Method for predicting acid and aluminum resistance of sugarcane by using GxE deep learning model

Abstract

The invention discloses a method for predicting the acid and aluminum resistance of sugarcane by using a GxE deep learning model. The method comprises the following steps: integrating sugarcane planting data sources; carrying out preliminary feature screening by adopting XGBoost + SHAP; feature redundancy removal is carried out through MIC + hierarchical clustering; a generator is used, and a conditional expression Wasserstein GAN is designed to generate high-quality synthetic data; a three-level evaluation process of MMD global inspection, KL divergence core gene verification and environment-gene dynamic correlation analysis is adopted; obtaining noise reduction data based on a hierarchical self-encoding architecture of noise injection and robust reconstruction; capturing time dynamic characteristics of gene expression; inhibiting or enhancing specific gene expression; and predicting and outputting the acid and aluminum resistance probability of the sugarcane through full connection. According to the method, the accuracy and interpretability of predicting the acid aluminum resistance of the sugarcane are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of predicting aluminum tolerance in sugarcane, and particularly to a method for predicting aluminum tolerance in sugarcane by applying a GxE deep learning model. Background Art

[0002] In the main sugarcane planting areas of China, the problems of soil acidification and high aluminum stress are generally faced. Among them, aluminum usually exists in the form of non-toxic aluminosilicates or aluminum oxides. However, when the soil pH value is lower than 5.5, a large amount of active aluminum (Al 3+ ) is dissolved out, which has a direct toxic effect on the sugarcane roots, resulting in limited inhibition of water and nutrient absorption, and ultimately leading to a decrease in yield and quality. Traditionally, the evaluation of aluminum tolerance mainly relies on field experiments and detection of conventional physiological indexes. However, these methods are restricted by factors such as large measurement errors, insignificant responses, and many environmental interferences, resulting in insufficient stability and sensitivity of the evaluation results. To alleviate the acid-aluminum stress, soil conditioners such as lime are mostly used in current production. However, this method has poor improvement effect on deep soil, high long-term use cost, and may damage the soil ecology. Therefore, breeding or screening sugarcane varieties with aluminum tolerance has become an economic solution. In recent years, with the development of molecular biology and intelligent algorithms, the comprehensive evaluation of aluminum tolerance by integrating genomic data and environmental variables has gradually become a research hotspot, providing new ideas and technical support for the screening of aluminum-tolerant varieties.

[0003] Currently, some studies have attempted to use multiple physiological and morphological indexes to judge the aluminum tolerance of crops through principal component analysis and weighted comprehensive evaluation methods. Such methods reflect the differences among different varieties under aluminum toxicity stress to a certain extent. However, the evaluation process often focuses on static physiological data and fails to fully consider the dynamic changes of the complex and non-linear interaction between genes and the environment. In addition, traditional data processing methods have low robustness to insufficient sample size and environmental variable interference, resulting in the prediction results being easily affected by external factors and being unstable. The aluminum tolerance trait is significantly affected by the gene-environment interaction, and the existing methods fail to organically integrate molecular level and environmental information. Therefore, there are problems such as low accuracy and limited applicability in the process of tolerance prediction and variety screening. In recent years, dynamic models based on deep learning have shown the advantage of efficiently extracting hidden features in other fields, but their application in the evaluation of aluminum tolerance in sugarcane is still in the exploratory stage. There is an urgent need for a new method to overcome the above deficiencies and further improve the reliability of evaluation and the accuracy of prediction. Summary of the Invention

[0004] In order to overcome the above-mentioned shortcomings of the prior art, the present invention proposes a method for predicting aluminum tolerance in sugarcane by applying a GxE deep learning model, which significantly improves the accuracy and interpretability of predicting aluminum tolerance in sugarcane.

[0005] The technical solution adopted by the present invention to solve its technical problems is as follows: A method for predicting the aluminum tolerance of sugarcane using a GxE deep learning model, comprising the following steps:

[0006] (1) Integrate the data sources of sugarcane cultivation, specifically including the following sub-steps:

[0007] (1-1) Collect RNA-seq transcriptome data of different sugarcane varieties under aluminum stress conditions, filter out low-expressed genes and perform standardization processing using Z-score to extract gene expression features highly correlated with aluminum tolerance traits;

[0008] (1-2) Collect and construct a standardized environmental dataset, and simultaneously measure the tolerance physiological indicators of sugarcane under different aluminum stress levels; the standardized environmental dataset includes key environmental variables; the key environmental variables include soil pH value, aluminum content, rainfall, soil organic matter content, temperature; the tolerance physiological indicators include: root length, chlorophyll content, plant biomass;

[0009] (1-3) Use data preprocessing methods such as missing value filling, normalization, and denoising for all data to ensure data consistency and reliability and eliminate the interference of outliers on the model;

[0010] (2) Use XGBoost+SHAP for preliminary feature screening, combine the global segmentation contribution and local marginal effect to accurately screen genes and environmental variables strongly related to aluminum tolerance traits. The whole process includes the first stage and the second stage;

[0011] In the first stage, based on the Gain index of the XGBoost model, quantify the global contribution degree of features during the decision tree splitting, and initially eliminate redundant features for screening through dynamic threshold setting;

[0012] In the second stage, for the screened features, use the SHAP value to quantify their local marginal contribution to the prediction result;

[0013] (3) Perform feature redundancy removal through MIC+hierarchical clustering, and calculate the maximum information coefficient MIC between features;

[0014] (4) Use a generator to design a conditional Wasserstein GAN to generate high-quality synthetic data;

[0015] (5) Adopt a three-level evaluation process of MMD global test, KL divergence core gene verification, and environment-gene dynamic association analysis to ensure that the generated data simultaneously meets statistical distribution consistency and biological rationality;

[0016] (6) Hierarchical auto-encoding architecture based on noise injection and robust reconstruction, designing a training strategy in combination with the characteristics of sugarcane gene data for hierarchical DAE noise reduction to obtain noise-reduced data;

[0017] (7) Screening the noise-reduced data in step (6) where T is the number of time steps and p is the number of genes; and inputting the screened noise-reduced data into the GRU-Attention network to capture the time dynamic characteristics of gene expression;

[0018] (8) Inputting environmental variables where q is the environmental factor, using the LSTM model to capture the environmental dynamic time series, calculating the regulation intensity of the environment on gene expression, and inhibiting or enhancing the expression of specific genes;

[0019] (9) Through gene feature modulation and environment-gene splicing performing feature interaction, and fully connecting to predict and output the probability of aluminum tolerance in sugarcane.

[0020] Furthermore, the calculation of the Gain index is shown in the following formula ①:

[0021]

[0022] where g and h are the first-order and second-order gradients respectively, and λ is the regularization coefficient;

[0023] Furthermore, the SHAP value is calculated according to the following formula ②

[0024]

[0025] where x \j represents the conditional expectation excluding the feature f j , and the features with the top 95% contribution of the SHAP value are selected.

[0026] Furthermore, in the said step (3), calculating the maximum information coefficient MIC between features specifically according to formula ③ and formula ④ to capture linear and non-linear associations:

[0027]

[0028] where I(f j , f k ) is the mutual information, G is the grid division method, n G , m G are the number of rows and columns of the grid; for the screened feature set, calculate the MIC values of all feature pairs to form a symmetric matrix M m×m , and the matrix element M jk = MIC(f j , fk ) Convert the MIC matrix into a distance matrix D, where D jk = 1 - M jk ④, Use the Ward variance minimization method for agglomerative hierarchical clustering, and the merging strategy is to minimize the incremental variance within the cluster:

[0029]

[0030] where n A , n B are the number of samples in clusters A and B, μ A , μ B are the cluster centers; Determine the number of clusters according to the cutting height of the dendrogram, set the threshold τ = 0.6, if the average MIC between features within the cluster ≥ 0.6, then strong collinearity is considered to exist; In each cluster, select the feature with the highest SHAP mean as the representative, and the rest are marked as redundant and removed;

[0031] Furthermore, in the step (4), the input of the generator is the Gaussian noise vector z and the conditional vector c, and synthetic gene expression data G(z|c) is generated through a fully connected network; The input of the discriminator is the real / synthetic data x and the conditional vector c, and the output is the data real score D(x|c); Use the formula ⑥ as the objective function for training:

[0032]

[0033] where, λ1 = 10 is the standard gradient penalty coefficient, λ2 = 5 is the sparse perception penalty weight, is the linear interpolation sample of the real data and the generated data, I(x i = 0) is the indicator function, which takes 1 when the expression value of gene i is 0, otherwise takes 0; If the pH value in the generated data exceeds the range [4.0, 5.5] or the AL3+ concentration > 50 μM, then remove the sample.

[0034] Furthermore, in the step (5), the MMD global test includes the step of calculating MMD according to formula ⑦:

[0035]

[0036] If MMD 2 ≥ 0.05, then retrain the generator; where n is the number of real samples, m is the number of generated samples, is the kernel function, and σ takes the median of the Euclidean distances between real data samples; Define the core gene set G core = {g1, g2,..., g 20} and calculate the KL divergence between the generated data and the real data according to formula ⑨:

[0037]

[0038] If DKL < 0.1, it is determined that the distribution of core genes conforms to the biological law; for the environmental-gene dynamic association analysis, the AL in the generated data is extracted. 3+ The concentration and the SbMATE gene expression sequence are used to calculate the Spearman rank correlation coefficient ρ. If ρ ≥ 0.6, it is determined that the generated data retains the true biological association characteristics.

[0039] Further, in step (6), two types of noise are imposed on the input data. One is random masking: the conventional gene expression values are set to zero with a probability of 30%, and the masking rate of core genes is reduced to 15% to reduce the loss of key signals. The other is adaptive Gaussian noise: the noise intensity is dynamically adjusted according to σ = 0.1 × log2(FPKM g +1) to avoid over-smoothing of highly expressed genes. The encoder and decoder are set. The encoder is set as a three-layer fully connected network, specifically p → 0.8p → 0.5p, and the activation function is LeakyRelu with α = 0.2; the low-dimensional feature representation is extracted; the decoder uses a symmetric structure, and the output layer uses the Sigmoid function to reconstruct the denoised data. The formula is According to the formula Calculate the weighted reconstruction loss:

[0040]

[0041] where the weight ω of the core gene g = 2.0, and the rest ω g = 1.0; the Adam optimizer is used, and the learning rate of this Adam optimizer is 1 × 10-3; the momentum is β1 = 0.9, β2 = 0.999; the batch size is 64, and the early stopping strategy is adopted.

[0042] Further, the network architecture of the GRU-Attention network includes:

[0043] Bidirectional GRU layer: the number of hidden units h gru = 128, capturing forward and backward temporal dependencies: Output concatenation:

[0044] Multi-head attention mechanism: the number of heads N head = 8, calculating the scaled dot-product attention according to the formula where the input H gru is split into The output attention weight matrix A is used for feature enhancement, and residual connection and layer normalization are performed, and finally the enhanced features are output

[0045] Further, in step (8), the regulation intensity of the environment on gene expression is calculated as follows: According to the formula The output is the gating weight g ∈ [0, 1] p .

[0046] Further, the training method in step (9) is as follows: Set up a fully connected prediction layer: The hidden layer is 256 - dimensional, the activation function is ReLU, the output layer uses Sigmoid activation, and the output is the probability y of acid - aluminum tolerance. Use the binary cross - entropy loss function:

[0047]

[0048] During the training process, select the AdamW optimizer, set grid search for hyperparameter tuning, and stop when the AUC of the validation set does not improve for 15 consecutive epochs.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] The AUC of the deep - learning model combining GxE established by the present invention reaches 0.92, which is significantly better than XGBoost (AUC = 0.78) and the fully connected neural network (AUC = 0.65). The F1 - score is increased to 0.87, and the specificity reaches 0.91, with better computational efficiency. In this way, the probability of sugarcane's acid - aluminum tolerance is finally obtained, which is significantly improved in terms of indicators such as AUC, F1 - score, and specificity compared with the traditional model, significantly improving the accuracy and interpretability of sugarcane acid - aluminum tolerance prediction. Description of the Drawings

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments.

[0052] Figure 1 It is a flowchart of a method for predicting sugarcane acid - aluminum tolerance by applying a GxE deep - learning model according to an embodiment of the present invention. Detailed Embodiments

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention.

[0055] In the description of the present invention, it should be noted that unless otherwise clearly specified and limited, the terms "installed", "connected", "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations. In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] As Figure 1 shown, a method for predicting aluminum tolerance of sugarcane by applying a GxE deep learning model includes the following steps:

[0057] (1) Integrate multiple data sources such as sugarcane genome, environmental factors, and physiological phenotype information to construct a high-precision aluminum tolerance prediction model. Specifically, it includes the following sub-steps:

[0058] First, collect RNA-seq transcriptome data of different sugarcane varieties under aluminum stress conditions, filter out low-expression genes and perform standardization processing using Z-score to extract gene expression characteristics highly related to aluminum tolerance traits. Secondly, collect and construct a standardized environmental data set, including key environmental variables: soil pH value, aluminum content (AL 3+ concentration), rainfall, soil organic matter content, temperature. At the same time, measure the tolerance physiological indexes of sugarcane under different aluminum stress levels. The tolerance physiological indexes include: root length, chlorophyll content, plant biomass. To ensure data quality, all data are subjected to data preprocessing methods such as missing value filling, normalization, and denoising to ensure data consistency and reliability and eliminate the interference of outliers on the model.

[0059] (2) Use XGBoost+SHAP for preliminary feature screening, combine global segmentation contribution and local marginal effect, and accurately screen genes and environmental variables strongly related to aluminum tolerance traits. The whole process includes the first stage and the second stage;

[0060] In the first stage, based on the Gain metric of the XGBoost model, the global contribution degree of features during the splitting of decision trees is quantified. Through dynamic threshold setting (cumulative contribution degree reaches 95%), redundant features are initially removed. The Gain calculation is shown in the following formula ①:

[0061]

[0062] where g and h are the first-order and second-order gradients respectively, and λ is the regularization coefficient.

[0063] In the second stage, for the filtered features, the SHAP value is used to quantify their local marginal contribution to the prediction result. The SHAP value is calculated according to the following formula ②:

[0064]

[0065] where x \j represents the conditional expectation excluding feature f j . Select the features with the top 95% SHAP contributions.

[0066] (3) Feature redundancy removal is performed through MIC + hierarchical clustering. Calculate the maximum information coefficient MIC between features to solve the collinearity problem. Calculate the maximum information coefficient (MIC) between features according to formula ④ to capture linear and non-linear associations:

[0067]

[0068] where I(f j , f k ) is the mutual information, G is the grid partitioning method, n G , m G are the number of rows and columns of the grid. For the filtered feature set, calculate the MIC values of all feature pairs to form a symmetric matrix M m×m , and the matrix element M jk = MIC(f j , f k ). Convert the MIC matrix to a distance matrix D, where D jk = 1 - M jk ④. Use Ward's method of minimizing variance within clusters for agglomerative hierarchical clustering, and the merging strategy is to minimize the incremental variance within clusters:

[0069] where n A , n B are the number of samples in clusters A and B, μ A , μ Bis the cluster center. The number of clusters is determined according to the cut height of the dendrogram, and the threshold τ = 0.6 is set. If the average MIC among the features within the cluster ≥ 0.6, it is considered that there is strong collinearity. In each cluster, the feature with the highest SHAP mean is selected as the representative, and the rest are marked as redundant and removed.

[0070] (4) For the characteristics of high dimensionality and small samples of RNA-seq data, a generator is used to design a conditional Wasserstein GAN (cWGAN) to generate high-quality synthetic data and improve the generalization ability of the model. The input of the generator is the Gaussian noise vector z and the conditional vector c (including environmental variables and key gene labels), and synthetic gene expression data G(z|c) is generated through a fully connected network; the input of the discriminator is the real / synthetic data x and the conditional vector c, and the real score D(x|c) of the data is output. Formula ⑥ is used as the objective function for training:

[0071]

[0072] Among them, λ1 = 10 is the standard gradient penalty coefficient, and λ2 = 5 is the sparse perception penalty weight. is the linear interpolation sample of the real data and the generated data, and I(x i = 0) is the indicator function, which takes 1 when the expression value of gene i is 0, and 0 otherwise. If the pH value in the generated data exceeds the range of [4.0, 5.5] or the AL3+ concentration > 50 μM, the sample is removed.

[0073] (5) Adopt a three-level evaluation process of MMD global test, KL divergence core gene verification, and environment-gene dynamic association analysis to ensure that the generated data meets both statistical distribution consistency and biological rationality. Calculate MMD according to formula ⑦:

[0074]

[0075] If MMD 2 ≥ 0.05, retrain the generator. Among them, n is the number of real samples, m is the number of generated samples, is the kernel function, and σ takes the median of the Euclidean distances between real data samples. Define the core gene set G core = {g1, g2,..., g 20} and calculate the KL divergence between the generated data and the real data according to formula ⑨:

[0076]

[0077] If DKL < 0.1, it is determined that the distribution of core genes conforms to biological laws. The environment-gene dynamic association analysis extracts AL in the generated data 3+Calculate the Spearman rank correlation coefficient ρ for the concentration and the SbMATE gene expression sequence. If ρ ≥ 0.6, it is determined that the generated data retains the true biological association characteristics.

[0078] (6) Hierarchical autoencoder architecture based on noise injection and robust reconstruction. Combining the characteristics of sugarcane gene data, design a training strategy to perform hierarchical DAE denoising to obtain denoised data. First, apply two types of noise to the input data. One is random masking: set the conventional gene expression values to zero with a probability of 30%, and reduce the masking rate of core genes (such as SbMATE and SbSTOP1) to 15% to reduce the loss of key signals. The other is adaptive Gaussian noise: the noise intensity is dynamically adjusted according to σ = 0.1 × log2(FPKM g +1) to avoid over-smoothing of highly expressed genes. Then, set up the encoder and decoder. The encoder is set as a three-layer fully connected network (p → 0.8p → 0.5p), and the activation function is LeakyRelu(α = 0.2). Extract the low-dimensional feature representation; the decoder uses a symmetric structure, and the output layer uses the Sigmoid function to reconstruct the denoised data, and the formula is ⑩. Finally, calculate the weighted reconstruction loss according to the formula :

[0079]

[0080] where the weight ω of the core gene g = 2.0, and the rest ω g = 1.0. Use the Adam optimizer (learning rate 1×10-3, momentum β1 = 0.9, β2 = 0.999), the batch size is 64, and the early stopping strategy.

[0081] (7) Screen the denoised data in step (6) (T is the number of time steps, p is the number of genes) input into the GRU-Attention network to capture the temporal dynamic characteristics of gene expression. The network architecture includes:

[0082] 1. Bidirectional GRU layer. The number of hidden units h gru = 128, capturing forward and backward temporal dependencies: Output concatenation:

[0083] 2. Multi-head attention mechanism. The number of heads N head = 8, calculate the scaled dot-product attention according to the formula where the input H gru , split into Output the attention weight matrix A, use residual connection and layer normalization after feature enhancement, and finally output the enhanced feature

[0084] (8) Input environmental variables (where q is the environmental factor), use the LSTM model to capture the dynamic time series of the environment, and according to the formula Calculate the regulation intensity of the environment on gene expression, and the output is the gating weight g ∈ [0, 1] p , inhibiting or enhancing the expression of specific genes.

[0085] (9) Through gene feature modulation and environment-gene splicing Perform feature interaction, and the fully connected prediction outputs the probability of aluminum tolerance in sugarcane, that is, the probability of aluminum tolerance.

[0086] The training method is as follows: Set the fully connected prediction layer: the hidden layer is 256-dimensional, the activation function is ReLU, the output layer uses Sigmoid activation, and the output is the probability of aluminum tolerance y. Use the binary cross-entropy loss function:

[0087]

[0088] During training, select the AdamW optimizer, set grid search to optimize hyperparameters (learning rate, weight decay, and Dropout), and stop when the AUC of the validation set does not improve for 15 consecutive epochs.

[0089] Select sugarcane varieties ROC22 (aluminum-tolerant type) and Badila (sensitive type) as the research objects, and collect a total of 200 samples, including 100 aluminum-tolerant types and 100 sensitive types, covering 5 growth stages: seedling stage, tillering stage, jointing stage, booting stage, and maturity stage. The detection data includes RNA-seq gene expression (25,000 genes), environmental variables (pH 4.0 - 5.5, exchangeable Al3+ concentration 0 - 50 μM), and phenotypic data (aluminum tolerance score, based on relative root elongation rate).

[0090] After data preprocessing, use the XGBoost-SHAP method to screen out 300 core features, including 285 genes and 15 environmental variables. The key genes include SbMATE (Gain ranking 1st, SHAP mean = 0.153), SbSTOP1 (Gain ranking 2nd, SHAP mean = 0.142), and the key environmental variables include soil pH (SHAP mean = 0.098), Al 3 + concentration (SHAP mean = 0.085). Use the GAN model to generate 1000 synthetic samples, and through quality assessment (MMD 2= 0.038, SbMATE KL divergence = 0.085), 820 high-quality samples were selected. In addition, the DAE noise reduction method was used to improve the data quality. After denoising, the signal-to-noise ratio of SbMATE increased from 15.2 dB to 30.4 dB, and the number of differentially expressed genes decreased by 68% (1,205 → 387, FDR < 0.05).

[0091] The prediction model adopts the GRU-Attention architecture, in which the bidirectional GRU (h = 128) combines 8-head self-attention (d_model = 256) to process RNA-seq data, the LSTM (hidden unit 64) encodes environmental variables, and the fully connected layer (256 → 1, Sigmoid activation) makes the final prediction. The training uses the AdamW optimizer (lr = 3e-4, weight decay = 1e-4), Dropout = 0.3, Label Smoothing = 0.1, and the early stopping strategy is to terminate if there is no improvement in the validation loss within 15 epochs. The experimental results show that the AUC of this model reaches 0.92, which is significantly better than XGBoost (AUC = 0.78) and the fully connected neural network (AUC = 0.65). The F1-score is improved to 0.87, and the specificity reaches 0.91, with better computational efficiency. In this way, the probability of aluminum tolerance in sugarcane is finally obtained, which is significantly improved in terms of indicators such as AUC, F1-score, and specificity compared with the traditional model, significantly improving the accuracy and interpretability of the prediction of aluminum tolerance in sugarcane.

[0092] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A method for predicting aluminum tolerance in sugarcane using a GxE deep learning model, characterized in that, Including the following steps: (1) Integrate the data sources of sugarcane cultivation, specifically including the following sub-steps: (1-1) Collect RNA-seq transcriptome data of different sugarcane varieties under acid-aluminum stress conditions, filter out low-expression genes and perform normalization using Z-score to extract gene expression features highly related to acid-aluminum tolerance traits; (1-2) Collect and construct a standardized environmental dataset, and measure the tolerance physiological indicators of sugarcane under different acid-aluminum stress levels at the same time; the standardized environmental dataset includes key environmental variables; the key environmental variables include soil pH value, aluminum content, rainfall, soil organic matter content, temperature; the tolerance physiological indicators include: root length, chlorophyll content, plant biomass; (1-3) Use data preprocessing methods of missing value filling, normalization, and denoising for all data to ensure data consistency and reliability and eliminate the interference of outliers on the model; (2) Use XGBoost+SHAP for preliminary feature screening, combine global segmentation contribution and local marginal effect, and accurately screen genes and environmental variables strongly related to acid-aluminum tolerance traits. The whole process includes the first stage and the second stage; In the first stage, based on the Gain index of the XGBoost model, quantify the global contribution degree of features when splitting decision trees, and preliminarily screen out redundant features through dynamic threshold setting; In the second stage, for the screened features, use the SHAP value to quantify their local marginal contribution to the prediction result; (3) Perform feature redundancy removal through MIC+hierarchical clustering, and calculate the maximum information coefficient MIC between features; (4) Use a generator to design a conditional Wasserstein GAN to generate high-quality synthetic data; (5) Adopt a three-level evaluation process of MMD global test, KL divergence core gene verification, and environment-gene dynamic association analysis to ensure that the generated data meets both statistical distribution consistency and biological rationality; (6) Based on a hierarchical autoencoder architecture of noise injection and robust reconstruction, design a training strategy in combination with the characteristics of sugarcane gene data for hierarchical DAE denoising to obtain denoised data; (7) Screen the denoised data in step (6). Where T is the number of time steps and p is the number of genes; and input the screened denoised data into the GRU-Attention network to capture the temporal dynamic features of gene expression. (8) Input environmental variables Where q is the environmental factor, and the LSTM model is used to capture the dynamic time series of the environment, calculate the regulation intensity of the environment on gene expression, and inhibit or enhance the expression of specific genes; (9) Modulating through gene characteristics and environment-gene splicing Perform feature interaction, and the fully connected prediction outputs the probability of aluminum tolerance in sugarcane.

2. The method for predicting aluminum tolerance of sugarcane by applying a GxE deep learning model according to claim 1, wherein The calculation of the Gain index is shown in the following formula ①: Where g and h are the first-order and second-order gradients respectively, and λ is the regularization coefficient.

3. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, wherein Calculate the SHAP value according to the following formula ② Among them, x \j represents the conditional expectation excluding feature f j and select the features with the top 95% SHAP value contributions.

4. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that, In the step (3), calculate the maximum information coefficient MIC between features according to formula ④ to capture linear and non-linear associations: where I(f j , f k ) is the mutual information, G is the grid division method, n G , m G are the number of rows and columns of the grid; for the filtered feature set, calculate the MIC values of all feature pairs to form a symmetric matrix M m×m , and the matrix element M jk = MIC(f j , f k ); convert the MIC matrix to a distance matrix D, where D jk = 1 - M jk ④, and use the Ward minimum variance method for agglomerative hierarchical clustering, and the merging strategy is to minimize the increase in within-cluster variance: where n A , n B is the number of samples in clusters A and B, μ A , μ B is the cluster center; determine the number of clusters according to the cutting height of the dendrogram, set the threshold τ = 0.6, if the average MIC among the features within the cluster ≥ 0.6, then strong collinearity is considered to exist; in each cluster, select the feature with the highest SHAP mean as the representative, and the rest are marked as redundant and removed.

5. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that, In the step (4), the input of the generator is the Gaussian noise vector z and the conditional vector c, and synthetic gene expression data G(z|c) is generated through a fully connected network; the input of the discriminator is the real / synthetic data x and the conditional vector c, and the output data real score D(x|c) is output; use formula ⑥ as the objective function for training: Among them, λ1 = 10 is the standard gradient penalty coefficient, and λ2 = 5 is the sparse perception penalty weight. is the linear interpolation sample of the real data and the generated data, I(x i = 0) is the indicator function that takes 1 when the expression value of gene i is 0 and 0 otherwise; if the pH value in the generated data exceeds the range of [4.0, 5.5] or the AL3+ concentration > 50 μM, then this sample is removed.

6. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that In the step (5), the MMD global test includes the step of performing MMD calculation according to formula ⑦: If MMD 2 ≥ 0.05, then retrain the generator; where n is the number of real samples and m is the number of generated samples, is the kernel function, and σ takes the median of the Euclidean distances between real data samples; define the core gene set G core ={g1, g2, …, g 20}, calculate the KL divergence between the generated data and the real data according to Equation ⑨: If DKL < 0.1, it is determined that the distribution of core genes conforms to biological laws; environmental-gene dynamic association analysis extracts the AL in the generated data 3+ concentration and the SbMATE gene expression sequence, calculates the Spearman rank correlation coefficient ρ, and if ρ ≥ 0.6, it is determined that the generated data retains the true biological association characteristics.

7. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that, Step (6), apply two types of noise to the input data. One is random masking: set the conventional gene expression values to zero with a probability of 30%, and reduce the masking rate of core genes to 15% to reduce the loss of key signals. The other is adaptive Gaussian noise: the noise intensity is dynamically adjusted according to σ = 0.1×log2(FPKM g +1) to avoid over-smoothing of highly expressed genes; perform the settings of the encoder and decoder. The encoder is set as a three-layer fully connected network, specifically p→0.8p→0.5p, and the activation function is LeakyRelu with α = 0.2; extract the low-dimensional feature representation; The decoder uses a symmetric structure, and the output layer uses the Sigmoid function to reconstruct the denoised data. The formula is According to the formula calculate the weighted reconstruction loss: Among them, the weight ω of the core gene g = 2.0, and the rest ω g = 1.0; The Adam optimizer is adopted, and the learning rate of this Adam optimizer is 1×10-3; The momentum is β1 = 0.9, β2 = 0.999; The batch size is 64, and the early stopping strategy is adopted.

8. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that The network architecture of the GRU-Attention network includes: Bidirectional GRU layer: number of hidden units h gru = 128, capturing forward and backward temporal dependencies: Output concatenation: Multi-Head Attention Mechanism: Number of Heads N head = 8, calculate scaled dot-product attention according to the formula where the input H gru is split into Output the attention weight matrix A, perform feature enhancement using residual connection and layer normalization, and finally output the enhanced feature 9. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that In step (8), calculate the regulation intensity of the environment on gene expression, specifically: According to the formula The output is the gating weight g ∈ [0, 1] p .

10. A method for predicting aluminum tolerance of sugarcane using a GxE deep learning model according to claim 1, characterized in that, In step (9), the training method is as follows: Set up a fully connected prediction layer. The hidden layer has 256 dimensions, the activation function is ReLU, the output layer uses Sigmoid activation to output the acid-resistant aluminum probability y, and the binary cross-entropy loss function is used: During the training process, the AdamW optimizer was selected, and grid search was set for hyperparameter tuning, stopping when the AUC of the validation set did not improve for 15 consecutive epochs.

Images