A method for interpreting causal genes in crop stress traits based on large models

Abstract

This invention belongs to the field of bioinformatics, specifically a method for interpretable mining of causal genes related to crop stress traits based on a large model. In this method, labels from conserved and exploratory modes are merged with gene expression matrices to construct a unified graph structure learning input matrix. A large language model is used to drive causal structure discovery thought chains to perform causal structure learning. Under the prior space constraints of conserved and exploratory modes, an optimal directed acyclic graph is generated, and the standardized direct causal effect of each candidate gene on the target trait is calculated based on this graph. By triggering the causal mechanism explanation thought chain, functional annotation information and quantitative causal effects in the thought chain reasoning are extracted to infer the regulatory direction, intensity, and potential molecular regulatory pathways of candidate genes, generating biologically meaningful causal explanations. This invention can achieve accurate identification of candidate causal genes related to crop stress traits and provide causal inferences with molecular regulatory mechanism explanation capabilities.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method for interpretable mining of causal genes for crop stress traits based on large models. Background Technology

[0002] Accurately identifying potential candidate causal genes causing crop phenotypic changes under crop stress conditions is one of the major challenges currently facing the fields of crop molecular biology and molecular breeding. With the rapid development of high-throughput sequencing and phenomics technologies, a large amount of transcriptome data and quantitative trait loci (QTLs) of major food crops such as rice and wheat under various stress conditions have been accumulated, providing a rich data foundation for the prediction and mining of candidate causal genes.

[0003] Based on the accumulation of data, a large number of computational methods have been proposed for the prediction of candidate genes related to crop traits.

[0004] Currently, most mainstream crop gene prediction methods still rely primarily on differential expression analysis, genome-wide association studies (GWAS), co-expression networks, and various machine learning models. However, these methods mainly describe the statistical correlation between genes and traits, rather than strict causal relationships. This often results in researchers obtaining a large set of candidate genes, but struggling to accurately identify the target causal genes that truly influence phenotypes. This limitation severely restricts the practical application value of these traditional methods in precision crop breeding design and functional gene validation.

[0005] To overcome the limitations of traditional association analysis, some studies have begun to introduce causal learning methods to predict candidate causal genes. For example, Bayesian network models based on directed acyclic graphs and their conditional dependencies have been applied to rice phenotypic dimensionality reduction and causal inference of maize agronomic traits. However, purely data-driven causal discovery methods are still severely constrained by data quality. In recent years, with the rapid development of large language models, some studies have begun to combine thought chain (CoT) reasoning with causal structure learning methods. These methods utilize the reasoning capabilities of large models to extract prior knowledge in specific domains to assist in causal discovery. For example, traditional causal algorithms are decomposed into a series of reasoning subtasks executed step-by-step by the large language model through CoT.

[0006] In the field of candidate causal gene mining for crop stress traits, mainstream methods mainly rely on differential expression analysis, GWAS, and various machine learning models for predicting candidate causal genes. In addition, some studies have also attempted to introduce causal inference methods such as Bayesian networks to analyze trait associations. However, these existing technologies still have the following drawbacks in practical applications:

[0007] (1) Difficulty in revealing true causal relationships and high false positive rate of candidate genes. Current mainstream crop gene mining methods mainly reveal the statistical correlation between genes and traits, rather than causal relationships. In the complex process of crop stress response, gene regulation is accompanied by a large number of cascade effects. It is difficult to distinguish the target causal gene that truly drives the stress resistance phenotype from downstream pararelated genes by simply relying on statistical correlation. This limitation often leads researchers to obtain a large set of candidate genes with a very high false positive rate, which brings great difficulties to the subsequent expensive and time-consuming functional gene verification and makes it difficult to accurately guide breeding design;

[0008] (2) Pure data-driven algorithms are limited by data quality and sample noise. Although some studies have introduced pure data-driven causal discovery algorithms (such as PC algorithms or traditional Bayesian networks), these methods are extremely dependent on data quality. Crop transcriptome data are usually characterized by "high dimensionality, small sample size" and high technical noise. In the absence of specific domain biological prior knowledge constraints, pure data-driven models are easily interfered with by sample noise, often only generating overly conservative network structures or producing a large number of spurious causal associations. This directly leads to low recall of target causal genes and extremely unstable network inference results;

[0009] (3) Lack of interpretability. Most existing deep learning and complex machine learning models are "black box" architectures, which can usually only output prediction scores but cannot provide transparent explanations of the mechanisms. Crop stress tolerance involves multi-level signal transduction and metabolic reorganization, and biologists need to clarify the physiological pathways through which genes affect phenotypes. The lack of transparent inferences about gene regulatory pathways and mechanisms of action severely limits the scientific credibility of prediction results, making it difficult for researchers to design targeted validation experiments and molecular breeding strategies. Summary of the Invention

[0010] The purpose of this invention is to provide an interpretable method for mining causal genes of crop stress traits based on large models, aiming to solve the technical problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution.

[0012] A method for interpretable mining of causal genes for crop stress traits based on large models includes the following steps:

[0013] Transcriptome samples and functional annotation information related to crop stress were obtained, and a set of trait-associated genes was retrieved based on the functional annotation information. A TPM-normalized gene expression matrix that eliminates batch effects was constructed, and the sample processing conditions were converted into binary trait labels.

[0014] By combining a large language model with a thought chain prompting framework, multi-dimensional causal reasoning is performed on the gene to be queried and the set of genes associated with traits. Genes with complete logical chains or direct evidence are marked as conservative modes, while genes that are supported only in a single dimension are marked as exploratory modes. The labels of conservative and exploratory modes are merged with the gene expression matrix to construct a unified graph structure learning input matrix.

[0015] By driving causal structure learning through a large language model, the parent nodes of candidate directed edges are limited to the conserved mode and exploratory mode gene pools. A BIC-based scoring function is adopted and the structure containing conserved mode genes is given a prior reward weight. The optimal directed acyclic graph is output through greedy search. Then, the standardized direct causal effect of each candidate gene on the target trait is calculated based on the acyclic graph, including the effect size and regulatory direction.

[0016] Functional annotations are extracted from the reasoning chain, and combined with the standardized direct causal effect, a causal explanation containing the direction of regulation, intensity of action, and molecular pathway description is generated using a large language model.

[0017] Furthermore, in multi-dimensional causal reasoning, a binary judgment label and a corresponding concise reasoning basis are output for each reasoning dimension; where:

[0018] For any gene g, the decision label in the k-th dimension is represented as: ;

[0019] In the formula, g represents any gene to be evaluated; k represents the index number of the reasoning dimension; This represents the binary decision label of gene g in the k-th dimension;

[0020] A prior alignment module is introduced to unify and summarize multi-dimensional scattered labels into two coordinated prior modes: conservative mode and exploratory mode. The conservative mode is used to characterize biological causal clues with high confidence, while the exploratory mode is used to characterize genes with potential phenotypic influence clues.

[0021] Furthermore, in multidimensional causal reasoning, the BioCoT (Bio-CoT) cue word framework is applied to perform the following steps:

[0022] Based on the ID of the gene to be queried, obtain the chromosome location and determine whether the gene to be queried co-localizes with a known major QTL region for stress resistance;

[0023] Analyze the GO annotations and functional descriptions of genes to assess whether current genes possess the characteristics of core molecular mechanisms for stress resistance;

[0024] By comparing with the KEGG pathway database, determine whether the current gene is involved in a specific stress signal transduction network;

[0025] Assessing cross-species evolutionary conservation of genes based on orthologous information;

[0026] Search literature databases to determine whether there is experimental evidence of directly altering the target phenotype through genetic intervention;

[0027] The discrimination results of each dimension are output in the form of binary labels, and a structured prior information summary table containing labels of conservative mode or exploratory mode is generated.

[0028] Furthermore, the steps of merging the labels of conserved and exploratory modes with the gene expression matrix to construct a unified graph structure learning input matrix specifically include:

[0029] Querying genes Gene set associated with traits Perform BioCoT inference to generate isomorphic evidence labels and inference patterns;

[0030] Among them, the set of genes associated with traits Multiple independent rounds of BioCoT inference iterations were performed to evaluate the stability of the prior labels, remove labels with conflicting inference results, and extract the output results that remained consistent with high frequency in each round of iterations as the final biological prior information.

[0031] Query genes Gene set associated with traits Merge the genes to build an expanded gene set. The gene expression matrix corresponding to the expanded gene set G is concatenated and fused with the prior information labels extracted by BioCoT to generate a unified expression matrix. Where n is the number of samples. Let G denote the set of real numbers, and let G denote the extended set of genes.

[0032] Furthermore, the step of restricting the parent nodes of candidate directed edges to only come from the conservative and exploratory mode gene pools includes:

[0033] Conserved pattern gene sets with high-confidence causal cues were extracted based on the extended gene set G. , Extracting exploration pattern gene sets with clues to potential trait influence

[0034] Search space constraints were introduced when generating the candidate graph structure, including: limiting that in newly added candidate causal edges, the parent node gene pointing to the target trait must be located in the gene pool jointly defined by the conservation and exploration modes. Inside;

[0035] The set of candidate new edges considered by the large language model in each iteration of the greedy search Represented as: ,in, Let E represent the set of potential causal edges pointing from a gene to a target trait, and let E represent the set of edges in the current graph structure. This represents a candidate directed edge from gene g to the target trait t. This represents a gene pool.

[0036] Furthermore, the steps of employing a BIC-based scoring function and assigning prior reward weights to structures containing conserved pattern genes include:

[0037] By modifying the BIC scoring function to adjust the relative priority of candidate graph structures, structures containing high-confidence biological genes are given a scoring bias.

[0038] The biological high-confidence gene set is the conserved pattern gene set. ;

[0039] The number of high-confidence biological genes in the edge set E that require positive rewards. Represented as:

[0040]

[0041] in, For indicator functions, when The value is 1 if the condition is true, and 0 otherwise. This represents a candidate edge pointing from the conserved pattern gene g to the target trait t;

[0042] The standard BIC scoring function is extended into a composite scoring function that incorporates prior reward terms. , represented as:

[0043]

[0044] in, The BIC score for the graph structure corresponding to the edge set E; These are preset hyperparameters used to adjust the strength of the prior reward weights. Indicates the number of genes.

[0045] Furthermore, the standardized direct causal effect is calculated through the following steps:

[0046] Based on the optimal directed acyclic graph, a logistic regression model is used to calculate the standardized direct causal effect coefficient of each candidate gene toward the target stress trait. The calculation model expression is:

[0047]

[0048] In the formula, Given a gene expression matrix X, the conditional probability that a crop sample exhibits the target stress trait is given. For binary trait labels; Let E represent the set of parent node genes in the edge set E that point to the target trait t; The normalized expression level of any gene i in the set of genes of the parent node; For the intercept term; Let be the logistic regression coefficient of any gene i in the parent node gene set pointing to the target trait t; when the candidate gene g to be evaluated belongs to the parent node gene set, i.e. When, its corresponding coefficient This is the standardized direct causal effect coefficient of candidate gene g pointing to target trait t; using The ratio was calculated. The strength and statistical uncertainty of the causal effect are assessed based on the confidence interval of the ratio; and the standardized direct causal effect coefficient is used as a basis for further analysis. The positive or negative sign of the gene indicates whether the candidate gene g is in a positive or negative regulatory direction on the target trait.

[0049] Furthermore, in the process of driving causal structure learning with a large language model, the Causal Structure Discovery Chain of Thought (CSCoT) cue word framework is introduced to guide the large language model to perform the following steps:

[0050] Load a unified expression matrix X and structured biological prior information containing labels for the conserved mode and the exploratory mode, and perform feature alignment between continuous expression data and discrete prior modes in the context space of the prompt word;

[0051] During the candidate topology generation phase, instructions are used to restrict the starting node of directed edges pointing to the target trait to be located in the candidate gene pool. Inside;

[0052] The large language model is driven to perform a greedy strategy-based graph structure search iteration, utilizing a composite scoring function in each iteration. Evaluate the quality of the graph structure and decide whether to add, delete, or reverse edges.

[0053] When determining the composite scoring function After reaching a local optimum, the optimal directed acyclic graph is locked, and a logistic regression model is used to calculate the standardized direct causal effect coefficient of each candidate gene toward the target stress trait. and ratio ratio .

[0054] Furthermore, the step of generating a causal explanation that includes descriptions of the regulatory direction, intensity of action, and molecular pathway specifically includes:

[0055] From the reasoning basis of the five reasoning dimensions output by BioCoT, extract the key support points of candidate causal genes, and identify the core mechanism module to which the candidate causal genes belong based on the key support points.

[0056] Standardized direct causality coefficients calculated by CSCoT and ratio ratio Logically align and map attributes with the core mechanism modules to establish the relationship between quantitative contributions and qualitative mechanisms;

[0057] By using association relationships to limit the evidence boundaries when generating explanations using large language models, the generated causal explanations are required to be deduced based on the key supporting points and quantitative causal effects, so as to deduce the regulatory pathways of the candidate causal genes on the target stress trait.

[0058] Furthermore, the step of generating a causal explanation that includes descriptions of regulatory direction, intensity of action, and molecular pathway also includes using the causal mechanism explanation thought chain (CECoT) cue word framework to guide the large model to perform the following steps:

[0059] Based on the inference records of BioCoT, the core supporting evidence and reasoning basis of the candidate causal genes under multiple inference dimensions are extracted, and their core mechanism modules are identified.

[0060] The standardized direct causal effect coefficient output by CSCoT and ratio ratio As known quantitative facts, semantic alignment is performed with the core mechanism module, and the large language model is used to assess the causal contribution strength between changes in gene expression levels and trait changes.

[0061] Under the constraints of the core mechanism module, and combined with the crop stress biology knowledge built into the large language model, the specific biological effects of the gene at the molecular and cellular levels are deduced, and the physiological state regulation mechanism mediated by the candidate causal gene in the stress response process is elucidated.

[0062] By deeply integrating the direction and intensity of biological and causal effects, the gene's role in promoting or inhibiting the target trait is determined, and a structured natural language causal explanation is ultimately generated. The causal explanation includes at least quantitative causal contribution indicators, traceable mechanistic clues, and descriptions of molecular regulatory pathways.

[0063] Compared with existing technologies, the beneficial effects of the present invention's method for mining causal genes for crop stress traits based on large-scale models include, but are not limited to:

[0064] This invention utilizes BioCoT to extract multidimensional biological priors and directly embeds them as structural space constraints and scoring correction terms into the optimization process of CSCOT. This deep integration effectively overcomes the shortcomings of pure data-driven algorithms that are susceptible to interference under small sample and high-noise transcriptome data, and significantly improves the recall rate and overall accuracy of candidate causal gene discovery.

[0065] This invention is based on the BIC-driven large model execution of multi-step causal directed acyclic graph construction. Under the strict constraints of exploratory and conservative dual-mode priors, it effectively eliminates pseudo-causal edges induced by noise from limited samples, and achieves accurate identification of target causal genes.

[0066] The CECoT module of this invention not only outputs candidate causal genes, but also performs joint analysis of the quantitative causal effect calculation results in the structure learning stage and the qualitative mechanism clues extracted in the prior stage. The model finally infers the molecular regulatory mechanism explanation of the candidate causal genes, providing a transparent and traceable causal logic chain for the experimental design and targeted verification of downstream candidate causal genes.

[0067] This invention fully leverages the broad-spectrum biological knowledge reserves and powerful general logical reasoning capabilities of large language models across disciplines. Combined with the prior alignment mechanism constructed in this invention, the system can achieve a dynamic optimal balance between known high-confidence knowledge clues and potential new mechanisms without retraining the underlying network for a single crop variety or specific stress trait. Attached Figure Description

[0068] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0069] In the attached diagram:

[0070] Figure 1 This is a flowchart illustrating the implementation of the method for explanatory mining of causal genes for crop stress traits based on a large model, as described in this invention.

[0071] Figure 2 This is a sub-flowchart of the present invention's method for mining causal genes for crop stress traits based on a large model;

[0072] Figure 3 This is a framework diagram of the method for explanatory mining of causal genes for crop stress traits based on a large model, as proposed in this invention. Detailed Implementation

[0073] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0074] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0075] The present invention provides a method for interpretable mining of causal genes for crop stress traits based on a large model. This method uses a biological prior-guided causal thought chain cue framework, taking the reasoning ability of a large language model as the core engine, and integrating biological prior knowledge, causal structure learning, and causal explanation generation into a unified framework. It constructs gene-level biological prior knowledge through retrieval-enhanced generation (RAG), and uses the thought chain to guide the large language model to transform the prior knowledge into direct constraints and scoring corrections for causal structure search. This achieves a deep integration of statistical relevance and biological rationality in transcriptome data-driven causal graph learning.

[0076] The mining method of this invention can not only accurately predict candidate causal genes, but also provide causal evidence at the gene level with biological explanation, solving the problems of traditional association analysis lacking causality and pure data-driven methods being limited by data quality.

[0077] The specific implementation of the embodiments of the present invention will be described below with reference to specific examples.

[0078] Please refer to Figure 1 In one embodiment of the present invention, a method for interpretable mining of causal genes for crop stress traits based on a large model is provided. The specific implementation of this method includes the following steps:

[0079] Step S101: Obtain transcriptome samples and functional annotation information related to crop stress, construct a TPM-normalized gene expression matrix that eliminates batch effects, and convert sample treatment conditions into binary trait labels;

[0080] In one specific implementation, in step S101, for the collection of multi-source stress data, transcriptome data of the target crop are collected from public databases (such as NCBI and CNCB) for specific crop stress tasks (such as rice salt tolerance, rice heat tolerance, and wheat drought resistance). All included datasets must have clearly defined sample data for stress treatment and control treatment. To ensure the interpretability and controllability of the comparison results, sample sets from the same variety and treated within the same experimental batch are preferred to minimize the interference of batch effects and genetic background differences on downstream causal inference.

[0081] Simultaneously, this invention utilizes stress trait keywords to retrieve and obtain gene function annotation information for the target species from relevant functional databases (such as RAP-DB and WheatIS). Based on this functional annotation information, this invention retrieves a set of trait-associated genes. .

[0082] Furthermore, the present invention preprocesses the collected transcriptome data;

[0083] Specifically, in the preprocessing, genes with expression levels close to zero in the vast majority of samples are first removed. This step aims to reduce technical noise and redundant dimensionality, while improving the stability of subsequent modeling. The set of transcriptome genes retained after filtering is defined as... Furthermore, a gene expression matrix that eliminates batch effects is constructed. For datasets that only provide expression matrices, a unified renormalization process is implemented based on their original standardized results to reduce expression distribution shifts caused by differences in processing procedures and enhance comparability across datasets.

[0084] Based on this, the present invention uniformly adopts TPM as the representation format for gene expression matrices;

[0085] Furthermore, the traits (i.e. treatment conditions) of the target stress samples and control samples were converted into binary trait labels. This forms a standardized input consisting of an expression matrix and its corresponding trait labels.

[0086] During the collection of multi-source data, this invention also simultaneously introduced and integrated literature data to construct a literature database. Specifically, the system uses the species name of the target crop (e.g., rice, wheat) and the corresponding abiotic stress traits (e.g., salt stress, drought stress) as core search keywords to conduct full-text or abstract searches in public medical and biological literature databases such as PubMed, comprehensively acquiring academic literature related to the target phenotype and gene genetic interventions (e.g., knockout, overexpression experiments). The acquired literature is then subjected to textual structured parsing and cleaning to extract gene regulatory logical relationships with direct experimental evidence, thereby completing the establishment of the literature database.

[0087] This invention further defines the trait-associated gene set and the query gene. This invention retrieves the initial trait-associated gene set from relevant functional databases (such as RAP-DB and WheatIS) using characteristic keywords. .

[0088] Furthermore, this invention further selects and extracts associated genes of relevant traits from databases such as GWAS Atlas, and constructs a GWAS associated gene set independent of the functional annotation distribution after standardization mapping, which is used for external validation of the model's generalization ability across evidence sources.

[0089] Furthermore, trait-related gene sets This is used to represent the set of trait-associated genes selected from transcriptome data that are related to the target stress trait. Based on this, an associated gene expression matrix is ​​constructed. Where n represents the number of samples. Represents the set of real numbers;

[0090] Meanwhile, the gene to be queried is input into the model. .

[0091] The present invention aims to identify whether a query gene is a candidate causal gene for a target trait under a given stress condition, and to output its corresponding mechanism explanation.

[0092] Please continue to refer to Figure 1 The present invention, based on a large-scale model, provides a method for the explanatory mining of causal genes for crop stress traits, and further includes the following steps:

[0093] Step S102: Using a large language model combined with a thought chain prompting framework, perform multi-dimensional causal reasoning on the gene to be queried and the known associated gene set. Genes with complete logical chains or direct evidence are marked as conservative modes, while those supported only in a single dimension are marked as exploratory modes.

[0094] Since crop stress traits usually involve multi-level signal transduction and regulatory network reorganization, it is difficult to fully capture their complex causal mechanisms by relying on a single data source. In step S102 of this invention, a large language model combined with RAG technology is used to transform external biological knowledge into structured prior constraints.

[0095] This invention performs automated bioinformatics retrieval on the input gene to be queried;

[0096] Specifically, using gene IDs as the primary index, the API interface of functional databases (such as RAP-DB or WheatIS) is invoked to obtain core annotation information, including gene names, chromosome locations, functional descriptions, and GO functional annotations. Subsequently, the retrieved results are aggregated, deduplicated, and standardized to construct a hierarchical gene information file. This process provides unified and easily parsed contextual input materials for subsequent reasoning in the large language model.

[0097] Based on the structured retrieval information described above, this invention constructs the BioCoT cue word framework to drive a large language model to perform multi-dimensional causal reasoning evaluation of genes. In this multi-dimensional causal reasoning, a binary judgment label and a corresponding concise reasoning basis are output for each reasoning dimension:

[0098] For any gene g, the decision label in the k-th dimension is represented as:

[0099] ;

[0100] In the formula, g represents any gene to be evaluated; k represents the index number of the reasoning dimension; This represents the binary decision label of gene g in the k-th dimension.

[0101] Specifically, the large language model performs a step-by-step chain-like inference evaluation around the index number k in the following five dimensions:

[0102] When k=1, it represents the QTL co-localization dimension, which is used to match gene locations with known major QTL intervals, providing preliminary screening of genetic background at the whole genome level;

[0103] When k=2, it represents the molecular mechanism relevance dimension, which is used to integrate GO annotation and functional description to assess whether a gene is involved in a core molecular mechanism.

[0104] When k=3, it represents the signaling pathway participation dimension, which is used to combine the KEGG pathway database to assess whether gene g is involved in the signal transduction pathway of stress response;

[0105] When k=4, it represents the homology conservation dimension, which is used to assess the evolutionary conservation and functional transferability of genes under different genetic backgrounds and environmental conditions by utilizing cross-species orthologous relationships and functional annotations.

[0106] When k=5, it represents the literature experimental evidence dimension, which is used to search the literature database to assess whether there is literature evidence of direct genetic intervention experiments altering the target phenotype.

[0107] To achieve the optimal balance between precision and recall in gene prediction, this invention introduces a priori alignment module, which unifies and summarizes multi-dimensional, scattered labels into two coordinated priori modes: conserved mode and exploratory mode; wherein:

[0108] Conservative patterns are used to characterize biological causal cues with high confidence; the indicator function of a conservative pattern. Represented as:

[0109]

[0110] In the formula, A conserved pattern indicator function for gene g; This indicates an indicator function that takes the value 1 when the mathematical condition within the parentheses is true, and 0 otherwise. This represents the function that takes the maximum value; here it is used to characterize the logical "OR" relationship between two variables. This represents the label used to determine gene g in the dimension of experimental evidence from the literature (i.e., k=5). This represents the minimum value of the label for gene g in the first 4 dimensions (k ranges from 1 to 4), which is used here to characterize the logical "AND" relationship of binary variables;

[0111] This formula indicates that when gene g has direct literature and experimental evidence to support it, or simultaneously meets the QTL colocalization dimension, molecular mechanism relevance dimension, signaling pathway participation dimension, and homology conservation dimension, then gene g is determined to be a conserved pattern gene.

[0112] Furthermore, the exploration pattern is used to characterize genes with potential phenotypic cues, and the indicator function of the exploration pattern... Represented as:

[0113]

[0114] In the formula, A function indicating the exploration mode of gene g; This indicates an indicator function that takes the value 1 when the mathematical condition within the parentheses is true, and 0 otherwise. This represents the sum of the judgment labels for gene g in the first four dimensions (i.e., k=1,2,3,4). This formula indicates that when gene g obtains positive support in at least one dimension of QTL colocalization, molecular mechanism relevance, signaling pathway participation, and homology conservation, an exploration mode label is assigned to gene g.

[0115] Please refer to Figure 2 In one embodiment of the present invention, in the thought chain reasoning of multi-dimensional causal evidence, the BioCoT cue word framework is used to guide the large language model to perform the following steps:

[0116] Step S201: Read the ID and chromosome location of the gene to be queried, and determine whether the gene co-localizes with a known major QTL region for stress resistance;

[0117] Step S202: Analyze the GO annotation and functional description of the gene to assess whether the current gene possesses the characteristics of the core molecular mechanism of stress resistance; and compare it with the KEGG pathway database to determine whether the current gene participates in a specific stress signal transduction network;

[0118] Step S203: Assess the cross-species evolutionary conservation of genes based on orthologous information, and search the literature database to determine whether there is experimental evidence of directly altering the target phenotype through genetic intervention;

[0119] Step S204: Integrate the basis of the above reasoning steps, output the discrimination results of each dimension in the form of binary labels, and generate a structured prior information summary table containing the conservative mode or exploratory mode labels.

[0120] This invention combines large language models with RAG technology to automatically retrieve gene annotation information and guide the model to perform chain reasoning around five dimensions: QTL colocalization, molecular mechanisms, signaling pathways, homology conservation, and literature and experimental evidence.

[0121] Building upon this foundation, this invention introduces a priori alignment module, which divides multidimensional reasoning evidence into "conservative modes" with high confidence and "exploratory modes" with potential causal clues. This mechanism overcomes the limitations of traditional methods that treat large models merely as external annotation tools, successfully transforming unstructured domain biological knowledge into structured prior constraint labels that can be directly used for algorithmic computation, thus providing a solid biological foundation for causal reasoning.

[0122] Please continue to refer to Figure 1 In a specific implementation of the present invention, the mining method provided in this embodiment further includes the following steps:

[0123] Step S103: Merge the labels of the conserved mode and the exploratory mode with the gene expression matrix to construct a unified graph structure learning input matrix;

[0124] Step S103 in this embodiment aims to execute a multi-source feature fusion strategy, which performs multivariate joint inference on the gene to be queried and its trait-related gene set to reduce the statistical bias caused by single gene analysis.

[0125] The specific construction process includes the following steps:

[0126] First, since the prior information of trait-associated genes determines the skeleton of the causal graph and directly constrains the candidate space for subsequent structure searches, this invention addresses the query genes... Gene set associated with traits Perform the same BioCoT steps to generate isomorphic evidence labels and inference patterns.

[0127] Secondly, to improve the reliability of the prior, this invention provides the trait-associated gene set... A priori tag stability assessment mechanism was introduced. Specifically, for the set of genes associated with the aforementioned trait... Perform multiple independent rounds of BioCoT inference iterations; during the iteration evaluation, actively remove labels with discrepancies in the inference results of the large language model, and extract the output results that maintain high frequency consistency in each iteration as the final biological prior information;

[0128] Subsequently, the queried gene Trait-associated gene set screened from transcriptome data Merge the genes to build an expanded gene set. ;

[0129] Finally, the expanded gene set The corresponding gene expression matrix is ​​then combined with the prior information labels extracted by BioCoT through feature concatenation and deep fusion to generate a unified expression matrix. Where n is the number of samples. Represents the set of real numbers;

[0130] This operation successfully unified knowledge-driven discrete prior labels with data-driven continuous representation data, constructing a unified graph structure learning input matrix, and providing the downstream CSCOT module with query-oriented genes. Standardized causal modeling input.

[0131] This invention utilizes thought chain to extract multidimensional biological priors and directly embeds them as structural space constraints and scoring correction terms into the optimization process of CSCOT. This deep integration effectively overcomes the shortcomings of pure data-driven algorithms that are easily interfered with in small sample and high-noise transcriptome data, and significantly improves the recall rate and overall accuracy of candidate causal gene discovery.

[0132] Please continue to refer to Figure 1 The mining method provided in this embodiment also includes the following steps:

[0133] Step S104: Drive causal structure learning with a large language model, limit the parent nodes of candidate directed edges to only come from the conserved mode and exploratory mode gene pools, adopt a scoring function based on Bayesian information criterion (BIC) and assign prior reward weights to structures containing conserved mode genes, output the optimal directed acyclic graph through greedy search, and then calculate the standardized direct causal effect of each candidate gene on the target trait based on the acyclic graph, including the effect size and regulatory direction.

[0134] Specifically, CSCOT performs causal structure learning driven by a large model language, decomposing the traditional BIC-based graph search process into interpretable multi-step decision-making. To preserve biological plausibility while enabling the exploration of new mechanisms, this invention introduces prior guidance through the following mathematical mechanism:

[0135] First, based on the unified pattern classification results output in step S102, this invention extracts a conserved pattern gene set with high-confidence causal clues based on the extended gene set G. , represented as: Simultaneously, we extracted the gene set of exploration patterns that contained clues to potential trait influences. , represented as: ;

[0136] To prevent causal structure learning from deviating from the known mechanism framework due to pure reliance on correlation signals, this invention introduces search space constraints when generating candidate graph structures, including: in newly added candidate causal edges, the parent node gene pointing to the trait must be limited to a gene pool jointly defined by conservatism and exploration modes. Therefore, the large model considers the set of candidate new edges in each step of the greedy search. Represented as:

[0137]

[0138] in, Let E represent the set of potential causal edges pointing from a gene to a target trait, and let E represent the set of edges in the current graph structure. Let represent candidate directed edges from gene g to target trait t. This constraint mechanism strictly limits the structure search to a subspace that satisfies biological prior feasibility, significantly reducing the probability of noise-induced pseudo-causal edges entering the optimal structure.

[0139] Based on the limited search space, this invention further adjusts the relative priority of candidate graph structures by modifying the BIC scoring function, so that structures containing high-confidence biological genes are positively biased.

[0140] The biologically high-confidence gene set is defined as a conserved pattern gene set. The number of high-confidence biological genes in the edge set E that need to be assigned positive rewards. Represented as:

[0141]

[0142] in, This is an indicator function that takes the value 1 when the mathematical condition within the parentheses is true, and 0 otherwise. This represents a candidate edge pointing from the conserved pattern gene g to the target trait t;

[0143] Based on this, the present invention extends the standard BIC scoring function into a composite scoring function that incorporates prior reward terms. , represented as:

[0144]

[0145] in, The BIC score for the graph structure corresponding to the edge set E; These are preset hyperparameters used to adjust the strength of the prior reward weights.

[0146] Compared to traditional learning methods that rely solely on BIC(E), the correction function in this embodiment innovatively embeds a structured biological prior reward term, which makes candidate edges with biological rationality obtain higher composite scores in the iterative selection of CSCOT and are more likely to be retained, significantly enhancing the algorithm's ability to identify real causal structures.

[0147] Furthermore, after completing the causal structure search of the optimal directed acyclic graph and determining the network topology based on CSCOT, the standardized direct causal effect is calculated through the following steps:

[0148] Based on the parent node set determined by the optimal directed acyclic graph, the standardized direct causal effect coefficient of each candidate gene pointing to the target stress trait is calculated using a logistic regression model. Its computational model expression is:

[0149]

[0150] In the formula, Given a gene expression matrix X, the conditional probability that a crop sample exhibits the target stress trait is given. For binary trait labels; Let E represent the set of parent node genes in the edge set E that point to the target trait t; The normalized expression level of any gene i in the set of genes of the parent node; For the intercept term; Let be the logistic regression coefficient of any gene i in the parent node gene set pointing to the target trait t; when the candidate gene g to be evaluated belongs to the parent node gene set, i.e. When, its corresponding coefficient This is the standardized direct causal effect coefficient of candidate gene g pointing to target trait t; using The ratio was calculated. The strength and statistical uncertainty of the causal effect are assessed based on the confidence interval of the ratio; and the standardized direct causal effect coefficient is used as a basis for further analysis. The positive or negative sign of the gene indicates whether the candidate gene g is in a positive or negative regulatory direction on the target trait.

[0151] This mechanism of transforming network topology into specific effect values ​​not only provides solid statistical support for identifying gene regulatory phenotypes, but also lays a quantitative foundation for the subsequent interpretation of highly interpretable biological mechanisms generated by the CECot module and for downstream targeted experimental validation.

[0152] Specifically, after completing the causal structure search of the optimal directed acyclic graph and determining the network topology based on CSCOT, this invention further evaluates the causal direction and quantitative effect of candidate causal genes on the target trait; and calculates the standardized direct effect of candidate genes on the target stress trait based on the structural parameters of the optimal directed acyclic graph. Effect estimates are generated in the causal structure learning module (i.e.,...). Based on the coefficients, these estimated values ​​are indexed to calculate the corresponding odds ratios. This calculation process not only quantifies the magnitude of the change in the log-odds ratio of traits caused by changes in gene expression, but also accurately assesses the strength of the causal effect and its statistical uncertainty by combining confidence intervals. This assessment mechanism, which transforms network topology into specific effect values, not only distinguishes whether a gene promotes or inhibits a specific phenotype, but also lays an indispensable quantitative foundation for the highly interpretable causal explanation of the biological mechanisms generated by the CECot module.

[0153] Furthermore, in the process of driving causal structure learning with a large language model, the CSCOT prompt word framework is introduced to guide the large language model to perform the following steps:

[0154] Load the unified expression matrix X and the structured biological prior information containing the labels of the conservative mode and the exploratory mode, and complete the feature alignment of continuous expression data and discrete prior modes in the context space of the prompt word;

[0155] During the candidate topology generation phase, instructions are used to restrict the starting node of directed edges pointing to the target trait to be located in the candidate gene pool. Inside;

[0156] The large language model is driven to perform a greedy strategy-based graph structure search iteration, and in each iteration, the composite scoring function is utilized. Evaluate the quality of the graph structure and decide whether to add, delete, or reverse edges.

[0157] When determining the composite scoring function After reaching a local optimum, the optimal directed acyclic graph is locked, and a logistic regression model is used to calculate the standardized direct causal effect coefficient of each candidate gene toward the target stress trait. and ratio ratio .

[0158] As can be seen, this invention, based on BIC-driven large model execution of multi-step causal directed acyclic graph construction, effectively eliminates pseudo-causal edges induced by noise from limited samples under the strict constraints of exploratory and conservative dual-mode priors, and achieves accurate identification of target causal genes.

[0159] The breakthrough of this invention lies in its substantial algorithmic extension of the classic BIC, constructing a novel composite scoring criterion that integrates biological priors. Its core lies in two prior guidance mechanisms:

[0160] First, there is a search space constraint, which strictly stipulates that when adding causal edges pointing to traits, the candidate parent node must come from the candidate causal gene set of the conservative mode or the exploratory mode, thereby eliminating pseudo causal edges that are affected by sample noise.

[0161] Secondly, the scoring function is reconstructed, overcoming the limitations of standard BIC which relies solely on data likelihood and network complexity. It innovatively embeds a structured biological prior reward term into the scoring objective function. This improved mechanism enables large models to explicitly numerically optimize network topologies that positively favor high-confidence biological mechanisms when driving structure search. This dual-constraint mechanism based on improved BIC effectively solves the technical challenge of pure data-driven methods easily deviating from known biological frameworks and generating numerous spurious causal associations in small-sample, high-dimensional transcriptome data.

[0162] In addition to the BIC-based scoring search method mainly used in this invention, this framework also supports the integration of other mainstream causal discovery algorithms to predict candidate causal genes, such as using constraint-based PC or FCI algorithms, using other scoring-based strategies such as GES, or introducing the NOTEARS algorithm based on continuous optimization. After combining the prior constraint mechanism of this invention, the above algorithms can all effectively achieve the same inventive purpose.

[0163] Please continue to refer to Figure 1 The mining method provided in this embodiment also includes the following steps:

[0164] Step S105: Extract functional annotations from the thought chain reasoning, combine them with the standardized direct causal effect, and use a large language model to generate a causal explanation that includes the direction of regulation, intensity of action, and molecular pathway description.

[0165] In one implementation of the present invention, the step of generating a causal explanation specifically includes:

[0166] From the reasoning evidence of the five inference dimensions output by BioCoT, key support points for candidate causal genes are extracted, and the core mechanism module to which the candidate causal genes belong is identified based on these key support points; the standardized direct causal effect coefficients calculated by the CSCOT module are then used to... and ratio ratio Logically align and map attributes with the core mechanism modules to establish the relationship between quantitative contributions and qualitative mechanisms;

[0167] This invention utilizes association relationships to limit the evidence boundaries when generating explanations using large language models, and mandates that the generated causal explanations must be deduced based on the key supporting points and quantitative causal effects, in order to deduce the regulatory pathways of the candidate causal genes on the target stress trait.

[0168] Furthermore, the step of generating a causal explanation also includes using the CECoT cue word framework to guide the large model to perform the following steps, specifically including: extracting the core supporting evidence and reasoning basis of the candidate causal gene under multiple reasoning dimensions based on the BioCoT inference record, and identifying its core mechanism module; and converting the standardized direct causal effect coefficient output by CECoT into a standard direct causal effect coefficient. and ratio ratio As known quantitative facts, semantic alignment is performed with the core mechanism module, and the causal contribution strength between changes in gene expression levels and trait changes is assessed using the large language model. Under the constraints of the core mechanism module, combined with the crop stress biology knowledge built into the large language model, the specific biological effects of the gene at the molecular and cellular levels are deduced, and the physiological state regulation mechanism mediated by the candidate causal gene in the stress response is elucidated. The direction and intensity of biological effects and causal effects are deeply integrated to determine whether the gene promotes or inhibits the target trait, and finally a structured natural language causal explanation is generated. The causal explanation includes at least quantitative causal contribution indicators, traceable mechanism clues, and descriptions of molecular regulatory pathways.

[0169] Step S105 of this embodiment aims to break the "black box" limitation of traditional machine learning models in crop gene prediction and proposes the gene-level interpretable CECoT framework. This module uses a large language model as an inference engine to jointly analyze the quantitative causal effects calculated in the structure learning stage and the qualitative mechanism clues extracted in the BioCoT stage, thereby deriving an explanation that is both consistent with statistical evidence and has biological mechanism rationality.

[0170] To reduce interference from irrelevant information and mitigate potential hallucinations that large language models may generate when generating explanations, this invention establishes clear evidence boundaries at the starting point of inference. Specifically, it extracts key supporting points for candidate causal genes across various dimensions from the inferential evidence output by BioCoT, identifying the core mechanism module most relevant to that gene. Subsequently, it logically aligns and maps the standardized direct causal effects calculated by CSCOT with the core mechanism module to limit the evidence boundaries when the large language model generates explanations.

[0171] After establishing the constraints of the core mechanism modules, the large-scale model further infers the biological effects mediated by the target genes, elucidating the specific regulatory mechanisms of these genes on physiological states and phenotypic expressions during crop stress responses. Finally, this invention deeply integrates the inferred biological effects with causal effects to generate a standardized causal biological explanation. This causal explanation not only covers the quantitative causal contribution of genes to traits but also traces their molecular regulatory pathways at the cellular and systemic levels, providing a reliable causal mechanism explanation with a transparent chain of evidence for targeted experimental verification of downstream candidate causal genes and molecular breeding design.

[0172] To translate statistical results into biological language, CECot's prompt-word instruction big language model performs causal explanatory reasoning that intersects qualitative and quantitative approaches; its thought process includes:

[0173] Evidence boundary constraints: The prompt words force the model to first extract the core support points in the BioCoT inference record, delineate the core mechanism module to which the candidate causal genes belong, and prevent the model from diverging and producing "illusions".

[0174] Quantitative and qualitative integration: The indicator model uses the quantitative causal effects (regulatory strength and positive / negative direction) output by CSCOT as known facts to match and explain the above mechanism modules;

[0175] Molecular pathway deduction: The model is required to further deduce the stress-resistance physiological pathway mediated by the gene and finally output a structured natural language explanation that clearly explains how the candidate causal gene promotes or inhibits the target phenotype through the cellular and systemic regulatory networks.

[0176] Therefore, the CECoT module of this invention not only outputs candidate causal genes, but also performs joint analysis of quantitative causal effect estimation in the structure learning stage and qualitative mechanism clues extracted in the prior stage. The model eventually automatically generates gene-level causal explanations, providing a transparent and traceable causal logic chain for experimental design and targeted verification of downstream candidate causal genes.

[0177] Most existing predictive models are black boxes lacking interpretability. This invention proposes CECoT, a gene-level interpretable mechanism. This mechanism extracts path coefficients and probability ratios calculated during the causal structure learning phase as quantitative causal effect indicators of the direction and intensity of a gene's influence on a target trait. Subsequently, this quantitative data, along with qualitative mechanistic evidence (such as functional annotations and signaling pathway regulatory nodes) generated during the BioCoT phase, is input into the CECoT module for systematic comparison and inference. Finally, a natural language explanation containing quantitative causal contributions and traceable mechanistic clues is automatically generated. This mechanism can not only determine the specific regulatory direction of candidate causal genes in promoting or inhibiting traits, but also clarify their role in specific molecular pathways, providing a reliable causal logic chain and decision support for the experimental verification of downstream candidate causal genes and precision breeding design.

[0178] In summary, this invention fully leverages the broad-spectrum biological knowledge reserves and powerful general logical reasoning capabilities of large language models across disciplines. Combined with the prior alignment mechanism constructed in this invention, the system can achieve a dynamic optimal balance between known high-confidence knowledge clues and potential new mechanisms without retraining the underlying network for a single crop variety or specific stress trait.

[0179] Validation results show that, thanks to the excellent versatility of the large model, the invented mining method maintains strong generalization ability and significant performance advantages in multiple independent and highly different task scenarios, such as rice salt tolerance, rice heat tolerance, and cross-species wheat drought resistance. This demonstrates that the framework can highly adapt to the complex and diverse needs of crop stress resistance candidate causal gene network mining.

[0180] To further verify the practical technical effectiveness, independent contributions of each core module, and robustness under different large language model foundations of the crop stress trait causal gene interpretability mining method based on large-scale models described in this invention, the embodiments of this invention conducted systematic benchmark tests and ablation experiments based on large-scale public datasets in the real world. The construction of the test datasets covered the rice salt-tolerant transcriptome dataset (PRJNA1010833), the rice heat-tolerant transcriptome dataset (PRJNA301554), and the wheat drought-resistant transcriptome dataset (PRJCA025815) for cross-species generalization ability verification.

[0181] In the experimental evaluation, this invention constructed a total of 6 independent candidate related gene sets for the functional annotation database (RAP-DB, WheatIS) retrieval set and GWAS feature set, and used genes mentioned in the literature that have been verified by direct functional perturbation experiments as standard causal truth labels.

[0182] Table 1. Verification benchmark dataset and truth label distribution table of the embodiments of the present invention

[0183]

[0184] Table 1 shows the specific distribution structure of the benchmark datasets and ground truth labels used in the verification of the embodiments of the present invention. In the benchmark method comparison and ablation evaluation experiments, the method of the present invention was compared with two mainstream benchmark methods and three ablation configurations in a unified quantitative manner. The first category is knowledge-driven large model benchmark methods, including the Zero-shot CoT method, the Few-shot CoT method, and the conventional large model causal hint method (CGI-GPT); the second category is pure data-driven causal structure learning algorithms, including the PC algorithm, the MMHC algorithm, and the NOTEARS algorithm. Ablation configurations include removing the biological prior cognition module, removing the causal structure discovery cognition module, and removing both modules simultaneously. Each method outputs candidate causal gene prediction results under a unified standardized input and evaluation protocol, and the mining performance is comprehensively evaluated by calculating precision, recall, accuracy, and F1 score.

[0185] Table 2. Performance Comparison Test Data of Different Algorithms under the Main Stress Tasks of this Invention

[0186]

[0187] Table 2 shows the comparative test data of the method of this invention in the core task scenarios. Analysis of the experimental data and test results in Table 2 shows that the method of this invention significantly outperforms traditional knowledge-driven benchmark methods and purely data-driven causal structure learning algorithms in identifying candidate causal genes across all test tasks. Particularly in the two core metrics of recall and F1 score, the method of this invention achieves the most significant performance gains, successfully improving the completeness of capturing true phenotypic regulatory causal genes while maintaining high precision. This performance advantage remains highly stable in both functional database and GWAS association dataset scenarios. This demonstrates the unique innovative advantage of this invention, which deeply integrates biological prior knowledge with data-driven causal graph learning, effectively capturing causal targets with true phenotypic regulatory effects from the whole genome or candidate pool.

[0188] Furthermore, this invention also conducted comparative tests on the robustness and compatibility of different underlying large language model foundations. The multi-step decision-making framework of this invention was deployed in models with different architectures, including GPT-5, GPT-4.1, GPT-4o, Gemini-3, DeepSeek-R1, and Qwen-3. Experimental results show that, thanks to the decoupling of complex graph structure search and its transformation into a verifiable and output-constrained thought chain subtask, all models exhibit a highly consistent performance gain trend under the unified framework. Among them, the advanced general-purpose large model GPT-5 achieved the best overall performance, with its median F1 score in the rice salt tolerance task being 18.0% higher than that of the earlier version GPT-4.1.

[0189] In the case study of biological function enrichment and causal mechanism interpretability verification, the candidate genes for rice salt tolerance identified by the method of this invention were subjected to GO biological process and KEGG pathway enrichment analysis. The results showed that the predicted genes were precisely enriched in core stress resistance cascade regulatory networks such as salt stress response, potassium ion transport, osmotic stress response, plant hormone signal transduction, and MAPK signaling pathway. Meanwhile, the candidate genes for rice heat tolerance were significantly enriched in key protective mechanisms maintaining cellular protein homeostasis, such as heat response, cellular heat response, and protein folding and refolding, demonstrating extremely high biological rationality. In a specific application embodiment of this invention, the candidate causal gene OS08G0558200 (OsGRX20), which ranked high in the system output, was used as a specific object for causal explanation chain tracing.

[0190] Based on locking the optimal directed acyclic graph, the system of this invention accurately calculates the standardized direct causal effect coefficient of gene expression level pointing to the target salt tolerance phenotype using a logistic regression model. =4.103. Subsequently, the CECot prompt framework was triggered, treating this quantitative statistical effect value as a known fact. Semantic alignment and logical reasoning were then performed with the qualitative functional mechanisms of the gene, such as reactive oxygen species scavenging, extracted in the previous stage. This automatically generated a natural language causal explanation containing quantitative contribution indicators, traceable mechanism clues, and descriptions of molecular regulatory pathways. This automatically generated structured explanation clearly elucidates the cellular and systemic regulatory networks through which the candidate gene positively promotes crop salt tolerance phenotypes, highly consistent with known biological facts. This fully demonstrates the significant application value of the method of this invention in practical precision breeding target screening and downstream experimental design.

[0191] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for interpretable mining of causal genes for crop stress traits based on large-scale models, characterized in that, Includes the following steps: Transcriptome samples and functional annotation information related to crop stress were obtained, and a set of trait-associated genes was retrieved based on the functional annotation information. A TPM-normalized gene expression matrix that eliminates batch effects was constructed, and the sample processing conditions were converted into binary trait labels. By combining a large language model with a thought chain prompting framework, multi-dimensional causal reasoning is performed on the gene to be queried and the set of genes associated with traits. Genes with complete logical chains or direct evidence are marked as conservative modes, while genes that are supported only in a single dimension are marked as exploratory modes. The labels of conservative and exploratory modes are merged with the gene expression matrix to construct a unified graph structure learning input matrix. By driving causal structure learning through a large language model, the parent nodes of candidate directed edges are limited to the conserved mode and exploratory mode gene pools. A BIC-based scoring function is adopted and the structure containing conserved mode genes is given a prior reward weight. The optimal directed acyclic graph is output through greedy search. Then, the standardized direct causal effect of each candidate gene on the target trait is calculated based on the acyclic graph, including the effect size and regulatory direction. Functional annotations are extracted from the reasoning chain, and combined with the standardized direct causal effect, a causal explanation containing the direction of regulation, intensity of action, and molecular pathway description is generated using a large language model.

2. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 1, characterized in that, In multi-dimensional causal reasoning, a binary judgment label and a corresponding concise reasoning basis are output for each reasoning dimension; where: For any gene g, the decision label in the k-th dimension is represented as: ； In the formula, g represents any gene to be evaluated; k represents the index number of the reasoning dimension; This represents the binary decision label of gene g in the k-th dimension; A prior alignment module is introduced to unify and summarize multi-dimensional scattered labels into two coordinated prior modes: conservative mode and exploratory mode. The conservative mode is used to characterize biological causal clues with high confidence, while the exploratory mode is used to characterize genes with potential phenotypic influence clues.

3. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 2, characterized in that, In multidimensional causal reasoning, BioCoT is applied to perform the following steps: Based on the ID of the gene to be queried, obtain the chromosome location and determine whether the gene to be queried co-localizes with a known major QTL region for stress resistance; We analyze the GO annotations and functional descriptions of genes to assess whether the current gene possesses the characteristics of a core molecular mechanism for stress resistance; and compare it with the KEGG pathway database to determine whether the current gene participates in a specific stress signal transduction network. Based on orthologous information, we assess the cross-species evolutionary conservation of genes and search literature databases to determine whether there is experimental evidence of directly altering the target phenotype through genetic intervention. The discrimination results of each dimension are output in the form of binary labels, and a structured prior information summary table containing labels of conservative mode or exploratory mode is generated.

4. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 3, characterized in that, The steps involved in merging the labels of conserved and exploratory modes with the gene expression matrix to construct a unified graph structure learning input matrix include: Query genes Gene set associated with traits Perform BioCoT inference to generate isomorphic evidence labels and inference patterns; Among them, the set of genes associated with traits Multiple independent rounds of BioCoT inference iterations were performed to evaluate the stability of the prior labels, remove labels with conflicting inference results, and extract the output results that remained consistent with high frequency in each round of iterations as the final biological prior information. Query genes Gene set associated with traits Merge the genes to build an expanded gene set. The gene expression matrix corresponding to the expanded gene set G is spliced ​​and fused with the prior information labels extracted by BioCoT to generate a unified expression matrix X.

5. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 4, characterized in that, The steps to restrict the parent nodes of candidate directed edges to only come from the conservative and exploratory gene pools include: Conserved pattern gene sets with high-confidence causal cues were extracted based on the extended gene set G. Simultaneously, explorer pattern gene sets with potential trait influence clues are extracted. ; Search space constraints were introduced when generating the candidate graph structure, including: limiting that in newly added candidate causal edges, the parent node gene pointing to the target trait must be located in the gene pool jointly defined by the conservation and exploration modes. Inside; The set of candidate new edges considered by the large language model in each iteration of the greedy search Represented as: ,in, Let E represent the set of potential causal edges pointing from a gene to a target trait, and let E represent the set of edges in the current structure. This represents a candidate directed edge from gene g to the target trait t. This represents a gene pool.

6. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 5, characterized in that, The steps of employing a BIC-based scoring function and assigning prior reward weights to structures containing conserved pattern genes include: By modifying the BIC scoring function to adjust the relative priority of candidate graph structures, structures containing high-confidence biological genes are given a scoring bias. The biological high-confidence gene set is the conserved pattern gene set. ; The number of high-confidence biological genes in edge set E that require positive rewards Represented as: ； in, For indicator functions, when The value is 1 if the condition is true, and 0 otherwise. This represents a candidate edge pointing from the conserved pattern gene g to the target trait t; The standard BIC scoring function is extended into a composite scoring function that incorporates prior reward terms. , represented as: ； in, The BIC score for the graph structure corresponding to the edge set E; These are preset hyperparameters used to adjust the strength of the prior reward weights. Indicates the number of genes.

7. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 6, characterized in that, Standardized direct causal effects are calculated using the following steps: Based on the optimal directed acyclic graph, a logistic regression model is used to calculate the standardized direct causal effect coefficient of each candidate gene toward the target stress trait. The calculation model expression is: ； In the formula, Given a gene expression matrix X, the conditional probability that a crop sample exhibits the target stress trait is given. For binary trait labels; Let E represent the set of parent node genes in the edge set E that point to the target trait t; The normalized expression level of any gene i in the set of genes of the parent node; For the intercept term; The logistic regression coefficient of any gene i in the set of genes of the parent node pointing to the target trait t; When the candidate gene g to be evaluated belongs to the set of genes of the parent node, that is When, its corresponding coefficient This is the standardized direct causal effect coefficient of candidate gene g pointing to target trait t; using The ratio was calculated. The strength and statistical uncertainty of the causal effect are assessed based on the confidence interval of the ratio; and the standardized direct causal effect coefficient is used as a basis for further analysis. The positive or negative sign of the gene indicates whether the candidate gene g is in a positive or negative regulatory direction on the target trait.

8. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 7, characterized in that, In the process of driving causal structure learning with a large language model, the CSCOT prompt word framework is introduced to guide the large language model to perform the following steps: Load a unified expression matrix X and structured biological prior information containing labels for the conserved mode and the exploratory mode, and perform feature alignment between continuous expression data and discrete prior modes in the context space of the prompt word; During the candidate topology generation phase, instructions are used to restrict the starting node of directed edges pointing to the target trait to be located in the candidate gene pool. Inside; The large language model is driven to perform a greedy strategy-based graph structure search iteration, utilizing a composite scoring function in each iteration. Evaluate the quality of the graph structure and decide whether to add, delete, or reverse edges. When determining the composite scoring function After reaching a local optimum, the optimal directed acyclic graph is locked, and a logistic regression model is used to calculate the standardized direct causal effect coefficient of each candidate gene toward the target stress trait. and ratio ratio .

9. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 8, characterized in that, The steps for generating a causal explanation that includes descriptions of regulatory direction, intensity of action, and molecular pathways specifically include: From the reasoning basis of the five reasoning dimensions output by BioCoT, extract the key support points of candidate causal genes, and identify the core mechanism module to which the candidate causal genes belong based on the key support points. The standardized direct causality coefficient calculated by the CSCOT module and ratio ratio Logically align and map attributes with the core mechanism modules to establish the relationship between quantitative contributions and qualitative mechanisms; By using association relationships to limit the evidence boundaries when generating explanations using large language models, the generated causal explanations are required to be deduced based on the key supporting points and quantitative causal effects, so as to deduce the regulatory pathways of the candidate causal genes on the target stress trait.

10. The method for interpretable mining of causal genes for crop stress traits based on a large model according to claim 9, characterized in that, The step of generating a causal explanation that includes descriptions of regulatory direction, effect intensity, and molecular pathway also includes using the CECoT cue word framework to guide the large model to perform the following steps: Based on the inference records of BioCoT, the core supporting evidence and reasoning basis of the candidate causal genes under multiple inference dimensions are extracted, and their core mechanism modules are identified. The standardized direct causal effect coefficient output by CSCoT and ratio ratio As known quantitative facts, semantic alignment is performed with the core mechanism module, and the causal contribution strength between changes in gene expression levels and trait changes is assessed using the large language model. Under the constraints of the core mechanism module, and combined with the crop stress biology knowledge built into the large language model, the specific biological effects of the gene at the molecular and cellular levels are deduced, and the physiological state regulation mechanism mediated by the candidate causal gene in the stress response process is elucidated. By deeply integrating the direction and intensity of biological and causal effects, the gene's role in promoting or inhibiting the target trait is determined, and a structured natural language causal explanation is ultimately generated. The causal explanation includes at least quantitative causal contribution indicators, traceable mechanistic clues, and descriptions of molecular regulatory pathways.

Images