Method for modeling forest genotype-environment interaction based on multi-modal deep learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- RES INST OF FOREST RESOURCE INFORMATION TECHN CHINESE ACADEMY OF FORESTRY
- Filing Date
- 2025-10-29
- Publication Date
- 2026-07-21
Smart Images
Figure CN121483367B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of forest tree breeding technology, specifically to a method for modeling forest tree genotype-environment interactions based on multimodal deep learning. Background Technology
[0002] Forest tree breeding technology is of great significance to ecological construction and timber resource supply. Its growth status directly affects ecological balance and the stable development of the timber industry. With the continuous advancement of science and technology, intelligent breeding technology has gradually emerged. The combination of artificial intelligence and high-throughput technology has provided new possibilities for analyzing genotype-environment interactions. Forest tree phenotypic traits, such as tree height, diameter at breast height, crown width, and biomass, are the result of complex interactions between genotype and environment. Accurately analyzing this interaction mechanism and achieving precise phenotypic prediction are crucial for the selection of superior forest tree varieties and efficient cultivation management, and are key to promoting the sustainable development of the forestry industry.
[0003] Traditional forest tree breeding methods rely primarily on long-term artificial selection, which is time-consuming, inefficient, and fails to meet the current demands for diverse timber resources and rapid environmental adaptation. Furthermore, traditional manual phenotyping methods are extremely inefficient and subjective in data collection, making them unsuitable for large-scale breeding. Current forest tree breeding faces key bottlenecks such as a scarcity of superior germplasm resources, limited genetic diversity of new varieties, and unclear genotype-environment interaction mechanisms, resulting in severely insufficient phenotypic prediction accuracy. In addition, existing machine learning models, when applied to forest tree breeding, face challenges due to the long growth cycle of trees, making data collection difficult, model validation and adjustment complex, and algorithms need to adapt to the long growth cycle and related environmental and genetic changes. Moreover, most machine learning models are black-box characteristics with poor interpretability, failing to provide clear biological evidence for breeding decisions and severely hindering the sustainable development of the forestry industry. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a forest tree genotype-environment interaction modeling method based on multimodal deep learning. This method systematically collects multimodal data on forest tree genotypes, environment, and phenotypes. After preprocessing and fusion, it constructs a genotype-environment interaction algorithm and a phenotype prediction model using innovative algorithmic formulas. Specifically, it uses a high-throughput phenotype platform, whole-genome sequencing technology, and soil nutrient detection equipment to collect multimodal data and performs targeted preprocessing on various data types. The multimodal data is then fused, and a deep neural network is constructed based on the fused data to extract the nonlinear interaction features between genotypes and environmental factors, developing a genotype-environment interaction algorithm. Based on this interaction algorithm, various strategies are introduced to construct and optimize the phenotype prediction model, solving the problems of long breeding cycles, low efficiency, and insufficient phenotype prediction accuracy in traditional forest tree breeding. This significantly improves the accuracy of forest tree phenotype prediction and breeding efficiency, providing key technical support for precision forest tree breeding and efficient cultivation management.
[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: a forest tree genotype-environment interaction modeling method based on multimodal deep learning, which includes the following specific steps:
[0006] Multimodal data acquisition: High-throughput phenotyping platform, whole-genome sequencing technology and soil nutrient detection equipment were used to collect SNP locus genotype data of forest trees, key soil physicochemical environmental data and growth-related morphological and biomass parameter phenotypic data.
[0007] Multimodal data preprocessing and fusion: After calibrating and optimizing phenotypic data, reducing the dimensionality of environmental data and removing collinearity, and standardizing genotype data for quality control, the three types of data are fused to obtain the fused multimodal feature matrix;
[0008] Genotype-environment interaction algorithm development: Based on fused data, a deep neural network is constructed to extract the nonlinear interaction features between genotype and environment. A multi-level model is constructed by combining a hybrid linear architecture, and the interaction algorithm is developed after optimizing the complexity.
[0009] Phenotypic prediction model construction and optimization: Based on the interaction algorithm, feature contrastive learning and knowledge transfer are introduced, multi-task learning is used to optimize the model, and probability optimization and feature contribution quantification are combined to construct a phenotypic prediction model with biological interpretability.
[0010] Furthermore, the genotype data is obtained by whole-genome resequencing of DNA from tree leaf samples to obtain SNP locus information; the environmental data consists of key physicochemical indicators of soil in the tree growth plots, collected in multiple plots with multiple biological replicates for each plot, and the average of the measurement results is taken; the phenotypic data consists of morphological and biomass parameters related to tree growth, which are obtained from three-dimensional data through a combination of aerial overhead equipment and ground mobile acquisition equipment, and the three-dimensional data is segmented, labeled and parameter extracted by individual trees using professional three-dimensional point cloud processing software to obtain the final phenotypic data.
[0011] Furthermore, in the multimodal data preprocessing and fusion step, the three types of data are fused, and the fusion formula is as follows: ,in, The fused multimodal feature matrix has dimensions of . , For the sample size, For the fused feature dimensions, , , The basic weight coefficients for genotype, environmental, and phenotypic characteristics are determined through cross-validation optimization. This is the preprocessed genotype feature matrix, obtained from SNP loci after quality control, standardization, and dimensionality reduction. Each row represents the genotype information of one forest tree sample, and each column represents a selected SNP locus or genotype principal component. This is the preprocessed environmental feature matrix, obtained from key soil physicochemical indicators after collinearity removal and principal component analysis. Each row represents comprehensive information about the growth environment of a single forest sample, and each column represents a comprehensive environmental factor. This is the preprocessed phenotypic feature matrix, where each row represents the phenotypic information of one forest tree sample, and each column represents one phenotypic trait. These are the attention mechanism weighting coefficients. It is a cross-modal attention feature.
[0012] Furthermore, in the genotype-environment interaction algorithm development steps, a deep neural network based on a multilayer perceptron architecture is constructed based on the fused multimodal data. This network receives the fused genotype, environment, and phenotype data, and designs a multi-dimensional feature interaction extraction layer. The interaction feature extraction of genotype and environmental factors in the nonlinear space is achieved through the genotype-environment nonlinear interaction feature extraction formula. A hierarchical feature fusion strategy is adopted, establishing genotype-specific and environment-response modules in the network hidden layers respectively. This is combined with a cross-modal feature association mechanism to capture the nonlinear relationship between genotype, environment, and phenotype, while using a hybrid linear architecture as the core. The basic framework uses phenotypic data as the core response variable, genotype main effects, environmental main effects, and their interaction as fixed effects, and plot spatial heterogeneity and population structure principal components as random effects to construct a multi-level model. The model complexity is optimized by using a loss function formula for the multi-level interaction model. A hybrid feature extraction-perception architecture is adopted, embedding association analysis significance parameters into the genotype module to construct an environmental feature extractor based on time-series feature learning, and coupling an environmental factor threshold response module to improve the algorithm's ability to analyze complex genotype-environment interactions. Finally, a high-performance genotype-environment interaction algorithm is developed.
[0013] Furthermore, in the genotype-environment interaction algorithm development steps, the interaction features of genotype and environmental factors in the nonlinear space are extracted using a genotype-environment nonlinear interaction feature extraction formula, which is as follows: ,in, It is a genotype-environment interaction feature matrix with dimension 1. , As the interaction feature dimension, It is an activation function used to introduce non-linear relationships. Linear transformation weight matrices for genotype characteristics and environmental characteristics, respectively, with dimensions of... , , For genotype characteristics, As an environmental characteristic dimension, It is a feature concatenation operation that concatenates the linearly transformed genotype features with environmental features in a dimensional manner, preserving the independent information of the two types of features. It is a bias vector used to adjust the baseline level of the feature output. It is a multilayer perceptron module, containing two layers of linear transformation and one layer of ReLU activation, used for nonlinear mapping of the genotype-environment feature similarity matrix. and It is a similarity matrix between genotype characteristics and environmental characteristics. It is the scaling factor.
[0014] Furthermore, in the genotype-environment interaction algorithm development steps, the model complexity is optimized through a multi-level interaction model loss function formula, which is as follows: ,in, It is the total loss value of the interaction algorithm, used to measure the difference between the model's predicted values and the true phenotypic values. It is the sample size. It is the first The true phenotypic values of each sample It is the first The phenotypic predicted value of each sample is derived from the interaction feature matrix. Obtained through linear mapping, It is the L2 regularization coefficient, used to suppress... , Overfitting , They are respectively , The L2 norm is the square root of the sum of the squares of the elements in the weight matrix. It is the interaction feature regularization coefficient.
[0015] Furthermore, in the phenotypic prediction model construction and optimization steps, based on the interaction algorithm, a feature contrast learning strategy is introduced to optimize the feature representation space. Combined with a knowledge transfer framework, a multi-task deep neural network is constructed using temporal feature learning. By sharing the genotype-environment interaction feature layer, specific prediction units are independently set for each phenotypic trait, and an adaptive weight allocation mechanism is designed. Through the adaptive weight formula for multi-task phenotypic prediction, the adaptive weight allocation mechanism is designed to dynamically balance the loss function contributions of different phenotypic tasks, optimize the model's comprehensive prediction performance for multiple phenotypic traits, and simultaneously introduce probabilistic optimization methods to adjust the model's hyperparameters. An improved key factor contribution quantification formula is used to quantify the contribution of key SNP sites and environmental driving factors to the target phenotypic trait, screen factors that significantly contribute to the phenotype, and combine gene function annotation information to mine the core regulatory factors affecting phenotypic formation. Combined with probabilistic methods, a hierarchical probability model is constructed using the posterior probability formula of the Bayesian hierarchical model. The posterior probability is used to estimate the strength of the genotype-environment interaction effect, improve the statistical power of small samples, and establish a forest tree phenotypic prediction model with clear biological interpretability.
[0016] Furthermore, in the phenotypic prediction model construction and optimization steps, an adaptive weight allocation mechanism is designed using a multi-task phenotypic prediction adaptive weight formula, the formula of which is: ,in, It is the first The prediction task weights for each phenotypic trait are used to dynamically balance the loss contributions of different phenotypic tasks. It is the first The predicted loss value for each phenotypic trait. It is the first The predicted loss value for each phenotypic trait. It is an exponential function used to amplify the impact of differences in loss between different tasks on the weights.
[0017] Furthermore, in the phenotypic prediction model construction and optimization steps, the contribution of key SNP sites and environmental driving factors to the target phenotypic trait is quantified by an improved key factor contribution quantification formula, the formula of which is: ,in, It is the first The SNP locus and the first The combined contribution of individual environmental factors It is the first In the nth sample, retain the nth The first SNP site and the second Phenotypic predictions for each environmental factor It is the first In the sample, the first one was removed. The first SNP site and the second Phenotypic predictions for each environmental factor It is the first The SNP locus and the first The correlation coefficients of the environmental factors are used to correct the overestimation of their contributions caused by collinearity.
[0018] Furthermore, in the phenotypic prediction model construction and optimization step, a hierarchical probability model is constructed using the posterior probability formula of the Bayesian hierarchical model, the formula of which is: ,in, Model parameters The posterior probability distribution is reflected in the observed data. Below, parameters The credibility of It is a set of model parameters. It is an observation dataset. It is the likelihood function, representing the likelihood of the parameter The observed data The probability of is calculated using the following formula: ,in For parameters Next Phenotypic predicted values for each sample To predict the variance of the error, It is a parameter The prior probability distribution, It is the marginal likelihood function, that is, for all parameters Likelihood value after integration.
[0019] Compared with existing technologies, this multimodal deep learning-based forest tree genotype-environment interaction modeling method has the following advantages:
[0020] I. This invention systematically collects multimodal data on tree genotypes, environment, and phenotypes. After preprocessing and fusion, it constructs a genotype-environment interaction algorithm and a phenotype prediction model. This model can accurately predict tree phenotypes, significantly shorten the breeding screening cycle, and greatly improve the accuracy and efficiency of breeding selection. It effectively solves the shortcomings of traditional breeding methods in terms of time and efficiency, enabling breeding work to respond more quickly and accurately to market demands and environmental changes, and providing strong support for the sustainable development of the forestry industry.
[0021] Second, this invention accurately captures the nonlinear relationship between genotype, environment, and phenotype by extracting nonlinear interaction features. Combined with the ability of loss function optimization models to fit complex interaction patterns, it can clearly elucidate the genotype-environment interaction mechanism, making up for the shortcomings of traditional linear models in terms of analytical capabilities. This not only provides solid theoretical support for the study of forest tree growth and development patterns, but also helps researchers to deeply understand the performance mechanism of forest trees in different environments, and helps to develop forest tree varieties that are more adapted to specific environments.
[0022] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0024] Figure 1 The flowchart shows a method for modeling forest tree genotype-environment interactions based on multimodal deep learning.
[0025] Figure 2 This is a flowchart illustrating the development steps of a genotype-environment interaction algorithm for a multimodal deep learning-based modeling method for forest tree genotype-environment interactions. Detailed Implementation
[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0027] This invention provides a method for modeling forest tree genotype-environment interactions based on multimodal deep learning. It systematically collects multimodal data on forest tree genotypes, environment, and phenotypes. After preprocessing and fusion, a creative algorithmic formula is used to construct a genotype-environment interaction algorithm and a phenotype prediction model. Specifically, a high-throughput phenotype platform, whole-genome sequencing technology, and soil nutrient detection equipment are used to collect multimodal data. Targeted preprocessing is performed on various types of data, and the multimodal data are fused to mine data correlation features. Based on the fused data, a deep neural network is constructed to extract the nonlinear interaction features between genotypes and environmental factors, and a genotype-environment interaction algorithm is developed. Based on this interaction algorithm, multiple strategies are introduced to construct and optimize the phenotype prediction model. This addresses the problems of long breeding cycles, low efficiency, and insufficient phenotype prediction accuracy in traditional forest tree breeding, significantly improving the accuracy of forest tree phenotype prediction and breeding efficiency, and providing key technical support for precision breeding and efficient cultivation management of forest trees.
[0028] The following are specific embodiments of the forest tree genotype-environment interaction modeling method based on multimodal deep learning provided by this invention:
[0029] like Figure 1As shown, data collection was conducted using a high-throughput phenotyping platform, whole-genome sequencing technology, and soil nutrient detection equipment. American black poplars planted in different geographical regions were selected as research subjects, with sample sizes ranging from 300 to 1610 trees per region to ensure representativeness and diversity. Whole-genome resequencing of the DNA from collected American black poplar leaf samples was performed using the Illumina high-throughput sequencing platform. Ten trees were randomly selected from each plot, with three biological replicates per tree to ensure SNP calling accuracy. SNP locus information was ultimately obtained to form genotype data. For the soil in the American black poplar growing plots, a SU-LF intelligent soil tester was used to measure key physicochemical indicators such as pH, water content, and available nitrogen / phosphorus / potassium. Five biological replicates were set at each collection point, and the mean of the measurement results was taken to obtain environmental data. A multi-view approach combining UAVs and backpack lidar was used to acquire three-dimensional data of morphological and biomass parameters related to American black poplar growth, such as tree height, diameter at breast height (DBH), crown width, and biomass. The three-dimensional data were then segmented, labeled, and parameters extracted from individual trees to obtain the final phenotypic data.
[0030] The three types of collected data were preprocessed separately. For phenotypic data, a multiple regression model was established based on ground-measured tree height and diameter at breast height (DBH). Parameters such as crown width and branch angle extracted by UAVs and LiDAR were calibrated. A random forest model was used to correct point cloud segmentation bias, and CloudCompare's DBSCAN semi-automatic clustering algorithm was used to optimize segmentation boundaries. Abnormal point cloud clusters were manually verified and marked to ensure accurate individual tree identification, thus completing the phenotypic data calibration and optimization. For environmental data processing, variance expansion factor (VIF) analysis was performed on soil pH, nitrogen, phosphorus, and potassium indicators, and data with VIF > 5 were removed. For highly collinear variables, the remaining indicators were converted into comprehensive environmental factors through principal component analysis, explaining more than 80% of the variance, thus achieving dimensionality reduction and collinearity elimination of environmental data. In genotype data processing, low-quality loci with MAF < 0.05 and deletion rate > 10% were removed, and sequencing data from three biological replicates were merged. Reliability was ensured through a Kappa coefficient > 0.8 consistency test. PCA or STRUCTURE analysis was used to identify population stratification. The first three principal components were included as covariates in subsequent models to complete genotype data quality control and standardization. After preprocessing, the data was fused according to the multimodal data fusion formula. The three types of data are fused to obtain the fused multimodal feature matrix. ,in , , The determination was made through cross-validation optimization. These are the weighting coefficients for the attention mechanism.
[0031] Based on the fused multimodal data, a deep neural network based on a multilayer perceptron architecture is constructed, combined with a hybrid linear model (MLM) infrastructure, such as... Figure 2 As shown, the genotypes that receive the fusion ,environment With phenotype The data is designed based on a multi-head self-attention mechanism, using a genotype-environment nonlinear interaction feature extraction formula. To achieve the extraction of interactive features between genotype and environmental factors in a nonlinear space, whereby... For activation function, , The weight matrix is a linear transformation matrix. For feature splicing operations, This is the bias vector.
[0032] A hierarchical feature fusion strategy is adopted, and genotype-specific modules and environmental response modules are established in the hidden layer of the network. The genotype module uses positional encoding to embed GWAS significance p-values, and the environmental module constructs an LSTM-based temporal feature extractor and couples it with a soil moisture threshold response module. Combined with a cross-modal feature association mechanism, the nonlinear relationship between genotype, environment and phenotype is captured.
[0033] Using phenotypic data as the core response variable, the main effects of genotype (SNP loci), the main environmental effects (comprehensive environmental factors after PCA dimensionality reduction), and their interaction term were analyzed. As fixed effects, plot spatial heterogeneity and principal components of population structure are treated as random effects to construct a multi-level model. The loss function formula of the multi-level interaction model is then used. Optimize model complexity, where, It is the sample size. It is the first The true phenotypic values of each sample It is the first Phenotypic predicted values for each sample It is the L2 regularization coefficient, used to suppress... , Overfitting , They are respectively , The L2 norm is used, and a hybrid Transformer-MLP feature extraction-perception architecture is adopted to improve the algorithm's ability to analyze complex genotype-environment interaction relationships. Finally, a genotype-environment interaction algorithm suitable for American black poplar is developed.
[0034] Based on the interaction algorithm developed above, a feature contrastive learning strategy is introduced to optimize the feature representation space. A knowledge transfer framework is used to address the generalization problem under small sample data. A multi-task deep neural network is constructed using temporal feature learning. This network shares a genotype-environment interaction feature layer and independently sets specific prediction units for each phenotypic trait, such as tree height, diameter at breast height (DBH), crown width, and biomass. An adaptive weight allocation mechanism is designed, and an adaptive weight formula for multi-task phenotypic prediction is used. Dynamically balance the contribution of loss functions to tasks with different phenotypes, where For the first The prediction task weights for each phenotypic trait , The predicted loss values for the corresponding phenotypic traits are used to optimize the model's overall predictive performance for multiple phenotypic traits. A probabilistic optimization method is introduced to adjust the model's hyperparameters, and an improved key factor contribution metric is used. Quantifying the contributions of key SNP sites and environmental driving factors to the target phenotypic trait, among which For the first The SNP locus and the first The combined contribution of individual environmental factors To determine the correlation coefficient between the two factors, we screened for significant contributing factors. Combining gene function annotation information, we identified core regulatory factors influencing the phenotypic formation of American black poplar. Finally, using probabilistic methods, we applied a Bayesian hierarchical model with a posterior probability formula. Construct a hierarchical probability model, in which For parameters The posterior probability distribution, Let be the likelihood function. Using a prior probability distribution, the strength of the genotype-environment interaction effect is estimated through posterior probability to improve the statistical power of small samples. A phenotypic prediction model of American black poplar with clear biological interpretability is established. This model can be used to predict the phenotypic performance of different genotypes of American black poplar under different environmental conditions, thereby screening out superior American black poplar genotypes adapted to specific environments.
[0035] The collected American black poplar multimodal dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter learning, and the validation set was used for hyperparameter tuning (such as the weight coefficients of the attention mechanism). L2 regularization coefficient The test set is used to evaluate the model's generalization ability. During the partitioning process, the distribution ratio of samples in each region across the three datasets is kept consistent to avoid the impact of regional bias on the validation results. Meanwhile, cross-validation is used to further verify the model's stability. The training set and validation set are merged and then randomly divided into 5 mutually exclusive subsets. Each time, 4 subsets are selected as training data and 1 subset is selected as validation data. After repeating this process 5 times, the average performance index is taken as the basic validation result of the model.
[0036] The evaluation indicators are set from three dimensions: prediction accuracy, stability, and biological rationality.
[0037] Prediction accuracy indicators: The coefficient of determination, root mean square error, and mean absolute error are used to evaluate the prediction accuracy of the model for phenotypic traits such as tree height, diameter at breast height, crown width, and biomass. The coefficient of determination is used to measure the model's ability to explain phenotypic variation, and its value ranges from [0,1]. The closer it is to 1, the better the prediction effect. The root mean square error and mean absolute error are used to measure the deviation between the predicted value and the actual value. The smaller the value, the more accurate the prediction.
[0038] Stability metrics: Calculate the coefficient of variation of each performance metric in cross-validation. The smaller the coefficient of variation, the smaller the performance fluctuation of the model on different subsets of data, and the stronger the stability.
[0039] Biological rationality indicators: By comparing the overlap between the key SNP sites quantified by the model and functional genes related to forest growth, the higher the overlap, the stronger the biological significance of the core regulatory factors discovered by the model. At the same time, the influence trend of environmental factors on phenotypic prediction results is analyzed. If the trend is consistent with the conclusions of forest physiological research, it indicates that the model has biological interpretability.
[0040] Two sets of comparative experiments were set up: the traditional machine learning model group: random forest, support vector regression and ordinary linear regression were selected as the comparative models, the same preprocessed data were used for phenotypic prediction, and the evaluation index was consistent with the model of this invention.
[0041] Two simplified versions of the invention model were constructed: a fusion model without the attention mechanism and a multi-task model without knowledge transfer. By comparing the performance differences between the simplified model and the complete model, the role of core modules such as the attention mechanism and knowledge transfer was verified.
[0042] Verification results and analysis of prediction accuracy: The model of this invention achieved determination coefficients of 0.89, 0.86, 0.84, and 0.82 for tree height, diameter at breast height (DBH), crown width, and biomass, respectively; root mean square errors (RMSEs) as low as 0.32m, 0.28cm, 0.45m, and 0.31kg, respectively; and mean absolute errors (MAEs) of 0.25m, 0.21cm, 0.38m, and 0.24kg, respectively. Compared with traditional machine learning models, the determination coefficients are improved by an average of 15%-34%, and the RMS and MAEs are reduced by an average of 20%-40%, indicating that the prediction accuracy of the model of this invention is significantly better than that of traditional models.
[0043] The coefficients of variation for each phenotypic trait prediction index in the model of this invention are all less than 5%, while the coefficients of variation for traditional models are generally between 8% and 15%. This indicates that the model of this invention has stronger stability under different data distributions and is less affected by data fluctuations.
[0044] Biological validity results: Among the key SNP sites mined by the model, 12 overlapped with previously reported genes related to the growth of American black poplar, with an overlap rate of 35%, which is significantly higher than the overlap rate of randomly selected SNPs; and the model prediction results showed that the available nitrogen content in the soil was positively correlated with tree height and biomass, consistent with physiological conclusions, proving that the model has biological interpretability.
[0045] Validation of the core modules: The determination coefficients of each phenotype in the fusion model without the attention mechanism decreased by an average of 8%-12%, and the determination coefficients of the multi-task model without knowledge transfer decreased by 10%-15% in small sample datasets. This shows that the attention mechanism can effectively improve the quality of multimodal data fusion, and the knowledge transfer module can enhance the generalization ability of the model in small sample scenarios. Both are key components of the model.
[0046] The model of this invention was applied to the actual breeding and screening of American black poplar. Based on the model's prediction results, 100 superior genotypes were screened. A two-year field planting experiment was conducted in three different environmental plots. The results showed that the average tree height, diameter at breast height (DBH), and biomass of the screened superior genotypes were 18%-25%, 15%-22%, and 16%-23% higher than those of the randomly selected control genotypes in each environment. The advantages were more obvious in the low-nitrogen soil environment. This indicates that the model can effectively screen out superior genotypes adapted to different environments, significantly shorten the breeding and screening cycle, and greatly improve the breeding efficiency, thus verifying the application value of the model in actual forestry breeding.
[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for modeling forest tree genotype-environment interactions based on multimodal deep learning, characterized in that, The method includes the following specific steps: Multimodal data acquisition: High-throughput phenotyping platform, whole-genome sequencing technology and soil nutrient detection equipment were used to collect SNP locus genotype data of forest trees, key soil physicochemical environmental data and growth-related morphological and biomass parameter phenotypic data. Multimodal data preprocessing and fusion: After calibrating and optimizing phenotypic data, reducing dimensionality and removing collinearity from environmental data, and standardizing genotype data for quality control, the three types of data are fused to obtain the fused multimodal feature matrix. The fusion formula is as follows: ,in, The fused multimodal feature matrix has dimensions of . , For the sample size, For the fused feature dimensions, , , The basic weight coefficients for genotype, environmental, and phenotypic characteristics are determined through cross-validation optimization. This is the preprocessed genotype feature matrix, obtained from SNP loci after quality control, standardization, and dimensionality reduction. Each row represents the genotype information of one forest tree sample, and each column represents a selected SNP locus or genotype principal component. This is the preprocessed environmental feature matrix, obtained from key soil physicochemical indicators after collinearity removal and principal component analysis. Each row represents comprehensive information about the growth environment of a single forest sample, and each column represents a comprehensive environmental factor. This is the preprocessed phenotypic feature matrix, where each row represents the phenotypic information of one forest tree sample, and each column represents one phenotypic trait. These are the attention mechanism weighting coefficients. It is a cross-modal attention feature; Genotype-Environment Interaction Algorithm Development: Based on fused multimodal data, a deep neural network with a multilayer perceptual architecture is constructed. This network receives fused genotype, environment, and phenotypic data and designs a multi-dimensional feature interaction extraction layer. The genotype-environment nonlinear interaction feature extraction formula is used to extract interactive features between genotype and environmental factors in a nonlinear space. A hierarchical feature fusion strategy is adopted, establishing genotype-specific and environment response modules in the network's hidden layers. A cross-modal feature association mechanism is combined to capture the nonlinear relationship between genotype, environment, and phenotype. Simultaneously, a hybrid linear architecture is used as the basic framework, with phenotypic data as the core response variable, genotype main effects, environment main effects, and their interaction term as fixed effects, and plot spatial heterogeneity and population structure principal components as random effects to construct a multi-level model. The model complexity is optimized using a multi-level interaction model loss function formula. A hybrid feature extraction-perceptual architecture is adopted, embedding association analysis significance parameters in the genotype module to construct an environment feature extractor based on time-series feature learning, coupled with an environmental factor threshold response module. This enhances the algorithm's ability to analyze complex genotype-environment interaction relationships, ultimately resulting in a high-performance genotype-environment interaction algorithm. Phenotypic Prediction Model Construction and Optimization: Based on the interaction algorithm, a feature contrastive learning strategy is introduced to optimize the feature representation space. Combined with a knowledge transfer framework, a multi-task deep neural network is constructed using temporal feature learning. By sharing the genotype-environment interaction feature layer, specific prediction units are independently set for each phenotypic trait, and an adaptive weight allocation mechanism is designed. Through the adaptive weight formula for multi-task phenotypic prediction, the adaptive weight allocation mechanism is designed to dynamically balance the loss function contributions of different phenotypic tasks, optimize the model's comprehensive prediction performance for multiple phenotypic traits, and introduce probabilistic optimization methods to adjust the model's hyperparameters. An improved key factor contribution quantification formula is used to quantify the contribution of key SNP sites and environmental driving factors to the target phenotypic trait, screen factors that significantly contribute to the phenotype, and combine gene function annotation information to discover the core regulatory factors affecting phenotypic formation. Combining probabilistic methods, a hierarchical probability model is constructed using the posterior probability formula of the Bayesian hierarchical model. The strength of the genotype-environment interaction effect is estimated through posterior probability to improve the statistical power of small samples and establish a forest tree phenotypic prediction model with clear biological interpretability.
2. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, The genotypic data was obtained by whole-genome resequencing of DNA from tree leaf samples to acquire SNP locus information; the environmental data consisted of key physicochemical indicators of the soil in the tree growth plots, collected from multiple plots with multiple biological replicates for each plot, and the average of the measurement results was taken; the phenotypic data consisted of morphological and biomass parameters related to tree growth, which were acquired from multiple perspectives using a combination of aerial overhead equipment and ground mobile acquisition equipment, and the 3D data was segmented, labeled, and extracted using professional 3D point cloud processing software to obtain the final phenotypic data.
3. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, In the genotype-environment interaction algorithm development steps, the interaction features of genotype and environmental factors in the nonlinear space are extracted using a genotype-environment nonlinear interaction feature extraction formula, which is as follows: ,in, It is a genotype-environment interaction feature matrix with dimension 1. , As the interaction feature dimension, It is an activation function used to introduce non-linear relationships. Linear transformation weight matrices for genotype characteristics and environmental characteristics, respectively, with dimensions of... , , For genotype characteristics, As an environmental characteristic dimension, It is a feature concatenation operation that concatenates the linearly transformed genotype features with environmental features in a dimensional manner, preserving the independent information of the two types of features. It is a bias vector used to adjust the baseline level of the feature output. It is a multilayer perceptron module, containing two layers of linear transformation and one layer of ReLU activation, used for nonlinear mapping of the genotype-environment feature similarity matrix. and It is a similarity matrix between genotype characteristics and environmental characteristics. It is the scaling factor.
4. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, In the genotype-environment interaction algorithm development steps, the model complexity is optimized through a multi-level interaction model loss function formula, which is as follows: ,in, It is the total loss value of the interaction algorithm, used to measure the difference between the model's predicted values and the true phenotypic values. It is the sample size. It is the first The true phenotypic values of each sample It is the first The phenotypic predicted value of each sample is derived from the interaction feature matrix. Obtained through linear mapping, It is the L2 regularization coefficient, used to suppress... , Overfitting , They are respectively , The L2 norm is the square root of the sum of the squares of the elements in the weight matrix. It is the interaction feature regularization coefficient.
5. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, In the phenotypic prediction model construction and optimization steps, an adaptive weight allocation mechanism is designed using a multi-task phenotypic prediction adaptive weight formula, the formula of which is: ,in, It is the first The prediction task weights for each phenotypic trait are used to dynamically balance the loss contributions of different phenotypic tasks. It is the first The predicted loss value for each phenotypic trait. It is the first The predicted loss value for each phenotypic trait. It is an exponential function used to amplify the impact of differences in loss between different tasks on the weights.
6. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, In the phenotypic prediction model construction and optimization steps, the contribution of key SNP loci and environmental driving factors to the target phenotypic trait is quantified by an improved key factor contribution quantification formula, the formula of which is: ,in, It is the first The SNP locus and the first The combined contribution of individual environmental factors It is the first In the nth sample, retain the nth The first SNP site and the second Phenotypic predictions for each environmental factor It is the first In the sample, the first one was removed. The first SNP site and the second Phenotypic predictions for each environmental factor It is the first The SNP locus and the first The correlation coefficients of the environmental factors are used to correct the overestimation of their contributions caused by collinearity.
7. The forest tree genotype-environment interaction modeling method based on multimodal deep learning according to claim 1, characterized in that, In the phenotypic prediction model construction and optimization steps, a hierarchical probability model is constructed using the posterior probability formula of the Bayesian hierarchical model. The formula is as follows: ,in, Model parameters The posterior probability distribution is reflected in the observed data. Below, parameters The credibility of It is a set of model parameters. It is an observation dataset. It is the likelihood function, representing the likelihood of the parameter The observed data The probability of is calculated using the following formula: ,in For parameters Next Phenotypic predicted values for each sample To predict the variance of the error, It is a parameter The prior probability distribution, It is the marginal likelihood function, that is, for all parameters Likelihood value after integration.