A method for detecting whole genome gene-gene and gene-environment interaction based on an interpretable genetic information neural network

Abstract

This invention discloses a genome-wide gene-gene and gene-environment interaction detection method based on an interpretable genetic information neural network, comprising the following steps: Step 1, training a deep neural network model using genotype and phenotypic data to obtain a trained deep neural network model; Step 2, inputting the genomic data of a sample and using the trained deep neural network model to predict the phenotypic value of the sample; Step 3, calculating the xPI value of each genomic variant site and performing genome-wide association analysis and genetic locus detection using the xPI values. This invention accurately predicts target traits by constructing a deep neural network model and combining genotype and phenotypic data of the sample. Through post-training network analysis, xAI-GWAS not only quantifies the importance of each genotype data point but also detects additive effects, gene interaction effects, and gene-environment interaction effects of genotype data.

Description

Technical Field

[0001] This invention relates to a whole-genome gene-gene and gene-environment interaction detection method based on an interpretable genetic information neural network, belonging to the fields of plant and animal breeding and artificial intelligence technology. Background Technology

[0002] Identifying phenotypic differences and causal relationships between genes is a core challenge in biology, and genome-wide association studies (GWAS) are a primary tool for achieving this goal. Traditional GWAS methods mainly focus on additive gene effects, making it difficult to effectively analyze gene-gene and gene-environment interactions. However, current biological research reveals that polygenic effects often work synergistically through interconnected biological pathways. Related studies suggest that epistasis may be a significant cause of the "loss of heritability" phenomenon. Ignoring this interaction effect would severely distort heritability estimates and reduce the reproducibility of studies.

[0003] Gene-gene interactions and gene-environment interactions introduce complex nonlinear statistical relationships, posing a significant challenge to traditional GWAS methods. The current mainstream GWAS framework is the Linear Mixed Model (LMM), which, while able to correct for population structure and sample correlations through random effects, struggles to capture higher-order genetic effects (such as gene-gene interactions and gene-environment interactions) due to its linear assumptions. Methods such as Bayesian hierarchical models can model specific nonlinear patterns, but rely on explicitly specifying interaction terms and often fail to identify true interactions due to confounding factors; models that explicitly consider heterogeneity require large sample sizes, while latent variable methods face enormous computational burdens.

[0004] Complex traits are typically determined by multiple gene loci, small-effect genes, environmental factors, and their interactions. Traditional GWAS relies on single-marker significance tests (such as p-values), making accurate predictions difficult by summing significant loci. Genome-wide regulatory models suggest that marginal genes also regulate traits, increasing the need for efficient prediction methods. Since linear models cannot characterize higher-order nonlinear effects, researchers have begun to explore machine learning and deep neural networks to improve predictive capabilities.

[0005] While existing research has attempted to leverage machine learning to uncover nonlinear associations using methods such as feature importance, these methods are highly dependent on model design and validation. For example, some studies have indicated that the ability of random forest algorithms to detect interactions may stem more from marginal effects than from genuine gene interactions. Although deep neural networks (DNNs) have shown potential in phenotypic modeling, there is currently no theoretically complete or empirically sufficient method to reliably identify intergenetic or gene-environment interactions.

[0006] Therefore, there is an urgent need for a nonlinear modeling framework with a solid theoretical foundation and practical applications to overcome the performance bottleneck of traditional GWAS in detecting complex interactions, and to provide interpretable biological meaning while ensuring prediction accuracy. This is precisely the technical problem that this invention aims to solve. Summary of the Invention

[0007] The purpose of this invention is to provide a genome-wide gene-gene and gene-environment interaction detection method based on an interpretable genetic information neural network. The aim is to accurately predict target traits by constructing a deep neural network model and combining genotype and phenotypic data of the sample. Through post-training network analysis, this method not only quantifies the importance of each genotype data, but also detects the additive effects, gene interaction effects, and gene-environment interaction effects of genotype data.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] In a first aspect, the present invention provides a method for detecting genome-wide gene-gene and gene-environment interactions based on an interpretable genetic information neural network, comprising the following steps:

[0010] Step 1: Train a deep neural network model using genotype and phenotypic data to obtain a trained deep neural network model;

[0011] Step 2: Input the genomic data of the sample and use the trained deep neural network model to predict the phenotypic value of the sample;

[0012] Step 3: Calculate the xPI value for each genomic variant site, and use the xPI value to perform genome-wide association analysis and genetic locus detection.

[0013] Furthermore, the deep neural network model is an interpretable deep neural network driven by genetic prior information, which includes the following components:

[0014] (1) Multi-step decision module, each decision module is used to receive input features and generate feature representations through feature converter, and finally aggregate the feature representations of all modules through output module to generate final prediction results;

[0015] (2) Feature converter, used to generate multi-dimensional feature representations through a multi-layer neural network structure;

[0016] (3) Splitter, used to effectively select features and remove redundant features to improve model efficiency;

[0017] (4) Output module: aggregate the feature representations of all decision modules and process them through a fully connected layer to finally generate the prediction results.

[0018] Furthermore, the feature converter includes the following sub-modules:

[0019] Fully connected layers: perform linear transformations on the input, providing rich feature representations for subsequent nonlinear processing;

[0020] Batch normalization layer: Standardizes the input of each layer, improves training stability, and accelerates convergence;

[0021] Gated linear units: These units selectively output features through a gating mechanism, thereby enhancing the model's learning ability.

[0022] Residual connections: ensure the stability of network training and avoid gradient vanishing.

[0023] Furthermore, by designing a loss function, genetic prior information is effectively incorporated into model training. This loss function comprises two parts:

[0024] (1) Data-driven terms: Minimize the difference between model predictions and actual observations to ensure accurate fitting of phenotypic data;

[0025] (2) Genetic prior knowledge driven terms: By modeling the genetic relationship matrix and combining genetic prior information to optimize the genetic constraints in the training process, the model’s ability to identify complex genetic effects is enhanced.

[0026] Furthermore, the method for calculating the xPI value includes:

[0027] (1) Based on the SHAP method in game theory, the predictive contribution of each genetic marker to the target trait is quantified at the single-sample level;

[0028] (2) Summarize the SHAP values ​​of all samples, calculate the overall influence of each genetic marker in the dataset, and thus obtain the global feature importance ranking; the xPI value is calculated using the following formula:

[0029] (1)

[0030] in Let be the SHAP of feature Xj for the i-th sample, and N be the total number of samples; xPI index represents the overall contribution of feature Xj to the model prediction in the entire dataset.

[0031] Furthermore, the following variations can be used in xPI calculation:

[0032] (1) In formula (1), for The absolute value of the value is squared to enhance the influence of high-contribution genetic markers;

[0033] (2) In formula (1), for Perform function transformations to measure the contribution of each genetic marker;

[0034] (3) In formula (1), the contribution of each sample is ( absolute value Square value The function transformation of the value assigns different weights, and refines the contribution of each sample's xPI based on the characteristics of each sample, such as sample quality and sample statistical characteristics.

[0035] (4) Normalize xPI to make the contribution values ​​of each genetic marker comparable across different datasets;

[0036] (5) Apply statistical transformations to xPI to enhance the ability to identify small-effect or marginal-effect genetic markers.

[0037] Furthermore, the relationship between xPI and genetic effect values ​​was specifically evaluated using simulated data:

[0038] (2)

[0039] In this formula, `additive_effect` represents the additive effect, `interaction_effect` represents the interaction effect, and `α` and `β` are the weighting coefficients of the additive and interaction effects, respectively. xPI can accurately quantify the contribution of genetic markers to phenotypes by integrating additive and interaction effects, maintaining consistency across various genetic structures. Using this formula, the `α` and `β` values ​​of any detected gene-gene interaction effect can be calculated. By analyzing these values, the genetic structure of the trait and the specific ways in which genes act on the trait can be clarified.

[0040] Furthermore, a genome-wide significance threshold for xPIs was determined using a substitution test, with a false discovery rate controlled to be less than 0.05. Significant genetic markers were then screened based on this threshold. Combined with other biological data (such as LD decay distance, gene function annotation within specified intervals, and gene expression data), the genes corresponding to these genetic markers and their phenotypic impact were identified, and the influence of genetic loci on phenotypes was further analyzed.

[0041] In a second aspect, the present invention provides a computer device including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor, when executing the computer-readable instructions, implements the fast detection method as described in the first aspect.

[0042] Thirdly, the present invention provides a readable storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the fast detection method as described in the first aspect.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] (1) It can comprehensively detect genetic effects and simultaneously identify additive effects, gene-gene (phenotype) interactions and gene-environment interactions, overcoming the problem of insufficient sensitivity of traditional linear models to non-additive effects.

[0045] (2) Controlling population structure interference: By designing a dedicated loss function for deep neural networks, the population stratification / structure between samples can be effectively corrected, ensuring the reliability of the genetic effect detection results.

[0046] (3) High-precision trait prediction and interpretability: The "interpretable prediction index (xPI)" is proposed to quantify the contribution of each genetic marker to the prediction of the target trait, and can be analyzed into " "Form" enhances the transparency of the model.

[0047] (4) Empirical verification and discovery of new loci: It has demonstrated superior performance in mouse hybrid populations and human disease association studies. It can not only discover known associated loci, but also discover many new loci that were missed by traditional methods but have been confirmed by other studies. The authenticity of gene interaction loci was further verified by statistical interaction test.

[0048] (5) It provides a new framework for the study of nonlinear biological systems, opens up new avenues for analyzing the interactive genetic networks behind complex traits, and empowers subsequent research on biological mechanisms in multi-gene and multi-environmental scenarios. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the xAI-GWAS framework. (a) xAI-GWAS includes the deep neural network GPINN, which models complex relationships between genotype, environmental factors, and phenotype as fixed effects; and controls the confounding effects of population structure as random effects through a novel loss function; it utilizes the interpretable predictive index (xPI) to quantify the contribution of each genetic variant to the prediction, thus achieving model interpretability. (b) In the simulated data, the xPI value is related to… "A strong linear relationship is shown. (c) The network structure of the deep neural network GPINN and its predictive interpretability elucidate the relationship between genetic variants and traits."

[0050] Figure 2 These are the xPI and genetic effect fitting formulas under four genetic frameworks: (a) pure additive framework, (b) pure epistatic interaction framework, (c) mixed additive and epistatic framework, and (d) gene-environment interaction framework. In all scenarios, the xPI values ​​calculated by xAI-GWAS are strongly positively correlated with the absolute value of the effect (r > 0.9), indicating that it can robustly detect additive effects and higher-order genetic effects.

[0051] Figure 3 This is a performance comparison of xAI-GWAS and GEMMA under different genetic architectures. In ae: (a) Additive architecture: Accuracy of xAI-GWAS (blue) and GEMMA (red) at 15 heritability levels (H² = 0.25–0.95). Both methods achieved 90% accuracy at H² = 0.25 and reached the performance cap (100%) when H² ≥ 0.3. (b) Pure phenotypic interaction architecture: xAI-GWAS (blue) consistently outperformed GEMMA (red), and the performance difference was more pronounced at higher heritability levels. (c) Combinatorial additive-interaction architecture: xAI-GWAS (blue) showed a gradual increase in accuracy with increasing heritability (50% at H² = 0.25; 83% at H² = 0.8; 100% at H² ≥ 0.85), while GEMMA (red) was comparable to xAI-GWAS only at a few heritability values ​​(H² = 0.3 / 0.35 / 0.55). The performance gap was most significant at H² = 0.8 (xAI-GWAS: 83% vs. GEMMA: 33%), and GEMMA showed lower accuracy at all heritability levels (H² range from 0.25 to 0.95). (d) Multi-marker purely additive architecture. (e) Multi-marker epistatic interaction architecture.

[0052] Figure 4 xAI-GWAS is used to detect gene-environment interactions. A latent variable following a Gaussian distribution interacts with a specified genetic marker, while two genetic markers with only additive effects are simulated as background. (a) GEMMA can detect the two genetic markers with additive effects, but it cannot detect the genetic marker interacting with the environmental variable, regardless of the individual marker heritability of the interacting genetic marker. (b) xAI-GWAS successfully and accurately locates the genetic marker involved in this environmental interaction.

[0053] Figure 5 This study compares the detection accuracy of xAI-GWAS with GEMMA and XGBoost under different population structure influences: When the population structure influence is weak, both xAI-GWAS and XGBoost outperform GEMMA; when the population structure influence is strong, the detection performance of xAI-GWAS is consistent with that of GEMMA. This indicates that xAI-GWAS provides superior detection performance under weak population structure influences; while under strong population structure influences, xAI-GWAS performs comparably to GEMMA, demonstrating its strong robustness and adaptability in correcting for population structure.

[0054] Figure 6 This section describes the impact of hyperparameters on prediction accuracy and genetic marker detection. (a) Scatter plot of prediction R² versus detection accuracy for 10 causal genetic markers. (b) Scatter plot of prediction R² versus detection accuracy for 20 causal genetic markers. (c) Scatter plot of prediction R² versus detection accuracy for 30 test genetic markers. (d) Line graph of prediction R² during training iterations, comparing overfitting and normal fitting. (e) Line graph of detection accuracy during training iterations, comparing overfitting and normal fitting.

[0055] Figure 7 This is a genome-wide xPI value plot for gene discovery in a heterogeneous mouse dataset. (a) Alkaline phosphatase gene detection results, (b) High-density lipoprotein cholesterol gene detection results, (c) Total cholesterol gene detection results, and (d) Ear cannula area gene detection results. Genes identified by both xAI-GWAS and GEMMA are shown in gray, while genes identified solely by xAI-GWAS are shown in black. Overall, xAI-GWAS detects genes that cover the results of GEMMA and also identifies unique genes not detected by GEMMA in the total cholesterol and ear cannula area traits.

[0056] Figure 8 This is a genome-wide xPI value map for gene discovery in the WTCCC1 human disease dataset. (a) Crohn's disease gene detection results, (b) rheumatoid arthritis gene detection results, (c) type 1 diabetes gene detection results, (d) type 2 diabetes gene detection results. Genes identified by both xAI-GWAS and GEMMA are shown in gray, while genes identified solely by xAI-GWAS are shown in black. Overall, the genes detected by xAI-GWAS cover the results of GEMMA, and unique genes not detected by GEMMA can also be identified in Crohn's disease. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0058] This embodiment provides a genome-wide gene-gene and gene-environment interaction detection method based on an interpretable genetic information neural network, which mainly includes the following steps:

[0059] Step 1: Train a deep neural network model using genotype and phenotypic data to obtain a trained deep neural network model;

[0060] Step 2: Input the genomic data of the sample and use the trained deep neural network model to predict the phenotypic value of the sample;

[0061] Step 3: Calculate the xPI value for each genomic variant site, and use the xPI value to perform genome-wide association analysis (GWAS) and genetic locus detection.

[0062] This method, named xAI-GWAS, is a novel GWAS approach based on deep learning. It aims to accurately predict target traits by constructing a deep neural network model that combines genotype and phenotypic data from samples. Through post-training network analysis, xAI-GWAS not only quantifies the importance of each genotype but also detects additive effects, gene-gene interaction effects, and gene-environment interaction effects.

[0063] 1. Deep Neural Network Architecture

[0064] The core of the xAI-GWAS framework is the Genetic-Prior Informed Neural Network (GPINN), which aims to achieve accurate gene-phenotype prediction through multi-level feature representation generation and aggregation. GPINN effectively captures complex gene-phenotype relationships through multiple independent decision modules and feature transformation mechanisms.

[0065] 1.1 Overview of Model Architecture

[0066] GPINN consists of multiple independent decision modules. Each module receives the raw input features and generates feature representations. Finally, the output module aggregates these feature representations to generate prediction results. Each decision module includes two core components: a feature transformer and a splitter.

[0067] Feature transformer: Responsible for extracting meaningful feature representations from input data and generating multi-dimensional feature representations through a multi-layer architecture. Each layer includes:

[0068] (1) Fully connected layer: performs linear transformation on the input, providing rich feature representations for subsequent nonlinear processing;

[0069] (2) Batch normalization layer: standardizes the input of each layer, stabilizes the training process, and accelerates model convergence;

[0070] (3) Gated linear unit: Feature selection is performed through a gating mechanism to enhance the model's ability to learn complex features;

[0071] (4) Residual connections: Normalized residual connections ensure the stability of variance during network training and avoid gradient vanishing.

[0072] Splitter: Responsible for splitting the feature representation output by the feature converter, selecting only a portion for subsequent aggregation, removing redundant features, and improving model efficiency and accuracy.

[0073] Output module: Aggregates the feature representations from each decision step and processes them through a fully connected layer to finally generate the prediction result.

[0074] 1.2 Advantages of GPINN

[0075] (1) Multi-step decision-making and multi-level feature transformation can accurately capture the complex relationship between genes and phenotypes, and improve the accuracy of trait prediction;

[0076] (2) Batch normalization and residual connection design: enhances model stability and accelerates the convergence process;

[0077] (3) Splitter design: Through an efficient feature selection mechanism, redundant features are avoided from interfering with the model, thus improving computational efficiency.

[0078] 2. Predictive power and interpretability analysis of the model

[0079] To explain the decision-making mechanism of the GPINN model, this invention constructs the xAI-GWAS interpretability evaluation framework and proposes the xPI (eXplainable Prediction Index) metric to quantify the contribution of each genetic marker to the prediction of the target trait. This index is calculated based on the SHapley Additive exPlanations (SHAP) method in game theory, enabling the quantification of feature importance at the single-sample level and obtaining a global ranking of feature importance through aggregation.

[0080]

[0081] in Let be the SHAP of feature Xj in the i-th sample, and N be the total number of samples. Therefore, the xPI metric represents the overall contribution of feature Xj to the model's predictions across the entire dataset. xPI measures the average influence of a feature across the entire dataset by averaging the SHAP values ​​of that feature across all samples. This helps to understand the importance of each feature in the model's predictions. For a trained predictive deep neural network, we can compute the xPI value for each genetic marker across the entire genome, thereby identifying genetic variations that have a significant impact on phenotype. To emphasize its ability to complement the P-value and PRS of GWAS, we name this metric the "Interpretable Predictive Index." This approach has the potential to uncover genetic signals overlooked by traditional GWAS, thereby enhancing our understanding of the genetic structure of complex traits.

[0082] The following variations can be used in xPI calculation:

[0083] (1) To The absolute value of the value is squared to enhance the influence of high-contribution genetic markers;

[0084] (2) To Perform function transformations to measure the contribution of each genetic marker;

[0085] (3) Contribution to each sample ( absolute value Square value The function transformation of the value assigns different weights, and refines the contribution of each sample's xPI based on the characteristics of each sample, such as sample quality and sample statistical characteristics.

[0086] (4) Normalize xPI to make the contribution values ​​of each genetic marker comparable across different datasets;

[0087] (5) Apply statistical transformations to xPI to enhance the ability to identify small-effect or marginal-effect genetic markers.

[0088] To verify the effectiveness of xPI under different genetic architectures, this invention evaluated the relationship between xPI and genetic effect values ​​using simulated data. We simulated various genetic architectures (including pure additive effects, pure interaction effects, and the superposition of additive and interaction effects), and the results show that xPI can effectively capture these genetic effects and maintain a consistent relationship under different architectures:

[0089]

[0090] Wherein, additive_effect is the additive effect, interaction_effect is the interaction effect, and α and β are the weighting coefficients of the additive and interaction effects. xPI can accurately quantify the contribution of genetic markers to phenotypes by integrating additive and interaction effects and maintain consistency across multiple genetic architectures.

[0091] 3. Real-world genetic locus detection methods

[0092] In the process of genetic locus detection, xAI-GWAS uses a trained model to calculate the xPI value of each genomic variant site, and then performs GWAS analysis and genetic locus detection. The following is the unified detection workflow for xAI-GWAS:

[0093] 3.1 Data Input and Preprocessing

[0094] Input quality-controlled genotype matrices and phenotypic data, respectively, from mouse and human datasets (e.g., heterogeneous mouse datasets and WTCCC1 human disease datasets). Ensure that the phenotypic data has been freed from the influence of environmental and other non-genetic factors for more accurate genetic analysis.

[0095] 3.2 Calculate the xPI value

[0096] xAI-GWAS was used to calculate the xPI value for each genetic marker, quantifying the contribution of each marker to the phenotype. To improve the identification of complex genetic signals, both main effects and gene-gene (G×G) interaction terms were incorporated into the model.

[0097] 3.3 Significance Test

[0098] The significance threshold for genetic markers was determined through 1000 permutation tests, with a false discovery rate (FDR) controlled to be less than 0.05. Reliable significant genetic markers were then screened at the significance threshold to ensure that the detected genetic loci were statistically significant.

[0099] Based on the above methods, significant sites that were not detected by traditional methods were detected in mouse and human diseases.

[0100] 4. Theoretical Framework of the GPINN Model

[0101] 4.1 Core Theoretical Foundation: Loss Function Design for Genetic Information Integration and Optimization

[0102] The GPINN model proposed in this invention aims to model the complex relationships between genotype, environment, and phenotype by integrating genetic prior knowledge and constraints. By incorporating genetic prior information into the training process of the neural network, this model can not only learn from data but also optimize and enhance its interpretability. To achieve this goal, this invention designs a novel loss function that effectively integrates genetic information into model training. Specifically, the contribution to quantitative biological traits can be modeled through the following multivariate nonlinear prediction equation:

[0103]

[0104] Where y represents the phenotype, x represents all gene markers, f(.) is a nonlinear multivariate prediction equation, ɛ is the residual, and I is the identity matrix. This represents the residual variance.

[0105] When gene markers are sufficient to represent all genomic variations, the contribution of heredity to quantitative biological traits can be modeled as follows:

[0106] )

[0107] Where Z is the event matrix, usually the identity matrix, and G is the genetic relationship matrix (GRM). This represents the genetic variance. In this way, the genetic constraint can be analogous to the boundary equation in a Physical Information Neural Network (PINN) and integrated into the loss function.

[0108] Therefore, GPINN's loss function consists of two parts: one part is a data-driven term, which forces the neural network to fit the prediction equation; the other part is a gene-related knowledge-driven term, which forces the neural network to satisfy population structure constraints. The loss function can be expressed as:

[0109]

[0110] Data-driven items Used to minimize the difference between predicted and observed values, specifically in the form of:

[0111]

[0112] Where Ny represents the sample size or the degrees of freedom of the prediction equation. To enhance the accuracy of data fitting, GPINN controls data fitting by imposing a more stringent white noise condition, namely:

[0113]

[0114] Gene-related knowledge-driven items Used to enhance the influence of genetic constraints on the model, common forms include:

[0115]

[0116] Where λ is the weighting of the loss term. To replace the distance with a statistical distribution, we choose a genetic constraint of the following form:

[0117]

[0118] 4.2 Joint Integration of Loss Functions

[0119] The two terms above are integrated using a joint statistical distribution to avoid explicitly determining the parameters, and the loss function is rewritten as follows:

[0120]

[0121] Through further derivation, the loss function ultimately manifests as:

[0122]

[0123] Where matrix H represents the result of the genetic correlation matrix (GRM) processing, defined as:

[0124]

[0125] Where G is the genetic correlation matrix, Z is the design matrix, and δ is the variance ratio, calculated using the following formula:

[0126]

[0127] The loss function described above combines genetic information with the data-driven part and embeds genetic constraints into the network training process using a joint statistical distribution, thereby approximating the log-likelihood function in maximum likelihood estimation.

[0128] 4.3 Training through prior genetic knowledge

[0129] By incorporating genetic prior knowledge into the loss function, GPINN training no longer relies solely on data. This design extends physical information learning methods to the field of bioinformatics, effectively integrating biological knowledge into the neural network training process, thereby improving the modeling ability and interpretability of complex biological systems.

[0130] 4.4 Mathematical Equivalence

[0131] It is worth noting that, under the assumption of purely linear effects, the method of this invention is mathematically equivalent to the traditional linear mixture model (LMM). Therefore, this invention not only maintains the advantages of the traditional method, but also makes significant improvements in the flexibility and interpretability of the model.

[0132] The method involved in this embodiment can be applied in the following aspects:

[0133] (1) Analysis of the genetic mechanisms of complex diseases

[0134] xAI-GWAS can be applied to genetic research on complex diseases such as diabetes and cardiovascular diseases. By integrating genomic data and clinical phenotypes from large populations, it systematically identifies disease-related additive genetic effects, gene-gene interactions, and gene-environment interactions. It is particularly suitable for discovering nonlinear genetic effects that are difficult to detect using traditional GWAS methods, providing new clues for disease mechanism research.

[0135] (2) Agricultural breeding and trait improvement

[0136] In crop and livestock breeding, this method can accurately elucidate the genetic basis of important agronomic traits such as yield and stress resistance. By identifying key epistatic effect sites and gene-environment interaction sites, it guides marker-assisted selection and genomic selection, significantly improving breeding efficiency. It is particularly suitable for the genetic analysis of outcrossed populations and complex traits.

[0137] (3) Pharmacogenomics research

[0138] Applied to personalized medicine research, this technology predicts the efficacy and adverse reaction risks of specific drugs by analyzing genetic interaction networks related to drug responses. It can identify drug response-related genetic markers missed by traditional methods, providing a scientific basis for precision medicine.

[0139] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments or technical solutions of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0140] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any way, and all technical solutions obtained by equivalent substitution or other means fall within the scope of protection of the present invention. Parts not covered in this invention are the same as or can be implemented using existing technology.

Claims

1. A method for detecting genome-wide gene-gene and gene-environment interactions based on interpretable genetic information neural networks, characterized in that, Includes the following steps: Step 1: Train a deep neural network model using genotype and phenotypic data to obtain a trained deep neural network model; the deep neural network model is an interpretable deep neural network driven by genetic prior information, and it includes the following components: (1) Multi-step decision module, each decision module is used to receive input features and generate feature representations through feature converter, and finally aggregate the feature representations of all modules through output module to generate final prediction results; (2) Feature converter, used to generate multi-dimensional feature representations through a multi-layer neural network structure; (3) Splitter, used to effectively select features and remove redundant features to improve model efficiency; (4) Output module: aggregates the feature representations of all decision modules and processes them through a fully connected layer to finally generate the prediction results; Step 2: Input the genomic data of the sample and use the trained deep neural network model to predict the phenotypic value of the sample; Step 3: Calculate the xPI value for each genomic variant site, and perform genome-wide association analysis and genetic locus detection using the xPI values; wherein, the method for calculating the xPI value includes: (1) Based on the SHAP method in game theory, the predictive contribution of each genetic marker to the target trait is quantified at the single-sample level; (2) Summarize the SHAP values ​​of all samples, calculate the overall influence of each genetic marker in the dataset, and thus obtain the global feature importance ranking; the xPI value is calculated using the following formula: （1） in Let X be the feature of the i-th sample. j SHAP, where N is the total number of samples; xPI index represents feature X. j The overall contribution to the model's predictions across the entire dataset; Then, the genome-wide significance threshold of xPI was determined by substitution test, the false discovery rate was controlled to be less than 0.05, and significant genetic markers were screened based on the significance threshold. Combined with relevant biological data, the genes corresponding to the genetic markers that affect the phenotype were identified, and the influence of genetic loci on the phenotype was further analyzed.

2. The method for detecting genome-wide gene-gene and gene-environment interactions based on an interpretable genetic information neural network according to claim 1, characterized in that, The feature converter Includes the following sub-modules: Fully connected layers: perform linear transformations on the input, providing rich feature representations for subsequent nonlinear processing; Batch normalization layer: Standardizes the input of each layer, improves training stability, and accelerates convergence; Gated linear units: These units selectively output features through a gating mechanism, thereby enhancing the model's learning ability. Residual connections: ensure the stability of network training and avoid gradient vanishing.

3. The method for detecting genome-wide gene-gene and gene-environment interactions based on an interpretable genetic information neural network according to claim 1, characterized in that, By designing a loss function, genetic prior information is effectively incorporated into model training. This loss function comprises two parts: (1) Data-driven terms: Minimize the difference between model predictions and actual observations to ensure accurate fitting of phenotypic data; (2) Genetic prior knowledge driven terms: By modeling the genetic relationship matrix and combining genetic prior information to optimize the genetic constraints in the training process, the model’s ability to identify complex genetic effects is enhanced.

4. The method for detecting genome-wide gene-gene and gene-environment interactions based on an interpretable genetic information neural network according to claim 1, characterized in that, The following variation is used in xPI calculation: (1) In formula (1), for The absolute value of the value is squared to enhance the influence of high-contribution genetic markers; (2) In formula (1), for Perform function transformations to measure the contribution of each genetic marker; (3) In formula (1), the contribution of each sample is: absolute value Square value The function transformation of the value assigns different weights, and the contribution of xPI for each sample is refined according to the characteristics of each sample; (4) Normalize xPI to make the contribution values ​​of each genetic marker comparable across different datasets; (5) Apply statistical transformations to xPI to enhance the ability to identify small-effect or marginal-effect genetic markers.

5. The method for detecting genome-wide gene-gene and gene-environment interactions based on an interpretable genetic information neural network according to claim 1, characterized in that, The relationship between xPI and genetic effect values ​​was specifically evaluated using simulated data: （2） Wherein, additive_effect is the additive effect, interaction_effect is the interaction effect, α and β are the weighting coefficients of the additive effect and the interaction effect, and xPI can accurately quantify the contribution of genetic markers to phenotypes by combining additive and interaction effects, and maintain consistency under multiple genetic structures; using formula (2) to calculate the α and β values ​​of any detected gene-gene interaction effect, and by analyzing the α and β values, the genetic structure of the trait and the specific way in which genes act on the trait can be clarified.

6. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the computer-readable instructions, it implements the detection method as described in any one of claims 1 to 5.

7. A readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by one or more processors, the one or more processors cause the one or more processors to perform the detection method as described in any one of claims 1 to 5.

Images