This invention belongs to the field of genome prediction technology, specifically relating to a method and system for predicting yields based on multi-scale G×E interactions using RWKV. Background Technology

[0002] With global warming and the increasing frequency of extreme weather events, crop yields are facing significant threats. Simultaneously, the growing global population, decreasing arable land, and rising food demand make improving crop yields through breeding crucial for food security and agricultural production in my country and globally. Genomic selection, a vital component of modern breeding technology, accelerates genetic improvement by utilizing genome-wide genetic markers to predict individual phenotypes. Traditional genotype-environment (G×E) models typically assume fixed genotype effects unaffected by the environment. However, phenotypic traits are influenced not only by genetic factors but also by significant environmental conditions such as climate, soil, and pests. This genotype-environment interaction (G×E) often makes phenotypic selection results unstable, especially under variable environmental conditions. This phenomenon is known as genotype-environment interaction. G×E describes the phenotypic differences between different genotypes under various environmental conditions. In breeding practice, this means that some genotypes may perform well in specific environments but poorly or moderately in others. Therefore, understanding and utilizing G×E is essential for improving breeding efficiency and accuracy.

[0003] Deep learning models have achieved some success in G×E predictions, but most of the deep learning models used in current research have simple architectures, such as pure CNNs, MLPs, and SVMs. These simple models do not specifically process genotype and environmental data, and do not fully integrate the locus features of genotype data with the multi-scale characteristics of environmental data, making it difficult to handle the complex interactions between genotype and environment in real-world scenarios. Therefore, to better predict crop yields and other phenotypes and save time and money in field trials, it is necessary to propose more targeted deep learning models that can further capture the characteristic relationships between genotype and environmental data. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for predicting yields based on multi-scale G×E interactions using RWKV, so as to improve the accuracy of genome prediction tasks.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for predicting yield based on multi-scale G×E interaction of RWKV, comprising the following steps: S1: Acquire whole-genome data, multi-scale environmental data, and yield phenotypic data, and perform preprocessing; S2: Use convolutional neural networks to extract local features between genotype loci, and use multiple temporal and channel mixing modules of linear attention mechanism to capture global features between genotype loci; S3: Using a linear attention mechanism, multi-scale and temporal features of environmental factors are captured by type to obtain illumination environment features and surface environment features; S4: Using the select and concat mechanisms, cross-modal features are fused to obtain genotype-light environment features, light environment-land surface environment features, and land surface environment-genotype features; S5: Use a fully connected neural network to fit and regress the fused features to obtain the predicted yield value of the material.

[0006] According to the above scheme, the specific steps in step S1 are as follows: Whole genome data of breeding materials were obtained through sequencing, and the whole genome data was compared with breeding microarray data to obtain a length of [missing information]. l SNP locus sequence genotype data G= { g 1, g 2 …g l}, g i ∈{0, 1, 2}; Collect multi-scale environmental factor data related to light at crop planting sites, from planting time to harvest time; Collect yield-related agronomic phenotypic data for each planting environment.

[0007] According to the above scheme, the specific steps in step S2 are as follows: S21: Sequence of SNP sites G Inputting features into a CNN layer and compressing them through convolutional operations yields... ; S22: Output of the CNN layer The input is normalized by the BN layer to obtain the output. ; S23: The output after passing through multiple CNN layers and BN layers Inputting a max pooling layer and a Dropout layer, we obtain the following by reducing the data dimensionality and randomly dropping some neurons: ; S24: The result after convolution and pooling The input is an RWKV layer, which extracts genotype data features from a global dimension through a linear attention mechanism, thereby capturing the feature dependencies between distant SNP sites and obtaining genotype features. .

[0008] Furthermore, the specific steps in steps S21, S22, and S23 are as follows: set up W c，i，j These are the convolutional kernel weights of the CNN layer, representing the weights of the convolutional kernels of the CNN layer c The first channel i On the input channel, the first j Then, one parameter for: ; set up and These are the BN layers. c Mean and variance within each channel: , The genotype feature vector extracted by the BN layer from the CNN layer After standardization, the following is obtained: ; The genotype feature vectors obtained after multiple CNN and BN layers Inputting a max pooling layer reduces data dimensionality by selecting the most salient features within the pooling window; a dropout layer randomly discards some neurons during training to reduce the risk of overfitting; resulting in: ; The output genotype is obtained after multiple maximum pooling layers and Dropout layers. .

[0009] Furthermore, in step S24, the specific steps are as follows: The TimeShift operation merges the current feature vector with the previous feature vector to form a new vector; it shifts the first half of the input genotype sequence's features by one time step in the time dimension, thereby capturing the dependencies between adjacent time steps; let the input sequence... , feature dimensions C Divided according to location and ; Features of the first half In the time dimension t Shifting upwards by one time step is:

[0010] Features of the second half Defined as: , The offset first half and the unoffset second half are concatenated along the feature dimension to generate a new feature. : ; set up u r , u k and u v For learnable mixing coefficients, TimeShift is applied to the input genotype token matrix: , , ; The input to the attention mechanism in the temporal hybrid block is a weighted sum of the current token and the previous token, and its output acts as cross-attention in the standard transformer. The current recursive state is... wkv t for: ; The channel mixing block uses an RNN-like inference and decoding method to process the state. t The output is used as the state. t+1 The input is decoded recursively; let... u r ，u z For learnable mixing coefficients, output O t Includes time t Historical information, channel blending modes are shown below: , , The gating mechanism generates a gating factor through a sigmoid activation function, which is used to adjust the weights at different time steps to obtain the final genotype feature output. The genotype features to be retained are determined by a sigmoid activation function, and the available information is filtered using a ⊙ operation to serve as the vector representation of the next token. ; The output after passing through the CNN layer and the max pooling layer The input is processed through multiple stacked time mixing and channel mixing modules to generate genotype feature vectors. G* .

[0011] According to the above scheme, the specific steps in step S3 are as follows: enter d Within a day e Feature vector of each light-related environmental factor Ex (1) ∈ R d×e We used a linear attention mechanism to capture environmental characteristics throughout the entire growth period from multi-scale illumination data and multi-scale surface data.

[0012]

[0013]

[0014]

[0015] Obtain multi-scale illumination environment feature output vector Ex* ; enter d Within a day e Feature vector of each surface-related environmental factor Ey (1) ∈ R d×e We utilize a linear attention mechanism to capture environmental features throughout the entire growth period of multi-scale surface data, resulting in a multi-scale surface environmental feature output vector. Ey* .

[0016] According to the above scheme, the specific steps in step S4 are as follows: Let Concat represent the splicing operation, where features are fused to obtain the genotype-light environment feature vector. GE x Light environment - surface environment characteristic vector E x E y and surface environment-genotype feature vector GE y for: , .

[0017] According to the above scheme, the specific steps in step S5 are as follows: The fused feature vectors are uniformly input into a linear layer, and a vector of uniform dimension is obtained through linear mapping. The uniformly dimension-reduced feature vectors are then concatenated and mapped to the predicted output value of the model through a linear transformation. , , ; A multidimensional feature space was constructed by integrating genotype feature vectors, illumination environment feature vectors, and surface environment feature vectors.

[0018] A system for predicting yields based on multi-scale G×E interaction using RWKV is characterized by: The data acquisition submodule is used to acquire whole-genome data, multi-scale environmental data, and yield phenotypic data, and to perform preprocessing. The gene feature submodule is used to extract local features between genotype loci using a convolutional neural network, and to capture global features between genotype loci using multiple temporal mixing modules and channel mixing modules with a linear attention mechanism. The environmental features submodule is used to capture the multi-scale and temporal features of environmental factors by type using a linear attention mechanism, so as to obtain the illumination environment features and the surface environment features. The feature fusion submodule is used to fuse cross-modal features using the select and concat mechanisms to obtain genotype-light environment features, light environment-land surface environment features, and land surface environment-genotype features. The fitting and prediction submodule is used to fit and regress the fused features using a fully connected neural network to obtain the predicted yield value of the material.

[0019] A computer memory storing a computer program executable by a computer processor, the computer program performing a method for predicting yields based on RWKV multi-scale G×E interaction.

[0020] The beneficial effects of this invention are as follows: 1. The present invention relates to a method and system for predicting yield based on multi-scale G×E interactions using RWKV. First, a convolutional neural network is used to extract local features between genotype loci, and the linear attention mechanism RWKV is used to capture global dependencies between loci. Then, the temporal and channel mixing modules of RWKV are used to capture features between each environmental factor and multi-scale features throughout the entire reproductive period. Finally, cross-modal features are fused and linearly fitted to obtain the predicted yield value. By constructing an RWKV_GP model that integrates genotype feature vectors, illumination environment feature vectors, and land surface environment feature vectors, a multi-dimensional feature space is constructed. This multi-level feature fusion mechanism significantly improves the model's ability to represent complex patterns of genotype-environment interactions, analyzes the interaction effects between different environmental factors and genotypes, and improves the accuracy of genome prediction tasks.

[0021] 2. This invention demonstrates strong robustness in different experimental settings. Based on two maize datasets, the accuracy of RWKV_GP in predicting important agronomic traits related to yield was verified using two types of experiments (untested genotypes in an untested environment, and tested genotypes in an untested environment). The results of the three types of experiments show that the prediction accuracy of RWKV_GP is superior to six state-of-the-art statistical learning methods and four machine learning methods.

[0022] 3. This invention enables accurate prediction of yield values ​​for complex traits based on genomic and environmental data, thereby accelerating the efficiency of breeding and selection for maize crops.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of an embodiment of the present invention.

[0026] Figure 2 This is a flowchart of an embodiment of the present invention.

[0027] Figure 3 This is a PCA analysis diagram of the Maize976 population according to an embodiment of the present invention.

[0028] Figure 4 This is a performance evaluation chart of the ear weight phenotype prediction accuracy in cross-material experiments of the present invention in the Maize976 hybrid population, and the yield value prediction accuracy in the ZHENG58 hybrid progeny material and the MO17 hybrid progeny material.

[0029] Figure 5 This is a performance evaluation graph of the ear weight phenotype prediction accuracy in cross-material experiments of the present invention in the Maize507 inbred population, and the yield value prediction accuracy in inbred line population 1 and inbred line population 2.

[0030] Figure 6 This is a performance evaluation of the ear weight phenotype prediction accuracy in the Maize976 hybrid population and the yield value prediction accuracy in the ZHENG58 and MO17 hybrid progeny materials, based on cross-environmental experiments according to embodiments of the present invention.

[0031] Figure 7 This is a performance evaluation graph of the ear weight phenotype prediction accuracy in the Maize507 inbred population and the yield value prediction accuracy in inbred line population 1 and inbred line population 2 in cross-environmental experiments according to embodiments of the present invention.

[0032] Figure 8This is a performance evaluation chart of the Maize507 inbred line for predicting hybrid offspring with or without the participation of the paternal parent, according to an embodiment of the present invention.

[0033] Figure 9 This is a performance evaluation diagram of an embodiment of the present invention under different environmental factor treatments. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Example 1 See Figure 1 The specific steps of the RWKV-based multi-scale G×E interaction method for predicting yield are as follows: S1: Obtain whole-genome data from sequencing, environmental data throughout the maize growth cycle, and yield-related phenotypic data obtained from the field, and preprocess the data; the specific steps are as follows: Whole-genome data of breeding materials were obtained through sequencing. High-quality variant site data were obtained by comparing the whole-genome data with breeding microarray data, resulting in genotype data of length [length missing]. l SNP site sequence data G= { g 1, g 2 …g l}, g i ∈{0, 1, 2}; For environmental data, we collect multi-scale environmental factors related to light, such as light and temperature, and multi-scale environmental factors related to the land surface, such as wind speed, from planting time to harvest time at crop planting sites. For yield-related agronomic traits, collect yield-related data such as ear weight and ear length for each planting environment.

[0036] S2: In terms of genotype feature extraction, the RWKV_GP model first uses a convolutional neural network to extract local features between genotype loci, and then uses the RWKV mechanism to capture global features of genotype loci. G* The specific steps are as follows: S21: Sequence of SNP sites G Input the data into the CNN layer and perform feature compression through convolution operations to enhance the model's feature extraction capabilities. .

[0037] S22: Output of the CNN layer Input to the Batch Normalization (BN) layer and obtain the output. The BatchNormalization layer helps normalize the input to intermediate layers of the network, accelerating the training process and improving model stability. It helps reduce the vanishing gradient problem and ensures that the model converges to a good solution more quickly during training.

[0038] S23: After passing through multiple (e.g., two) convolutional layers and BN layers, the output is... It will enter the max pooling layer and the Dropout layer to get Pooling layers help reduce data dimensionality, computational complexity, and increase the model's receptive field. RWKV_GP uses two max-pooling layers, which reduce the size of the feature maps by selecting the maximum value in each pooling window to better capture features. Dropout layers are used to randomly drop some neurons during training to reduce the risk of overfitting. Dropout layers also help improve the model's generalization ability, allowing it to perform better on unseen data.

[0039] Steps S21, S22, and S23 include: SNP site sequence G Inputting the data into a convolutional CNN layer, and performing feature compression on the genotype data through convolution operations to enhance the model's feature extraction capability, yields... The specific steps are as follows: set up W c, i, j These are the convolution kernel weights, representing the weights of the convolution kernels. c The first channel i On the input channel, the first j One parameter.

[0040]

[0041] Subsequently, the genotype feature vector extracted from the convolutional layer is processed in the BN layer. Standardization is performed. Batch Normalization (BN) layers help normalize the inputs to intermediate layers of the network, reduce the vanishing gradient problem, and ensure that the model converges to a good solution faster during training. The specific operation of BN layers is shown below.

[0042]

[0043] in and They are the first c Mean and variance within each channel (calculated based on mini-batch samples).

[0044]

[0045] Finally, the genotype feature vectors extracted from the convolutional and batch normalization (BN) layers are input into the max pooling and dropout layers to obtain the feature vectors. The max pooling layer reduces the size of the feature map by selecting the most salient features within the pooling window, thereby reducing data dimensionality, computational complexity, and increasing the model's receptive field. The dropout layer is used to randomly discard some neurons during training to reduce the risk of overfitting. The implementation of the max pooling layer is shown below.

[0046]

[0047] After passing through four CNN and BN layers, and two max pooling layers, the input genotype data finally yields the output genotype. .

[0048] S24: The result obtained after the above convolution and pooling Inputting into the RWKV layer yields genotype characteristics. The RWKV module extracts genotype data features from a global perspective, thereby capturing the feature dependencies between distant SNP loci; the specific steps are as follows: RWKV_GP utilizes the genotype RWKV module for deep encoding of genotype features. This module employs a global feature extraction strategy to effectively capture long-range dependencies between SNP sites. The core of the RWKV module consists of a series of stacked residual blocks, each composed of time-mixing and channel-mixing modules with a cyclic structure. This provides a linear-cost attention-like mechanism and parallel training capabilities.

[0049] Temporal fusion blocks are a core component of RWKV, used to capture dependencies between different time steps in a sequence, effectively focusing on long-distance dependencies between genotype loci. Simultaneously, temporal fusion blocks borrow the multi-head attention mechanism and flexible global modeling capabilities of Transformers, and introduce the idea of ​​temporal recursion from RNNs, reducing computational complexity and enhancing the modeling ability for genotype data of equal length. Finally, by introducing a parameterized temporal weighting mechanism, the influence of specific time points in the input sequence is dynamically adjusted, and combined with the attention mechanism, efficient genotype data modeling is achieved.

[0050] The TimeShift operation is the core of the temporal fusion block. It refers to fusing the current feature vector with the previous feature vector to create a new vector, which can be used to enhance the model's sensitivity to local temporal context. By shifting the first half of the input genotype sequence's features one time step in the temporal dimension, the model can capture the dependencies between adjacent time steps, thereby improving its ability to model global genotype data features. By enriching the contextual information of the time series, it ensures that the information of the current token can recursively learn from past information.

[0051] In the TimeShift operation, let the input sequence be... , where: feature dimension C It is divided into two parts according to its location: and .

[0052] Features of the first half In the time dimension t Shifting upwards by one time step is defined as:

[0053] Features of the second half Defined as:

[0054] The offset first half and the unoffset second half are concatenated along the feature dimension to generate a new feature. .

[0055]

[0056] Temporal blending blocks replace the weighted sum computation in traditional attention mechanisms with hidden states, enhancing the modeling of dependencies within genotype sequences. By incorporating hidden states, temporal blending blocks can propagate and update information across sequences. Simultaneously, it improves upon Attention-Free Transformer (AFT), transforming the self-attention mechanism in the Transformer into an implementation utilizing RNN principles, achieving both linear speed and good performance. The temporal blending block first requires a TimeShift operation on the input genotype token matrix, as shown below. u r , u k and u v These are learnable mixing coefficients.

[0057]

[0058]

[0059]

[0060] The input to the attention mechanism in the temporal hybrid block is the weighted sum of the current token and the previous token, and its output acts as cross-attention in the standard transformer. The specific formula is shown below. wkv t This represents the current recursive state. The formula can be efficiently rewritten as a recursive expression with linear computational cost, enabling the model to handle infinite context lengths and effectively capture long-term dependencies in gene loci and environmental factors.

[0061]

[0062] Channel mixing blocks are designed to perform non-linear interactions and updates on feature channels at each time step, providing a refined processing of the feature representation at each time step. Unlike the Time-Mixing module, which deals with the dependencies of time series, the Channel-Mixing module focuses on the interactions within feature channels. By introducing gating mechanisms and non-linear transformations, it enhances the model's ability to capture complex features in genotype data.

[0063] Traditional self-attention mechanisms rely on linearly growing key-value caches, resulting in increased memory usage and time consumption as the sequence length increases. Genotype data typically has long sequence lengths, leading to low prediction efficiency for traditional attention mechanisms in genome prediction. To address these issues, the channel mixing block employs an RNN-like inference and decoding approach: in RNNs, states are typically... t The output is used as the state. t+1 The input is used in this way, allowing the RWKV module to decode recursively during inference. This mode leverages the advantage that each output tag depends only on the latest state, the size of which is constant and independent of the input sequence length. Therefore, RWKV has constant speed and memory footprint, making it more efficient at handling longer genotype sequences. The Channel-Mixing mode described above is shown below. u r , u z For learnable mixing coefficients, output O t Includes time t Historical information.

[0064]

[0065]

[0066] The gating mechanism generates a gating factor through a sigmoid activation function, which is used to adjust the weights at different time steps to obtain the final genotype feature output. The process is shown below. A sigmoid activation function is used to determine which genotype features should be retained, and the available information is filtered using the ⊙ operation to serve as the vector representation of the next token.

[0067]

[0068] The output after passing through the CNN and max pooling module The input is processed through multiple stacked Time-Mixing and Channel-Mixing modules to finally generate genotype feature vectors. G* .

[0069] S3: Regarding multi-scale environmental feature extraction, the environmental features are first categorized into two types: multi-scale illumination environmental features and multi-scale surface environmental features. Then, multiple temporal and channel mixing modules of the RWKV mechanism are used to capture the multi-scale and temporal features of each type of environmental factor, resulting in illumination environmental features and surface environmental features. Ex* and Ey* The specific steps are as follows: Given the multi-scale nature of environmental data, RWKV_GP can directly utilize the linear attention mechanism of the RWKV module to capture environmental features throughout the entire growth period of multi-scale illumination and surface data. The RWKV module extracts key feature information from the illumination and surface data and transforms it into final multi-scale illumination and surface data feature vectors.

[0070] The input to the environment RWKV module is a multi-scale illumination environment vector and a multi-scale surface environment vector. Among them, Ex (1) ∈ R d×e for d Within a day e Feature vectors of illumination-related environmental factors Ey (1) ∈ R d×e for d Within a day e Feature vectors of surface-related environmental factors. Based on multi-scale illumination environment data. Ex (1) For example, see the details below.

[0071]

[0072]

[0073]

[0074]

[0075] The RWKV module fully utilizes the time-series information in environmental data, using feature data from the past period for the calculation of current features. For illumination data, it includes time-series information on multiple meteorological variables such as temperature and sunlight intensity. For surface data, it considers the time-series evolution of factors such as soil moisture and wind speed. By incorporating this time-series data into the RWKV feature extraction process, the evolutionary trends of illumination and surface variables can be captured more comprehensively, ultimately yielding illumination environmental features. Ex* and surface environmental characteristics Ey* Output vector.

[0076] S4: Regarding the interaction between genotype and multi-scale environmental features, drawing on the idea of ​​cross-modal feature fusion, the RWKV_GP model uses the select and concat mechanisms to fuse genotype-light environment data. G-Ex Light environment - Surface environment Ex-Ey Surface environment - genotype G-Ey Three sets of feature vectors; the specific steps are as follows: Based on the idea of ​​fusing cross-modal features, RWKV_GP integrates genotype and environmental features. G* Lighting environment characteristics Ex* and surface environmental characteristics Ey* Feature fusion is performed on three types of features. This process aims to extract and fuse information between feature vectors through linear mapping, concatenation, and further combined feature transformations, thereby enhancing the model's final feature representation capability. Simultaneously, to ensure that the contributions of the three different types of data to the final result are consistent, this embodiment uniformly reduces the dimensionality of the three feature vectors to maintain the same feature length. This operation balances the influence of different data sources on the phenotype, ensuring that multi-dimensional data have similar weights in the model.

[0077] Let Concat represent the splicing operation, where feature fusion yields the genotype-light environment profile. GE x Light environment - Surface environment E x E y and surface environment - genotype GE y The three sets of feature vectors are shown in the following process:

[0078]

[0079] S5: Use a fully connected neural network to fit and regress the fused features to obtain the predicted yield value of the material. y The specific steps are as follows: After feature fusion, the three sets of feature vectors are input into a linear layer and linearly mapped to obtain vectors of uniform dimension. Finally, the three sets of dimensionality-reduced feature vectors are concatenated and linearly transformed to obtain the predicted output value of the model. The specific operation is as follows:

[0080]

[0081]

[0082] By integrating genotype feature vectors, illumination environment feature vectors, and land surface environment feature vectors, RWKV_GP constructs a multi-dimensional feature space. This multi-level feature fusion mechanism significantly enhances the model's ability to represent complex patterns of genotype-environment interactions, elucidates the interaction effects between different environmental factors and genotypes, and significantly improves the accuracy of RWKV_GP in genome prediction tasks.

[0083] This embodiment first uses a convolutional neural network to extract local features between genotype loci and employs a linear attention mechanism to capture global dependencies between loci. Then, it uses the temporal and channel mixing modules of RWKV to capture features between each environmental factor and multi-scale features throughout the entire reproductive period. Finally, it fuses cross-modal features and performs linear fitting to obtain the predicted yield value. By constructing the RWKV_GP model and integrating genotype feature vectors, illumination environment feature vectors, and land surface environment feature vectors, a multi-dimensional feature space is built. This multi-level feature fusion mechanism significantly improves the model's ability to represent complex patterns of genotype-environment interactions, analyzes the interaction effects between different environmental factors and genotypes, and achieves the function of improving the accuracy of genome prediction tasks.

[0084] Example 2 The steps in this embodiment are the same as in Embodiment 1, except that each step is applied to a specific instance. An experiment on genome prediction is conducted in a population of publicly published maize data (Maize507 inbred line, Maize976 hybrid). The relevant experiments are used to illustrate in detail the superior performance of the method of the present invention in the interaction between genotype and environment.

[0085] The results of the PCA analysis of the Maize976 population are as follows: Figure 3The red subpopulation represents the MO17 subpopulation, and the green subpopulation represents the ZHENG58 subpopulation. The first two principal components of this PCA collectively explain 92.06% of the genetic variation, with PC1 contributing 89.34% and PC2 contributing 2.72%.

[0086] This study compared the predictive performance of RWKV_GP with four statistical learning methods: the main-effect EAD+GE biased model based on optimal linear unbiased genomic prediction (EAD+GE(GB)), the main-effect EAW+GE response canonical model based on optimal linear unbiased genomic prediction (EADW+GW(GB)), the main-effect EAD+GE biased model based on Gaussian kernel (EAD+GE(GK)), and the main-effect EAW+GE response canonical model based on Gaussian kernel (EADW+GW(GK)). The predictive accuracy of the ear weight phenotype in cross-material experiments was compared between two populations (Maize507 inbred line and Maize976 hybrid) and their subpopulations with different parental backgrounds (MO17 and ZHENG58 parents). Figure 4 and Figure 5 ).

[0087] Experimental results based on all materials from the Maize976 hybrid population and the Maize507 self-pollinated population ( Figure 4 A and Figure 5 In study A), RWKV_GP demonstrated a significant advantage in prediction accuracy across materials. For the Maize976 hybrid population, RWKV_GP achieved an average prediction accuracy 22.4% higher than the control method. In the Maize507 self-pollinated population, RWKV_GP also significantly outperformed the control method, with an average improvement in prediction accuracy of 27.8%.

[0088] The Maize976 population was obtained by crossing 488 different inbred lines with two superior parents (ZHENG58 and MO17). Since this population contains multiple hybrids, each hybrid can be further divided into two different materials based on the ZHENG58 and MO17 parents. This parental difference can cause characteristic differences between different materials during model training, significantly affecting the model's predictive performance. To investigate whether RWKV_GP has better predictive ability for materials with similar genetic backgrounds, this study, while avoiding interference between parents with different genetic backgrounds, specifically focused on the predictive performance of materials with similar genetic backgrounds.

[0089] Figure 4(B and C) primarily showcase the cross-material EarWeight yield prediction results of RWKV_GP for the ZHENG58 and MO17 progeny materials in the Maize976 population. Comparisons revealed that RWKV_GP demonstrated a significant advantage in prediction accuracy in single-environment cross-material experiments within the MO17 and ZHENG58 subpopulations of the Maize976 population. For the MO17-related parental materials, RWKV_GP improved the accuracy by at least 22.1% compared to the comparative method. For the ZHENG58-related parental materials, RWKV_GP improved the accuracy by at least 33.0% compared to the comparative method.

[0090] In the Maize976 population, the maternal parents of the hybrids all originated from the Maize507 inbred line population. By selecting maternal parents from the Maize507 population who were crossed with the two parents, MO17 and ZHENG58, a dataset of inbred lines with different progeny backgrounds was obtained. Modeling the characteristic relationships of these materials and analyzing their association with parental characteristics allows for in-depth learning of the genetic characteristics of the inbred line population, while also analyzing the deep-seated relationships between parents and hybrid offspring.

[0091] Two subpopulations were constructed from the Maize507 population: inbred line population 1, constructed from the MO17 parent and its N parent materials, and inbred line population 2, constructed from the ZHENG58 parent and its M parent materials. Figure 5 (B and C) primarily present the experimental results of cross-material EarWeight yield prediction for inbred line populations 1 and 2 in RWKV_GP. Comparisons revealed that RWKV_GP demonstrated a significant advantage in prediction accuracy in single-environment cross-material experiments for inbred line populations 1 and 2 of the Maize507 population. For the MO17-related parental materials, RWKV_GP improved the accuracy by at least 24.0% compared to the comparative method. For the ZHENG58-related parental materials, RWKV_GP improved the accuracy by at least 44.8% compared to the comparative method.

[0092] In summary, to evaluate the predictive performance of RWKV_GP on materials with similar genetic backgrounds, this study selected subsets of the Maize976 and Maize507 datasets with different genetic backgrounds using the ZHENG58 and MO17 parents, respectively. This method not only helps improve the model's predictive accuracy within the same genetic background but also allows for a deeper examination of the model's adaptability to phenotypic changes. Through this segmentation, we can further verify whether RWKV_GP can effectively avoid feature interference under similar genetic backgrounds, thereby improving the overall stability and accuracy of predictions.

[0093] To evaluate the predictive performance of RWKV_GP in experiments with test genotypes in untested environments, this study used data from two populations (Maize507 inbred line and Maize976 hybrid) and their offspring populations (ZHENG58 and MO17) with different parental backgrounds. The accuracy of RWKV_GP in predicting ear weight phenotypes across environments was compared with that of three machine learning methods (Random Forest (RF), Gradient Boosting Decision Tree (XGBoost), and Stacking ensemble model), one deep learning method (Multilayer Perceptron (DL)), and four statistical learning methods. Figure 6 and 7 ).

[0094] In most environments, RWKV_GP demonstrated relatively stable prediction accuracy, while exhibiting the best prediction accuracy in some environments. For example, for the Maize976 population in the 2012 CQ environment, RWKV_GP achieved the highest prediction accuracy among all methods. For the Maize507 population in the YN(2011) environment (a challenging prediction condition), RWKV_GP significantly outperformed other comparative methods, improving the prediction performance by at least 2.87%. These results indicate that RWKV_GP has strong predictive ability across environments for hybrids, accurately capturing correlations between different environments and demonstrating high prediction accuracy and stability.

[0095] Figure 6 (B and C) mainly present the cross-environment EarWeight prediction results of RWKV_GP for the ZHENG58 and MO17 hybrid progeny materials in the Maize976 population. For the MO17 subpopulation in the HN (2011), CQ (2012), and HB (2012) environments, RWKV_GP significantly outperformed other comparative methods, improving by at least 3.49%, 4.25%, and 0.45% respectively. Similarly, for the ZHENG58 subpopulation in the CQ (2012), HB (2012), and YN (2012) environments, RWKV_GP also significantly outperformed other comparative methods, improving by at least 2.37%, 0.95%, and 1.02% respectively.

[0096] Two subpopulations were constructed from the Maize507 population: inbred line population 1, constructed from the MO17 parent and its N parent materials, and inbred line population 2, constructed from the ZHENG58 parent and its M parent materials. Figure 7(B and C) mainly present the cross-environment EarWeight prediction results of inbred line populations 1 and 2 in RWKV_GP. For inbred line population 1 in the challenging environments of HN(2011) and YN(2011), RWKV_GP significantly outperformed other comparative methods, improving by at least 7.07% and 4.005% respectively. Similarly, for inbred line population 2 in the YN(2011) environment, RWKV_GP also significantly outperformed other comparative methods, improving by at least 3.20%.

[0097] To verify whether the characteristic association between parental materials and offspring materials helps improve the prediction accuracy of offspring materials, this study set up a control experiment based on the participation of two paternal parents, MO17 and ZHENG58, in cross-material prediction of the Maize507 population, obtaining two different training models and predicting the yield values ​​of hybrid offspring. Figure 8 The terms "Paternal And Maternal" represent predictions obtained using a training population built on the father (MO17 or ZHENG58) and mother, while "Only Maternal" represents predictions obtained using only the mother. This process not only improves the model's feature extraction capabilities in complex genetic contexts but also provides an effective means to further explore the feature relationships between inbred lines and hybrid offspring, further revealing the complex nonlinear relationship between hybrid offspring and parents.

[0098] Figure 8 (A) compares the accuracy of predicting hybrid offspring with and without the inclusion of the MO17 paternal parent. The results show that RWKV_GP exhibits a significant predictive advantage in predicting hybrid offspring using a dataset that includes the paternal parent. Specifically, compared to the case without the MO17 paternal parent, the average accuracy of predicting hybrid offspring with the MO17 paternal parent is improved by 3.5277 times. This phenomenon indicates that including the paternal parent in the dataset of MO17-related parental materials can effectively capture the deep relationship between parents and offspring, thereby significantly improving the accuracy of hybrid materials.

[0099] RWKV_GP explores the potential advantages of genotype-environment data interactions in predicting complex traits; therefore, it is crucial to investigate whether environmental factors are included in the model and how to effectively handle them. To elucidate the role of environmental covariates in RWKV_GP, this study established a two-pronged control experimental framework: (1) completely removing environmental feature vectors using parametric ablation; and (2) replacing all environmental factors using zero-initialization masking. This control framework can quantitatively assess the contribution of environmental elements to model predictions. This example uses the Maize976 population's cross-material prediction experiment as an example to compare and analyze the different settings of the above two methods for handling environmental factors.

[0100] Figure 9 This paper presents the cross-material prediction results of RWKV_GP on all materials in the Maize976 population, using two different treatments of environmental factors. RWKV_NOENV represents the treatment method that completely removes environmental data, and RWKV_ENV_ZERO represents the treatment method that replaces environmental data with zeros. The results show that RWKV_GP improves prediction performance in most environments. Specifically, RWKV_GP achieves an average prediction accuracy improvement of 2.58% and 0.27% compared to RWKV_NOENV and RWKV_ENV_ZERO, respectively. This result indicates that RWKV_GP's treatment of environmental factors can capture the interaction between genotype and environmental factors at a deeper level, thereby improving the predictive ability for complex traits.

[0101] Studies on maize yield-related traits involve several key agronomic traits, among which ear weight, as a core indicator, shows significant correlations with secondary yield traits such as ear length, ear diameter, ear row number, and cob weight. These agronomic traits are not only closely related to maize yield, but the interactions between their genetic variation sites also exhibit a complex genetic structure. Therefore, compared to non-yield-related traits such as plant height, models predicting yield-related traits often struggle to achieve high accuracy. To further explore the ability of RWKV_GP to capture both complex and simpler characteristics of yield-related traits, this study conducted cross-material and cross-environment genome prediction experiments for traits such as ear length (Table 1).

[0102] Table 1. Genomic prediction experiments across materials and environments.

[0103] Experimental results showed that the prediction accuracy for spike length and grain weight traits was relatively low under most environmental conditions, with prediction accuracy generally below 0.5 (mainly around 0.3). In contrast, the prediction results for spike diameter and number of spike leaves showed a higher correlation, with prediction accuracy mostly between 0.5 and 0.6. These results indicate that SNP loci affecting plant height are more easily captured by prediction models than yield-related traits.

[0104] This embodiment demonstrates strong robustness across different experimental settings. Based on two maize datasets, two types of experiments (untested genotypes in an untested environment, and tested genotypes in an untested environment) were used to validate the accuracy of RWKV_GP in predicting important agronomical traits related to yield. The results of the three types of experiments show that RWKV_GP outperforms six state-of-the-art statistical learning methods and four machine learning methods in prediction accuracy.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0106] Example 3 This embodiment is used to implement the principle of the above method embodiment to build a multi-scale G×E interaction prediction system based on RWKV, including a data acquisition submodule, a gene feature submodule, an environmental feature submodule, a feature fusion submodule, and a fitting prediction submodule.

[0107] The data acquisition submodule is used to acquire whole-genome data, multi-scale environmental data, and yield phenotypic data, and to perform preprocessing. The gene feature submodule is used to extract local features between genotype loci using a convolutional neural network, and to capture global features between genotype loci using multiple temporal mixing modules and channel mixing modules with a linear attention mechanism. The environmental features submodule is used to capture the multi-scale and temporal features of environmental factors by type using a linear attention mechanism, so as to obtain the illumination environment features and the surface environment features. The feature fusion submodule is used to fuse cross-modal features using the select and concat mechanisms to obtain genotype-light environment features, light environment-land surface environment features, and land surface environment-genotype features. The fitting and prediction submodule is used to fit and regress the fused features using a fully connected neural network to obtain the predicted yield value of the material.

[0108] Each submodule is mainly used to implement the various steps of the method embodiment, which will not be elaborated here.

[0109] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0110] This embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of a method for predicting yields based on multi-scale G×E interaction of RWKV.

[0111] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, enable the processor to implement a method for predicting yields based on RWKV multi-scale G×E interaction.

[0112] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0113] Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to the flowchart of the method and computer program product according to Embodiment 1 and the block diagram of the device (system) according to Embodiment 3. It should be understood that each step or block in the flowchart or block diagram, as well as combinations of steps or blocks in the flowchart or block diagram, can be implemented by computer program instructions.

[0115] These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which are executable by the processor of the computer or other programmable data processing device, produce instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A system for predicting yields based on RWKV multi-scale G×E interaction, specifying the functions in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes or boxes Figure 1 The steps of the method for predicting yields based on RWKV multi-scale G×E interaction are specified in one or more boxes.

[0118] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting yield based on RWKV multi-scale G x E interaction, characterized by: The method comprises the following steps: S1: obtaining whole genome data, multi-scale environment data and yield phenotype data, and preprocessing; S2: using a convolutional neural network to extract local features between genotypic sites, and using multiple time mixing modules and channel mixing modules of a linear attention mechanism to capture global features between genotypic sites; S3: using a linear attention mechanism to capture multi-scale features and time sequence features of environmental factors by type, to obtain illumination environment features and surface environment features; S4: using a select mechanism and a concat mechanism to fuse cross-modal features to obtain genotypic-illumination environment features, illumination environment-surface environment features and surface environment-genotypic features; S5: using a fully connected neural network to fit and regress the fused features to obtain the predicted yield value of the material.

2. The method of claim 1, wherein the RWKV-based multiscale G x E interaction yield prediction is based on: In the step S1, the specific steps are: The whole genome data of the breeding material is obtained by sequencing, the whole genome data is compared with the breeding chip data to obtain SNP site sequence genotype data with a length of l G= { g 1, g 2 …g l}, g i ∈{0,1,2};​ Collecting multi-scale environmental factors related to illumination and multi-scale environmental factor data from planting time to harvesting time at crop planting sites; Collecting yield-related agronomic trait phenotype data for each planting environment.

3. The method of claim 1, wherein: In the step S2, the specific steps are: S21: sequence of SNP sites G Input CNN layer, compress features by convolution operation ; S22: normalizing the output of the CNN layer inputting the BN layer to obtain an output ; S23: output after passing through multiple CNN layers and BN layers input max-pooling layer and Dropout layer, obtain by reducing data dimension and randomly discarding part of neurons ; S24: the data features of the genotype are extracted from the global dimension by the linear attention mechanism of the input RWKV layer, and then the feature dependence between the long distances between the SNP sites is captured to obtain the genotype features input RWKV layer, through linear attention mechanism from global dimension to extract the data features of the genotype, and then capture the feature dependence between the long distances between the SNP sites to obtain the genotype features .

4. The method of claim 3, wherein the RWKV-based multiscale G x E interaction yield prediction is characterized by: In the steps S21, S22 and S23, the specific steps are: Let W c，i，j be the convolution kernel weights of the CNN layer, representing the c th input channel of the i th input channel of the j th parameter, then is given by: ； Let and are the mean and variance of the BN layer respectively c in the i-th channel ， The BN layer extracts a genotype feature vector for the CNN layer After standardization, the following is obtained: ； The genotype feature vector extracted through multiple CNN layers and BN layers An input max-pooling layer reduces the dimension of data by selecting the most significant features in the pooling window; a Dropout layer randomly discards some neurons during the training process to reduce the risk of overfitting; Obtained: ； After several max-pooling layers and Dropout layers, the output genotype is obtained .

5. The method of claim 3, wherein: In the step S24, the specific steps are: The current feature vector is fused with the previous feature vector as a current new vector through a TimeShift operation; the first half of the feature of the input genotype sequence is shifted in the time dimension by one time step, so as to capture the dependence between adjacent time steps; the input sequence , the feature dimension C is divided into and according to the position; First half features In time dimension t One time step up translation is: second half features defined as: ， The offset first half and the non-offset second half are spliced in the feature dimension to generate new features : ； Let u r , u k and u v be learnable mixing coefficients, TimeShift the input genotype token matrix: ， ， ； The input to the attention mechanism in the temporal mixing block is a weighted sum of the current token and the previous token, and its output plays the role of cross-attention in the standard transformer, the current recurrent state wkv t is: ； The channel mixing block adopts RNN-like push decoding manner, taking the output of state t as the input of state t+1 , to decode in a recursive manner; Let u r ，u z for the learnable mixing coefficients, the output O t contains history information to time t The channel mixing mode is specified as follows: ， ， The gating mechanism generates gating factors by a sigmoid activation function, which is used to adjust the weights of different time steps to obtain the final genotype feature output, wherein The genotype feature information to be retained is determined by a sigmoid activation function, and the available information is filtered by a operation to obtain a vector representation of the next token: ； The output after passing through the CNN layer and the max pooling layer The input is processed through multiple stacked time mixing and channel mixing modules to generate genotype feature vectors. G* .

6. The method of claim 1, wherein: In the step S3, the specific steps are: Input d Within a day e a feature vector of the individual light-related environmental factor Ex (1) ∈ R d×e , capturing the environmental characteristics of multi-scale light data and multi-scale ground data throughout the growing season using linear attention mechanism: Obtaining a multi-scale lighting environment feature output vector Ex* ; Input d Day e Feature vector of the individual surface-related environmental factor Ey (1) ∈ R d×e , capture the environmental characteristics of the multi-scale surface data throughout the growing season using linear attention mechanism: get multi-scale surface environmental feature output vector Ey* .

7. The method of claim 1, wherein: In the step S4, the specific steps are: Let Concat denote the concatenation operation, and let the fused features be the genotype-illumination-environment feature vector GE x , the illumination-environment-soil feature vector E x E y , and the soil-environment-genotype feature vector GE y is: ， 。 8. The method of claim 1, wherein: In the step S5, the specific steps are: Uniformly input the fused feature vector into a linear layer, and obtain a uniform dimension vector through linear mapping; splice the uniformly reduced feature vectors, and map them into the predicted yield value of the model output through linear transformation: ， ， ； A multi-dimensional feature space is constructed by integrating the genotypic feature vector, the illumination environment feature vector and the surface environment feature vector.

9. A system for predicting yield based on multi-scale GxE interaction according to the RWKV, characterized in that: a data acquisition submodule is configured to acquire whole genome data, multi-scale environment data and yield phenotype data, and perform preprocessing; a gene feature submodule is configured to use a convolutional neural network to extract local features between genotypic sites, and use multiple time mixing modules and channel mixing modules of a linear attention mechanism to capture global features between genotypic sites; an environment feature submodule is configured to use a linear attention mechanism to capture multi-scale features and time sequence features of environmental factors by type, to obtain illumination environment features and surface environment features; a feature fusion submodule is configured to use a select mechanism and a concat mechanism to fuse cross-modal features to obtain genotypic-illumination environment features, illumination environment-surface environment features and surface environment-genotypic features; a fitting and prediction submodule is configured to use a fully connected neural network to fit and regress the fused features to obtain the predicted yield value of the material.

10. A computer memory, characterized by: It has a computer program stored therein which can be executed by a computer processor, and the computer program executes the method for predicting yield based on multi-scale GxE interaction according to the RWKV as claimed in any one of claims 1 to 8.

Images