Pathogenic gene identification method and system based on multi-agent debate and medium
Through a multi-agent debate method, combined with genetic testing data and clinical phenotype data, and using the Bayesian update mechanism for multiple rounds of debate, the problems of sample dependence and lack of interpretability in existing methods are solved, and efficient and transparent pathogenic gene identification and analysis are achieved.
Patent Information
- Application Number
- CN202510812787.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-18
- Publication Date
- 2025-09-26
AI Technical Summary
Existing methods for identifying pathogenic genes rely on large amounts of training data and lack interpretability, resulting in low identification efficiency in rare genetic diseases and results that are difficult to be clinically accepted.
A multi-agent debate method is adopted, through the collaborative work of data agents and knowledge agents, multiple rounds of debate are carried out based on the Bayesian update mechanism, and genetic testing data and clinical phenotype data are combined to generate natural language pathogenic gene ranking results, providing a detailed reasoning process and chain of evidence.
It improves the accuracy and interpretability of pathogenic gene identification, enables effective identification in small sample scenarios, outputs structured reasoning analysis reports, and enhances the credibility and transparency of clinical applications.
Smart Images

Figure CN120708720A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of pathogenic gene identification, and in particular to a pathogenic gene identification method, system and medium based on multi-agent debate. Background Art
[0002] The identification of pathogenic genes, especially in rare genetic diseases, often relies on a multidisciplinary collaborative mechanism (Multidisciplinary Team, MDT for short). This mechanism is usually composed of clinicians, genetics experts, bioinformatics experts, etc., combining knowledge from different fields, conducting in-depth analysis and reasoning of cases from multiple perspectives, and ultimately reaching a more accurate diagnostic conclusion. However, the implementation of MDT is extremely dependent on high-level physician resource allocation and a large amount of time cost. Different medical institutions have significant differences in professional personnel structure, knowledge and experience accumulation, tool support capabilities, etc., resulting in inter-institutional heterogeneity in the identification results of pathogenic gene variants. In addition, the MDT mechanism is essentially still dependent on human participation, which is difficult to achieve scale and automation, resulting in low efficiency.
[0003] In recent years, with the widespread application of artificial intelligence in medicine, disease-causing gene identification tools such as diseaseGPS, Xrare, Exomiser, and LIRICAL have gradually emerged. These tools generally use statistical models or machine learning algorithms to score and rank potential disease-causing genes based on genetic variation and phenotypic data, providing clinical auxiliary diagnostic recommendations. These systems have significantly improved the screening efficiency and diagnostic coverage of rare diseases, becoming an important supplement to MDT.
[0004] However, these existing genetic disease pathogenic gene identification tools mainly analyze the harmfulness of pathogenic genes, including functional damage analysis caused by mutations, population frequency analysis, and whether there are functional experimental verifications, and then use scoring or machine learning methods to classify the pathogenicity of the aforementioned mutation characteristics (pathogenic, possibly pathogenic, uncertain, possibly harmless, harmless, etc.). Finally, based on the scores, a pathogenicity ranking list is given as the diagnostic result.
[0005] However, this technical strategy has two major problems: 1) Insufficient sample size. Methods that rely on neural networks or machine learning algorithms, such as Xrare and diseaseGPS, require a large number of samples for pre-training to ensure the accuracy and robustness of the model. However, the incidence of rare genetic diseases is low, and the sample size is scarce, making it impossible to obtain sufficient samples to support effective training of the model. This type of method can only rely on the Clinvar database or variant sites in reported cases to construct simulated cases for model training. However, the background genomes of real cases and simulated cases are often quite different, resulting in lower robustness of the model. 2) Insufficient interpretability. The pathogenicity of each gene can only be calculated mathematically based on various evidences, and the pathogenicity probability of the mutant gene can be ranked according to the calculated values. The reason for the ranking cannot be given from the pathogenic mechanism, making it difficult for clinicians to evaluate whether the results are accurate, and thus difficult to promote and apply in clinical practice.
[0006] CN119339921A discloses a disease diagnosis method, device, equipment and storage medium, which provides an intelligent disease diagnosis model to assist doctors in improving the accuracy and efficiency of clinical diagnosis, solves the problem of insufficient disease knowledge in general medical large models, and can fully tap the knowledge potential of large models. The model is guided to verify the accuracy of pre-diagnosis results according to current medical guidelines to ensure the interpretability of medical decisions. The model aims to improve the accuracy and efficiency of complex disease diagnosis by integrating multi-scale knowledge enhancement, multi-agent debate optimization and self-correction diagnosis collaborative verification. However, in this method, the task is split into different hierarchical relationships (low, medium and high), the higher-level inference is based on the lower-level inference results, and the final result is the result of multiple rounds of voting by different agents. The debate content and depth are insufficient, the debate process is relatively broad, and there is a lack of active mining of new information on the contradictions in the debate to support or refute the other party's point of view, and it mainly revolves around the diagnosis result itself. The accuracy and interpretability of the final conclusion need to be further improved. Summary of the Invention
[0007] The purpose of the present invention is to overcome the defects of existing pathogenic gene identification, such as dependence on training samples and lack of interpretability, and to provide a pathogenic gene identification method, system and medium based on multi-agent debate to improve the accuracy and interpretability of the identification results.
[0008] The purpose of the present invention can be achieved by the following technical solutions:
[0009] A method for identifying pathogenic genes based on multi-agent debate, comprising the following steps:
[0010] Data acquisition: Obtain genetic testing data and clinical phenotype data in natural language text format;
[0011] Data agent analysis and debate: The data agent analyzes the acquired genetic testing data and clinical phenotype data, calls the gene variant pathogenicity analysis operator to output a ranked list of candidate pathogenic genes, inputs this ranked list of candidate pathogenic genes into the self-trained first molecular genetics model, generates hypothesis evidence based on molecular genetics knowledge, formats the results and outputs them, obtaining the first pathogenic gene information;
[0012] Knowledge agent analysis and debate: The knowledge agent analyzes the first pathogenic gene information and clinical phenotype data and inputs them into the self-trained second molecular genetics model. Based on molecular genetics knowledge, it generates hypotheses, re-ranks the gene pathogenicity, formats the results, and outputs them, providing the reasons for the ranking in natural language to obtain the second pathogenic gene information;
[0013] Multi-agent debate: Based on the ranking results of the pathogenic gene information they each output, the knowledge agent and the data agent each speak in turn about the pathogenicity of the candidate pathogenic genes in the first or second pathogenic gene information. Each speech is based on the other party's speech and its own previous speech to debate. Among them, the data agent mines new information from genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and incorporates the newly mined information into the rebuttal argument for the knowledge agent to obtain new information and make judgments.
[0014] Result output: After each round of debate, the debate agent determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The debate agent reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking results. Otherwise, it enters the next round of debate.
[0015] The data agent performs data analysis on the acquired genetic testing data and clinical phenotype data by converting the input natural language text into an HPO-coded ID through retrieval enhancement generation to obtain a structured file containing abnormal phenotype information.
[0016] The gene variation pathogenicity analysis operator calculates the empirical similarity score, the maximum information score and the theoretical phenotypic similarity score based on the sequencing file containing the variation, and sorts the pathogenic genes from high to low according to the empirical similarity score. If the empirical similarity scores are the same, they are further sorted from high to low based on the maximum information score. If the maximum information scores are also the same, they are further sorted from high to low based on the theoretical phenotypic similarity score, and finally outputs a ranked list of candidate pathogenic genes.
[0017] The calculation method of the empirical similarity score SimE is:
[0018] SimE=GE i +PE i
[0019] Among them, GE i is the score of gene function damage directly and indirectly related to disease i, with a value range of [0,1]; PE i is the matching score between the patient and the known abnormal phenotype of disease i, which takes a value of 0 or 1, where 0 indicates no match and 1 indicates a match;
[0020] The calculation method of the maximum information score MIC is:
[0021] MIC=max(IC(t1,…,t n ))
[0022] Among them, IC(t1,…,t n ) is the patient’s abnormal phenotype t1,…,t n The information score of , n is the number of abnormal phenotypes;
[0023] The calculation method of the theoretical phenotypic similarity score SimT is:
[0024] SimT(t,T i )=ScoreP i +ScoreG i
[0025] Among them, ScoreP i ScoreG is the matching score between the patient's abnormal phenotype and the known phenotype of the disease. i It is the hypergeometric distribution test score between the theoretical abnormal phenotype corresponding to the patient's abnormal gene and the known phenotype of the disease.
[0026] The formatting and outputting of the results specifically includes: formatting the results according to the pathogenic gene identification results, molecular basis, phenotypic basis and argumentation basis to generate natural language text description results, wherein the pathogenic gene identification results are the ranking of candidate pathogenic genes.
[0027] The multi-agent debate process is implemented based on structured Bayesian updating.
[0028] During the multi-agent debate, the decision of each data agent in each round is an update of the correctness of the prior probability of the decision in the current round, which is expressed as:
[0029]
[0030] The decision of the knowledge agent in each round is an update of the correctness of the prior probability of the decision in the current round, which can be expressed as:
[0031]
[0032] Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and represent the specific evidence generated by data and knowledge agent after the tth round of debate; P(G i =1|H D ) and P(G i =1|H K ) represent the probability that the data agent and the knowledge agent believe that gene i causes disease; and They represent the probability that the data agent and the knowledge agent will update and provide specific update evidence in the first round; and They represent the probability of the data agent and the knowledge agent giving specific updated evidence when assuming that gene i is the pathogenic gene in the t-1 round of debate; and They represent the probability that the data agent and the knowledge agent will give specific updated evidence based on the other party's evidence in the t-1 round of debate, assuming that gene i is the pathogenic gene; and They represent the probability that the data agent and the knowledge agent, in the tth round of debate, assume that gene i is pathogenic and give specific updated evidence based on the evidence of both parties in the previous round during the multi-agent debate.
[0033] The debate ends, and the data agent and the knowledge agent satisfy the following inequality:
[0034]
[0035] Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and Respectively represent the specific evidence generated by the data and knowledge agents after the first round of debate; and They represent the probability that the data agent and the knowledge agent believe that gene i causes disease based on their preliminary analysis.
[0036] A pathogenic gene identification system based on multi-agent debate is used to implement the method described above. The system includes a data agent, a knowledge agent and a debate agent, wherein:
[0037] The data agent includes a first data parsing module, a gene variation pathogenicity analysis operator, a first molecular genetics macromodel, and a first result formatting module. The first data parsing module performs data parsing on acquired genetic testing data and clinical phenotype data in natural language text format. The gene variation pathogenicity analysis operator is used to output a ranked list of pathogenic gene candidates. The first molecular genetics macromodel uses the input ranked list of pathogenic gene candidates and clinical phenotypes to generate a hypothesis based on molecular genetics knowledge. The first result formatting module formats the output of the first molecular genetics macromodel to obtain first pathogenic gene information.
[0038] The knowledge agent includes a second data parsing module, a second molecular genetic model, and a second result formatting module. The second data parsing module performs data parsing on the first pathogenic gene information and clinical phenotype data, and inputs the data into the second molecular genetic model. The second molecular genetic model generates a hypothesis based on molecular genetic knowledge and re-ranks the pathogenicity of the genes. The second result formatting module formats the output of the second molecular genetic model and provides a ranking reason in natural language to obtain the second pathogenic gene information.
[0039] The debate agent includes a result analysis module, a debate guidance module and a result reasoning module. After each round of debate, the result analysis module determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The result reasoning module reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking result. Otherwise, it enters the next round of debate, and the debate guidance module guides the data agent and the knowledge agent to debate. The debate process is as follows: the data agent and the knowledge agent speak in turn on the pathogenicity of the candidate pathogenic genes in the second pathogenic gene information based on the ranking results in the pathogenic gene information they respectively output. Each speech is debated in combination with the other party's speech and its previous speech. Among them, the data agent mines new information from the genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and integrates the newly mined information into the rebuttal argument for the knowledge agent to obtain and judge new information.
[0040] A storage medium stores a program, which implements the method when the program is executed.
[0041] Compared with the prior art, the present invention has the following beneficial effects:
[0042] (1) This invention is based on the Bayesian method and multi-agent multi-round debate mechanism. Compared with traditional machine learning methods, it does not need to rely on large-scale training data, and solves the problem of model construction in scenarios with scarce data and small samples.
[0043] (2) The entire decision-making path of the present invention is described in natural language, fully presenting each round of reasoning process and the chain of evidence use. Compared with calculating gene pathogenicity scores (or classifications) based on evidence or performing pathogenicity classification based on machine learning methods, the process is transparent and the logic is explainable, which significantly enhances the usability and trustworthiness of the reasoning results in actual medical scenarios.
[0044] (3) The debate process of the present invention is jointly participated by data agents and knowledge agents. Data agents focus on the analysis and mining of genetic testing data and clinical phenotype data, and can extract key information from massive data; knowledge agents focus on using molecular genetics knowledge to analyze and sort pathogenic genes. This clear division of labor in the role setting of agents enables the debate to be carried out from the two key dimensions of data and knowledge, comprehensively and in-depth exploring the possibility of pathogenic genes, and avoiding omissions and biases that may result from single-perspective analysis. Data agents and knowledge agents each play their own advantages, identifying and analyzing pathogenic genes from the perspectives of data-driven and knowledge-driven, respectively, which can more accurately locate pathogenic genes and improve the accuracy of pathogenic gene identification.
[0045] (4) The debate process of this application not only focuses on the preliminary identification results of pathogenic genes, but also requires data agents and knowledge agents to speak in turn in the debate. Each speech must be combined with the other party's speech and their previous speech for in-depth analysis. This debate method encourages the agents to conduct a more in-depth analysis of the pathogenicity of pathogenic genes, including the biological significance, functional impact, and correlation with diseases of gene mutations, so as to explore deeper pathogenic mechanisms and potential genetic factors. Through in-depth analysis of the pathogenicity of pathogenic genes, this application can comprehensively evaluate the sequencing results of pathogenic genes and provide a more reliable basis for clinical diagnosis and treatment.
[0046] (5) During the debate, the data agent will proactively mine new information from genetic testing data and clinical phenotypic data based on the conflicting points in the debate to support or refute the views of the knowledge agent. This ability to dynamically mine new information makes the debate process highly adaptable and flexible, enabling the timely discovery of new evidence and clues to further improve the identification and analysis of pathogenic genes. During the debate, the knowledge agent can obtain and judge the new information mined by the data agent, and provide feedback and adjustments to its own views and ranking results based on this new information, thereby improving the accuracy of the analysis and judgment of pathogenic genes.
[0047] (6) After each round of debate, the debate agent will reason based on the consensus and disagreement between the data agent and the knowledge agent. By comprehensively analyzing these different opinions, it can more comprehensively reflect the complexity and uncertainty of pathogenic gene identification and output a structured reasoning analysis report and the final pathogenic gene ranking results. This structured report format can not only clearly display the various viewpoints and evidence during the debate process, but also provide clinicians with detailed pathogenic gene analysis and diagnostic basis, facilitating them to make more scientific and reasonable treatment decisions, and greatly improving interpretability. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 is a flow chart of the method of the present invention;
[0049] Figure 2 This is a system structure diagram of the present invention. DETAILED DESCRIPTION
[0050] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0051] Example 1
[0052] This embodiment provides a method for identifying pathogenic genes based on multi-agent debate. The method is implemented by a data agent, a knowledge agent, and a debate agent. Based on the Bayesian update strategy, the three agents engage in multiple rounds of debate and new evidence submission around candidate diagnostic hypotheses, continuously and autonomously updating the prior probability of each agent's decision, and realizing the dynamic evolution of the diagnostic hypothesis. When the three agents gradually converge and reach a consensus, the final result is given. When the data and knowledge agents cannot converge to a unified conclusion and the maximum number of debate rounds is reached, the debate agent performs knowledge reasoning based on the results of all the above debate rounds and gives the final decision. This method does not require sample pre-training, which solves the dependence of existing methods on sample size. At the same time, the entire process is carried out in natural language, all processes are transparent and traceable, and the identification results and reasoning paths are interpretable, which solves the problem of poor interpretability of existing methods. Through this method, multi-agent debate and identification of pathogenic genes for rare genetic diseases can be completed autonomously without human intervention, with a high degree of intelligence.
[0053] Specifically, such as Figure 1 As shown, the method includes the following steps:
[0054] S1, Data Acquisition: Obtain genetic testing data and clinical phenotype data in the form of natural language text.
[0055] S2, data agent analysis and debate: The data agent performs data analysis on the acquired genetic testing data and clinical phenotype data, calls the gene variation pathogenicity analysis operator to output a ranked list of pathogenic gene candidates, inputs the ranked list of pathogenic gene candidates into the self-trained first molecular genetic model, generates hypothesis basis based on molecular genetic knowledge, and formats the results and outputs them to obtain the first pathogenic gene information.
[0056] In this step, data parsing specifically involves converting the input natural language text into an HPO-coded ID through retrieval enhancement generation to obtain a structured file containing abnormal phenotypic information.
[0057] The gene variant pathogenicity analysis operator calculates the empirical similarity score, maximum information score, and theoretical phenotypic similarity score based on the sequencing file containing the variant, and sorts the pathogenic genes from high to low according to the empirical similarity score. If the empirical similarity scores are the same, they are further sorted from high to low based on the maximum information score. If the maximum information scores are also the same, they are further sorted from high to low based on the theoretical phenotypic similarity score, and finally outputs a ranked list of candidate pathogenic genes.
[0058] The calculation method of the empirical similarity score SimE is:
[0059] SimE=GE i +PE i
[0060] Among them, GE i is the score of gene function damage directly and indirectly related to disease i, with a value range of [0,1]; PE i is the matching score between the patient and the known abnormal phenotype of disease i, which takes a value of 0 or 1, where 0 indicates no match and 1 indicates a match.
[0061] The calculation method of the maximum information score MIC is:
[0062] MIC=max(IC(t1,…,t n ))
[0063] Among them, IC(t1,…,t n ) is the patient’s abnormal phenotype t1,…,t n is the information score of , and n is the number of abnormal phenotypes.
[0064] The theoretical phenotypic similarity score SimT is calculated as:
[0065] SimT(t,T i )=ScoreP i +ScoreG i
[0066] Among them, ScoreP i ScoreG is the hypergeometric distribution test score of the patient's abnormal phenotype and the known disease phenotype. i It is the hypergeometric distribution test score between the theoretical abnormal phenotype corresponding to the patient's abnormal gene and the known phenotype of the disease.
[0067] S3, knowledge agent analysis and debate: The knowledge agent analyzes the first pathogenic gene information and clinical phenotype data and inputs them into the self-trained second molecular genetic model. Based on molecular genetic knowledge, it generates hypothesis basis, re-ranks the gene pathogenicity, formats the results and outputs them, and gives the reasons for the ranking in natural language to obtain the second pathogenic gene information.
[0068] In step S2 and step S3, the results are formatted and outputted specifically as follows: the results are formatted to generate natural language text description results according to the pathogenic gene identification results, molecular basis, phenotypic basis and argumentation basis, wherein the pathogenic gene identification results are the ranking of candidate pathogenic genes.
[0069] S4, multi-agent debate: Based on the ranking results of the pathogenic gene information output by each, the knowledge agent and the data agent speak in turn on the pathogenicity of the candidate pathogenic genes in the first or second pathogenic gene information. Each speech is combined with the other party's speech and its own previous speech for debate. Among them, the data agent mines new information from the genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and integrates the newly mined information into the rebuttal argument for the knowledge agent to obtain and judge new information.
[0070] In this embodiment, the multi-agent debate process is implemented based on structured Bayesian updating. This embodiment is described in detail using the data agent as an example. The reasoning process of the knowledge agent can refer to the data agent, and this embodiment will not be repeated here.
[0071] First round of debate on data agents:
[0072]
[0073] Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and Respectively represent the specific evidence generated by the data and knowledge agents after the first round of debate; represents the probability of the data agent assuming that gene i is pathogenic and proposing specific updated evidence in the first round of multi-agent debate. It represents the probability that the data agent believes that gene i causes disease based on the preliminary analysis of the two agents; Represents the probability that the data agent updates and provides evidence for a specific update based on the initial analysis of the two agents.
[0074] The first round of update process of the data agent can be specifically as follows:
[0075]
[0076] The decision of the data agent in each round is an update of the correctness of the prior probability of the decision in the current round, which is expressed as:
[0077]
[0078] Based on this, we can deduce that the decision of each round of the knowledge agent is an update of the correctness of the prior probability of the current round of decision, which can be expressed as:
[0079]
[0080] Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and represent the specific evidence generated by data and knowledge agent after the tth round of debate; P(G i =1|H D ) and P(G i =1|H K ) represent the probability that the data agent and the knowledge agent believe that gene i causes disease; and They represent the probability that the data agent and the knowledge agent will update and provide specific update evidence in the first round; and They represent the probability of the data agent and the knowledge agent giving specific updated evidence in round t-1, assuming that gene i is a pathogenic gene; and They represent the probability that the data agent and the knowledge agent will give specific updated evidence based on the other party's evidence in the t-1 round of debate, assuming that gene i is the pathogenic gene; and They represent the probability that the data agent and the knowledge agent, in the tth round of debate, assume that gene i is pathogenic and give specific updated evidence based on the evidence of both parties in the previous round during the multi-agent debate.
[0081] S5, result output: After each round of debate, the debate agent determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The debate agent reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking results. Otherwise, it enters the next round of debate.
[0082] After the debate, the data agent and the knowledge agent satisfy the following inequality:
[0083]
[0084] Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and Respectively represent the specific evidence generated by the data and knowledge agents after the first round of debate; and They represent the probability that the data agent and the knowledge agent believe that gene i causes disease based on their preliminary analysis.
[0085] Example 2
[0086] This embodiment provides a pathogenic gene identification system based on multi-agent debate, which is used to implement the method of embodiment 1. Figure 2 As shown, the system includes data agent, knowledge agent and debate agent, among which,
[0087] The data agent includes a first data parsing module, a gene variation pathogenicity analysis operator, a first molecular genetics macromodel, and a first result formatting module. The first data parsing module performs data parsing on the acquired genetic testing data and clinical phenotypic data in the form of natural language text. The gene variation pathogenicity analysis operator is used to output a ranked list of pathogenic gene candidates. The first molecular genetics macromodel uses the input ranked list of pathogenic gene candidates to generate a hypothesis based on molecular genetics knowledge. The first result formatting module formats the output of the first molecular genetics macromodel to obtain first pathogenic gene information.
[0088] The knowledge agent includes a second data parsing module, a second molecular genetic model, and a second result formatting module. The second data parsing module performs data parsing on the first pathogenic gene information and clinical phenotype data, and inputs the data into the second molecular genetic model. The second molecular genetic model generates a hypothesis based on molecular genetic knowledge and re-ranks the pathogenicity of the genes. The second result formatting module formats the output of the second molecular genetic model and provides a reason for the ranking in natural language to obtain the second pathogenic gene information.
[0089] The debate agent includes a result parsing module, a debate guidance module, and a result reasoning module. After each round of debate, the result parsing module determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The result reasoning module reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking results. Otherwise, it enters the next round of debate, and the debate guidance module guides the data agent and the knowledge agent to debate. The debate process is as follows: based on the ranking results in the pathogenic gene information they each output, the knowledge agent and the data agent speak in turn on the pathogenicity of the candidate pathogenic genes in the first or second pathogenic gene information. Each speech is debated in combination with the other party's speech and its previous speech. Among them, the data agent mines new information from the genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and incorporates the newly mined information into the rebuttal argument for the knowledge agent to obtain and judge new information.
[0090] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0091] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0092] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.
Claims
1. A pathogenic gene identification method based on multi-agent debate, characterized in that: The following steps are involved: Data acquisition: Obtain genetic testing data and clinical phenotype data in natural language text format; Data agent analysis and debate: The data agent analyzes the acquired genetic testing data and clinical phenotype data, calls the gene variant pathogenicity analysis operator to output a ranked list of candidate pathogenic genes, inputs this ranked list of candidate pathogenic genes into the self-trained first molecular genetics model, generates hypothesis evidence based on molecular genetics knowledge, formats the results and outputs them, obtaining the first pathogenic gene information; Knowledge agent analysis and debate: The knowledge agent analyzes the first pathogenic gene information and clinical phenotype data and inputs them into the self-trained second molecular genetics model. Based on molecular genetics knowledge, it generates hypotheses, re-ranks the gene pathogenicity, formats the results, and outputs them, providing the reasons for the ranking in natural language to obtain the second pathogenic gene information; Multi-agent debate: Based on the ranking results of the pathogenic gene information they each output, the knowledge agent and the data agent each speak in turn about the pathogenicity of the candidate pathogenic genes in the first or second pathogenic gene information. Each speech is based on the other party's speech and its own previous speech to debate. Among them, the data agent mines new information from genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and incorporates the newly mined information into the rebuttal argument for the knowledge agent to obtain new information and make judgments. Result output: After each round of debate, the debate agent determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The debate agent reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking results. Otherwise, it enters the next round of debate.
2. A pathogenic gene identification method based on multi-agent debate according to claim 1, characterized in that: The data agent performs data analysis on the acquired genetic testing data and clinical phenotype data by converting the input natural language text into an HPO-coded ID through retrieval enhancement generation to obtain a structured file containing abnormal phenotype information.
3. The method for identifying pathogenic genes based on multi-agent debate according to claim 1, characterized in that: The gene variation pathogenicity analysis operator calculates the empirical similarity score, the maximum information score and the theoretical phenotypic similarity score based on the sequencing file containing the variation, and sorts the pathogenic genes from high to low according to the empirical similarity score. If the empirical similarity scores are the same, they are further sorted from high to low based on the maximum information score. If the maximum information scores are also the same, they are further sorted from high to low based on the theoretical phenotypic similarity score, and finally outputs a ranked list of candidate pathogenic genes.
4. The method for identifying pathogenic genes based on multi-agent debate according to claim 3, characterized in that: The calculation method of the empirical similarity score SimE is: SimE=GE i +PE i Among them, GE i is the score of gene function damage directly and indirectly related to disease i, with a value range of [0,1]; PE i is the matching score between the patient and the known abnormal phenotype of disease i, which takes a value of 0 or 1, where 0 indicates no match and 1 indicates a match; The calculation method of the maximum information score MIC is: MIC=max(IC(t1,…,t n )) Among them, IC(t1,…,t n ) is the patient’s abnormal phenotype t1,…,t n The information score of , n is the number of abnormal phenotypes; The calculation method of the theoretical phenotypic similarity score SimT is: SimT(t,T i )=ScoreP i +ScoreG i Among them, ScoreP i ScoreG is the matching score between the patient's abnormal phenotype and the known phenotype of the disease. i It is the hypergeometric distribution test score between the theoretical abnormal phenotype corresponding to the patient's abnormal gene and the known phenotype of the disease.
5. The method for identifying pathogenic genes based on multi-agent debate according to claim 1, characterized in that: The formatting and outputting of the results specifically includes: formatting the results according to the pathogenic gene identification results, molecular basis, phenotypic basis and argumentation basis to generate natural language text description results, wherein the pathogenic gene identification results are the ranking of candidate pathogenic genes.
6. The method for identifying pathogenic genes based on multi-agent debate according to claim 1, characterized in that: The multi-agent debate process is implemented based on structured Bayesian updating.
7. The method for identifying pathogenic genes based on multi-agent debate according to claim 6, characterized in that: During the multi-agent debate, the decision of each data agent in each round is an update of the correctness of the prior probability of the decision in the current round, which is expressed as: The decision of the knowledge agent in each round is an update of the correctness of the prior probability of the decision in the current round, which can be expressed as: Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and represent the specific evidence generated by data and knowledge agent after the tth round of debate; P(G i =1|H D ) and P(G i =1|H K ) represent the probability that the data agent and the knowledge agent believe that gene i causes disease; and They represent the probability that the data agent and the knowledge agent will update and provide specific update evidence in the first round; and They represent the probability of the data agent and the knowledge agent giving specific updated evidence when assuming that gene i is the pathogenic gene in the t-1 round of debate; and They represent the probability that the data agent and the knowledge agent will give specific updated evidence based on the other party's evidence in the t-1 round of debate, assuming that gene i is the pathogenic gene; and They represent the probability that the data agent and the knowledge agent, in the tth round of debate, assume that gene i is pathogenic and give specific updated evidence based on the evidence of both parties in the previous round during the multi-agent debate.
8. The method for identifying pathogenic genes based on multi-agent debate according to claim 6, characterized in that: The debate ends, and the data agent and the knowledge agent satisfy the following inequality: Among them, H D and H K Represent data agent and knowledge agent respectively; and represent the initial evidence of data agent and knowledge agent respectively; G i =1 means gene i is determined to be a pathogenic gene; and Respectively represent the specific evidence generated by the data and knowledge agents after the first round of debate; and They represent the probability that the data agent and the knowledge agent believe that gene i causes disease based on their preliminary analysis.
9. A pathogenic gene identification system based on multi-agent debate, characterized by: For implementing the method according to any one of claims 1 to 8, the system comprises a data agent, a knowledge agent and a debate agent, wherein: The data agent includes a first data parsing module, a gene variation pathogenicity analysis operator, a first molecular genetics macromodel, and a first result formatting module. The first data parsing module performs data parsing on acquired genetic testing data and clinical phenotype data in natural language text format. The gene variation pathogenicity analysis operator is used to output a ranked list of pathogenic gene candidates. The first molecular genetics macromodel uses the input ranked list of pathogenic gene candidates and clinical phenotypes to generate a hypothesis based on molecular genetics knowledge. The first result formatting module formats the output of the first molecular genetics macromodel to obtain first pathogenic gene information. The knowledge agent includes a second data parsing module, a second molecular genetic model, and a second result formatting module. The second data parsing module performs data parsing on the first pathogenic gene information and clinical phenotype data, and inputs the data into the second molecular genetic model. The second molecular genetic model generates a hypothesis based on molecular genetic knowledge and re-ranks the pathogenicity of the genes. The second result formatting module formats the output of the second molecular genetic model and provides a ranking reason in natural language to obtain the second pathogenic gene information. The debate agent includes a result analysis module, a debate guidance module and a result reasoning module. After each round of debate, the result analysis module determines whether the ranking results of the candidate pathogenic genes by the data agent and the knowledge agent are consistent. If they are consistent or the number of debate rounds reaches a given maximum threshold, the debate ends. The result reasoning module reasons based on the consensus and disagreement between the data agent and the knowledge agent, and outputs a structured reasoning analysis report and the final pathogenic gene ranking result. Otherwise, it enters the next round of debate, and the debate guidance module guides the data agent and the knowledge agent to debate. The debate process is as follows: the data agent and the knowledge agent speak in turn on the pathogenicity of the candidate pathogenic genes in the second pathogenic gene information based on the ranking results in the pathogenic gene information they respectively output. Each speech is debated in combination with the other party's speech and its previous speech. Among them, the data agent mines new information from the genetic testing data and clinical phenotype data based on the contradictions in the debate to support or refute the knowledge agent, and integrates the newly mined information into the rebuttal argument for the knowledge agent to obtain and judge new information.
10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.
Citation Information
Patent Citations
Disease diagnosis method and device, equipment and storage medium
CN119339921A
Electronic medical record-based difficult disease evidence-based method and system, and storage medium
CN117352159A
Multi-agent debate
CN119678162A
Cited By
New energy automobile charging state test method based on electromagnetic compatibility
CN121253921A
