A gene mutation pathogenicity automatic rating method and system based on multi-agent game and thought chain reasoning

Abstract

The present application relates to the technical field of the cross of bioinformatics, medical genetics and artificial intelligence, and discloses a gene mutation pathogenicity automatic rating method and system based on multi-agent game and thought chain reasoning, which comprises the following steps: searching data related to a target mutation site; based on the related data; performing multi-round iterative gene mutation pathogenicity rating by a proposer agent, a refuter agent and a decision maker agent; monitoring the rating conclusion output by the decision maker agent and the confidence score in real time, and when a preset dynamic truncation condition is met, terminating the multi-agent game step and outputting the final gene mutation pathogenicity rating conclusion. The present application can realize accurate and transparent rating of gene mutation pathogenicity by eliminating model hallucinations through adversarial games and analyzing unstructured documents by using thought chains.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of bioinformatics, medical genetics, and artificial intelligence, and in particular to an automated rating method and system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning. Background Technology

[0002] With the widespread adoption of Next-Generation Sequencing (NGS) technology, the amount of data generated by clinical genetic testing is growing exponentially. In the genetic disease diagnosis process, a core step is interpreting the pathogenicity of detected gene variants. Currently, the globally recognized interpretation standard is the Sequence Variation Interpretation Guidelines published by the American College of Medical Genetics and Genomics (ACMG) in conjunction with the Association for Molecular Pathology (AMP). These guidelines contain 28 criteria of evidence (such as PVS1, PS1, PM2, etc.), requiring interpreters to comprehensively consider multi-dimensional evidence, including population frequency, computational prediction, functional experiments, and pedigree co-segregation, to classify variants as "pathogenic," "probably pathogenic," "VUS (Very Undetermined Significance)," "probably benign," or "benign."

[0003] In early technological practices, this process primarily relied on manual intervention by genetic counselors. However, faced with a massive number of variant sites, manual interpretation was not only inefficient but also highly susceptible to subjective experience, leading to inconsistent ratings of the same variant across different laboratories. To address this efficiency bottleneck, the industry saw the emergence of the first generation of automated interpretation tools. These tools set up "rule-oriented" automated paths, using pre-defined logic code (such as Python scripts) to extract fields from structured databases (such as ClinVar and gnomAD) and perform ACMG entry matching based on hard-coded Boolean logic. For example, if the population frequency of a variant in the database is greater than 5%, the BA1 benign rule is automatically triggered.

[0004] While rule-based systems demonstrate some effectiveness in processing structured data, their underlying limitations are becoming increasingly apparent as precision medicine demands higher diagnostic accuracy. A significant amount of unstructured information exists in evidence of gene pathogenicity, particularly descriptions of functional studies, often scattered throughout free text within PubMed publications. For example, one publication might describe a 30% reduction in enzyme activity, while another might state that no significant abnormalities were observed in protein expression levels. Lacking semantic understanding capabilities, rule-based systems cannot parse such ambiguous or contradictory natural language descriptions, leading to the systematic underestimation of crucial evidence (such as PS3 / BS3 rules). Ultimately, this forces many variants to be classified as "VUS" (Very Undetermined Significance), failing to guide clinical decision-making.

[0005] To overcome the barriers to semantic understanding, researchers have begun to explore generative artificial intelligence techniques based on Large Language Models (LLMs). Some scholars have proposed a general architecture for medical and genetic analysis using agentic AI, attempting to replace manual document reading with the general reasoning capabilities of large models. However, in real-world clinical applications, this general agent architecture faces more subtle and critical mechanistic flaws—hallucination and confirmation bias.

[0006] From a mechanistic perspective, large single-model systems or general multi-agent systems lacking specific topological constraints are essentially generators that predict the next token based on probability. When the model is asked to find evidence of disease, its attention mechanism tends to focus on contexts supporting the disease, and may even try to "cater" the user's implicit expectations by piecing together irrelevant fragments when evidence is lacking. While some studies mention the application of agents, they lack adversarial verification mechanisms for conflicting evidence. When faced with a complex variant with both reports of disease and benign outcomes, general agents often cannot engage in multi-round debates and weighing of options like human expert groups. Instead, they tend to output the conclusion that appears most frequently in the model's training data, or get stuck in a logical dead loop.

[0007] Furthermore, existing multi-agent systems face a significant contradiction between computational power consumption and response time in engineering implementation. In high-throughput scenarios of gene testing, if full multi-agent deep inference is initiated for each variant, the token consumption and time cost will be astronomical, making it impossible for the system to complete the task within the clinical reporting cycle.

[0008] In summary, the field urgently needs an automated gene mutation pathogenicity rating system that can deeply understand unstructured literature evidence, eliminate model illusions through rigorous logical confrontation, and possess an engineered computing power control mechanism, in order to resolve the long-standing contradiction between accuracy, interpretability, and efficiency in existing methods. Summary of the Invention

[0009] This invention provides an automated rating method and system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, in order to overcome the shortcomings of existing technologies.

[0010] This invention provides an automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, comprising the following steps:

[0011] Data retrieval steps: Retrieve data related to the target variant site;

[0012] Multi-agent game steps: Based on the relevant data, the proposer agent, the refutator agent, and the decision-maker agent conduct multiple rounds of iterative gene mutation pathogenicity rating. Each round of the game includes:

[0013] Based on the currently retrieved chain of evidence, the proposing agent generates a first argument demonstrating that the target variant site is pathogenic.

[0014] The rebuttal agent searches for benign evidence and / or points out logical flaws in the first argument based on the relevant data, and generates a second argument refuting the pathogenicity of the target mutation site.

[0015] The decision-maker agent performs a weighted calculation by combining the first argument and the second argument, and outputs the rating conclusion and corresponding confidence score for the current round; wherein, in the N+1 round of the game, the first argument generated by the proposer agent includes the response to the second argument in the Nth round, so as to form a self-correcting feedback loop, where N is a positive integer;

[0016] Reasoning control steps: Monitor the rating conclusion and confidence score output by the decision-maker agent in real time. When the preset dynamic truncation condition is met, terminate the multi-agent game step and output the final gene mutation pathogenicity rating conclusion. The preset dynamic truncation condition includes: the confidence score in the current round is higher than a preset threshold, and / or the rating conclusions in a consecutive preset number of rounds are consistent.

[0017] According to the present invention, an automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning is provided. The data retrieval step further includes: identifying unstructured experimental data in the retrieved relevant data; parsing the natural language description of the unstructured experimental data using a large language model, and converting it into structured ACMG evidence fragments for use in the multi-agent game theory step.

[0018] According to the present invention, an automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning is provided. The first argument includes ACMG evidence entries supporting pathogenicity found by the proposer agent; the second argument includes ACMG evidence entries supporting benignity found by the refutation agent; and / or, a statistical significance review opinion on the literature cited in the first argument.

[0019] According to the present invention, an automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning is provided. The proposer agent, the refutator agent, and the decision-maker agent are all configured with a unified thought chain structure. The thought chain structure includes the following steps executed sequentially: retrieving evidence, matching ACMG evidence entries, evaluating the strength of evidence, and making a comprehensive judgment.

[0020] According to the present invention, an automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning is provided. The multi-agent game step further incorporates phenotypic expert agents and structural biology expert agents. Specifically, the phenotypic expert agent calculates the semantic similarity between the patient's clinical symptoms and the phenotype of the gene-related disease, and submits phenotypic supporting evidence to the decision-maker agent when the similarity exceeds a preset value. The structural biology expert agent analyzes whether the target mutation site disrupts the three-dimensional structural core region or active site of the protein, and submits structural supporting evidence to the decision-maker agent when disruption is confirmed.

[0021] According to the present invention, an automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning is provided. The decision-maker agent uses a weighted voting mechanism or a Bayesian weighting algorithm to assign different weights to the arguments and evidence output by the proposer agent, the refutator agent, the phenotypic expert agent, and the structural biology expert agent, and to comprehensively calculate the rating conclusion and corresponding confidence score for the current round.

[0022] The present invention provides an automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, which further includes: a state memory step: using a vector database to store the intermediate reasoning state, dialogue history, cited literature identifiers and thought chain reasoning process of each round of game, and generating an automated rating report containing a complete chain of evidence based on the stored data.

[0023] The present invention also provides an automated rating system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, comprising: a data retrieval module configured to perform the data retrieval steps as described in any of the above items;

[0024] A multi-agent game module, configured to execute the multi-agent game steps as described in any of the above items;

[0025] The inference control module is configured to perform the inference control steps as described in any of the above items.

[0026] In one implementation, the system further includes a state memory module configured to store the intermediate reasoning states, dialogue history, cited document identifiers, and thought chain derivation processes for each round of the game using a vector database, and to generate an automated rating report containing a complete chain of evidence based on the stored data.

[0027] The present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning as described in any of the preceding claims.

[0028] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning as described in any of the preceding claims.

[0029] It should be noted that the system / method of the present invention is not intended to directly obtain disease diagnosis results, but only to provide intermediate data / information processing.

[0030] The present invention provides an automated rating method and system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, which can bring at least the following beneficial effects:

[0031] By introducing an adversarial game architecture of "proposer agent - rebuttal agent," this invention constructs a self-correcting loop from a mechanistic perspective. The presence of the rebuttal agent forces the system to confront adverse evidence, thereby breaking the "echo chamber effect" of the monolithic model. Experimental data show that, compared to the monolithic CoT model, the hallucination rate of the system of this invention is reduced from 5.2% to 0.8%, ensuring the safety of medical-grade applications.

[0032] By combining RAG technology with a mandatory thought chain, the system of this invention can accurately identify experimental data (such as enzyme activity values ​​and protein expression levels) implicit in the literature. This significantly increases the recall rate of ACMG rules (such as PS3) that rely on functional experiments from 15% for rule-based tools to 68%, effectively solving the problem of VUS misjudgment caused by incomplete evidence extraction.

[0033] The dynamic truncation mechanism in the inference control module cleverly solves the challenge of implementing multi-agent systems in high-throughput clinical scenarios. For simple variants with clear evidence, the system can converge quickly within 1-2 rounds; only complex and contradictory variants undergo multiple rounds of in-depth debate. Tests show that this mechanism maintains a high accuracy of 92.4% while saving an average of approximately 40% in computing resources and time costs.

[0034] The mandatory thought chain structure ensures that every pathogenicity rating of a gene mutation in the system is verifiable. The final report not only includes conclusions but also fully preserves the entire logical process of "retrieval-matching-evaluation-judgment" as well as key turning points in the debate, greatly facilitating review by clinicians and establishing a foundation of trust for AI-assisted diagnosis. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This invention provides an architecture diagram of an automated rating system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning.

[0037] Figure 2 This is a schematic diagram of the core logic judgment and iterative control process of the multi-agent game module.

[0038] Figure 3 This is a schematic diagram showing the logical nodes of a mandatory thought chain structure.

[0039] Figure 4 The bar chart compares the performance of Example 1 with Comparative Example 1 (single agent), Comparative Example 2 (rule-guided system), and Comparative Example 3 (fixed rounds system) in two key dimensions: "ACMG rule activation accuracy" and "final rating consistency (Kappa coefficient)". The chart shows that Example 1 exhibits a significant advantage in activation rate on unstructured evidence rules such as PS3 (functional experiment) and PP4 (phenotypic matching), and is the only example to exceed the 0.90 threshold in Kappa coefficient. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0041] See Figure 1The system provided by this invention for implementing an automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning may include:

[0042] A. Data retrieval module, which is configured with external interfaces (such as ClinVar, gnomAD and PubMed external interfaces) and configured to retrieve data related to the target variant site (in the step of obtaining multi-source evidence, the Top-K value of vector retrieval is set to 5~20), identify unstructured data (including unstructured experimental data such as enzyme activity values) in the retrieved data related to the target variant site, and perform structure transformation on it.

[0043] B. A multi-agent game module, comprising a proposer agent, a refutation agent, and a decision-maker agent. The proposer agent, the refutation agent, and the decision-maker agent are configured to perform at least one round of gene mutation pathogenicity rating based on data related to the target mutation site retrieved by the data retrieval module (in this embodiment, each agent is constructed based on a large language model, the temperature coefficient of the large language model is set to 0~0.5, and the number of iteration rounds for gene mutation pathogenicity rating is set to 1~10 rounds).

[0044] Specifically, the proposer agent is configured to search for pathogenic evidence (PP, PS, PM, PVS types) to demonstrate that the target variant site is pathogenic, and is further configured to include in the arguments generated in the N+1th round a response to the rebuttal arguments generated by the rebuttal agent in the Nth round; the rebuttal agent is configured to search for benign evidence (BP, BS, BA types) and / or point out logical flaws in the proposer agent to refute the pathogenicity of the target variant site; the decision-maker agent is configured to rate the pathogenicity of the target variant site based on the arguments of the proposer agent and the rebuttal agent.

[0045] Specifically, the proposer agent, the refutation agent, and the decision-maker agent are configured with a unified thought chain structure, which includes the following steps executed sequentially: retrieving evidence, matching ACMG entries, evaluating the strength of the evidence, and making a comprehensive judgment.

[0046] In one embodiment, the multi-agent game-playing module may further include a phenotype expert agent and a structural biology expert agent. The phenotype expert agent calculates the semantic similarity between the patient's clinical symptoms and the phenotype of a gene-related disease, and submits phenotype-supporting evidence to the decision-maker agent when the similarity exceeds a preset value. The structural biology expert agent analyzes whether the target mutation site disrupts the three-dimensional structural core region or active site of the protein, and submits structural-supporting evidence to the decision-maker agent when disruption is confirmed. In this case, the decision-maker agent is configured to use a weighted voting mechanism to comprehensively determine the outputs of the proposer agent, the refuter agent, the phenotype expert agent, and the structural biology expert agent.

[0047] C. Inference Control Module: The inference control module is configured to coordinate the multi-agent game module to perform gene mutation pathogenicity rating rounds based on the rating conclusion of the decision-maker agent in each round of gene mutation pathogenicity rating, through preset conditions. The preset conditions include: terminating the multi-agent game module's gene mutation pathogenicity rating when the confidence score of the decision-maker agent's rating conclusion is higher than a preset threshold and / or the rating conclusions are consistent for a consecutive preset number of rounds (in this embodiment, the preset threshold for confidence score ranges from 0.8 to 0.99, and the consecutive preset number of rounds is two consecutive rounds).

[0048] D. State memory module, wherein the state memory module is configured to use a vector database to store the intermediate reasoning states and evidence chains of each round of the multi-agent game module.

[0049] The following examples 1-3, as well as Comparative Examples 1 (single agent), 2 (rule-oriented system), and 3 (fixed round system), will further illustrate the automated gene mutation pathogenicity rating system based on multi-agent game theory and thought chain reasoning provided by the present invention.

[0050] Example 1

[0051] This embodiment details an automated gene variant pathogenicity rating system deployed in a cloud-based high-performance computing environment. This system aims to address the technical challenges of low efficiency and poor consistency in interpreting variants of unknown significance (VUS) in clinical high-throughput sequencing scenarios.

[0052] like Figure 1As shown, the physical carrier of this system is a local server 100 configured with a high-performance graphics processing unit. In its specific construction, server 100 is equipped with eight NVIDIA A100 GPUs, each with 80GB of high-bandwidth video memory (HBM2e), achieving an inter-card interconnect bandwidth of 600GB / s via NVLink. It should be noted that the high-performance graphics processing unit is not limited to the NVIDIA A100; it can also be a TPU (Tensor Processing Unit), an NPU (Neural Processing Unit), or any parallel computing accelerator card with high-bandwidth video memory (such as HBM), as long as its video memory bandwidth and computing power meet the requirements of large model inference. This hardware topology is designed to support closed-source large language models with more than 175 billion parameters (such as GPT-4 Turbo-level API images or private deployment versions), ensuring that video memory throughput does not become a bottleneck during multi-agent concurrent inference.

[0053] The system software environment is based on the Linux Ubuntu 22.04 LTS operating system, with the underlying computing framework using PyTorch 2.1 and the upper-layer application logic orchestrated through the LangChain framework. The data retrieval module 200 is connected to an external knowledge base 500 via a fiber optic network, including ClinVar, gnomAD v4, PubMed summary repository, and the full repository.

[0054] During system initialization, the data retrieval module 200 is configured to listen for variant site input streams from the LIMS (Laboratory Information Management System). When a request such as NM_000059.3:c.406C>T (a site on the BRCA2 gene) is received, the system triggers the Retrieval Enhancement Generation (RAG) process.

[0055] The data retrieval module 200 does not simply perform keyword matching, but drives the vector retrieval engine to perform semantic mapping. It transforms the contextual information of variant sites into a 1536-dimensional dense vector and performs an approximate nearest neighbor (ANN) search in a vector database (such as Milvus or Pinecone). Here, the Top-K value is set to 10 to balance information coverage with the load on the context window. This module is specifically configured with a parser for unstructured experimental data, capable of recognizing natural language descriptions in the literature such as "enzyme activity reduced to 30% of wild type" or "yeast two-hybrid experiments showed a lack of interaction," and converting them into structured evidence fragments. The unstructured experimental data may also include expression level descriptions from Western blotting, localization descriptions from immunofluorescence experiments, and electrophysiological data from membrane potential experiments.

[0056] Furthermore, to achieve the transformation from the aforementioned natural language description to structured evidence fragments, the parser adopts a fixed two-stage information extraction process: the first stage is candidate fragment recall, which segments the retrieved literature text into sentences and locates sentences containing experimental conclusions based on a pre-defined experimental trigger dictionary (including but not limited to "% of WT", "residual activity", "yeast two-hybrid / Y2H", "interaction lost / abolished", etc.), while retaining one sentence before and after the hit sentence to form a context window; the second stage is structured extraction, which calls the large language model to output evidence_snippet according to a predefined JSON schema. The schema fixedly includes the fields of assay_type, readout, value, unit, direction, acmg_candidate, and source_span, where value and unit are normalized according to rules (e.g., "reduced to 30% of wild type" is converted to value=0.30, unit=%WT, direction=decrease; "interaction lost" is converted to direction=absent). To ensure traceability and prevent fabrication, the source_span must be a continuous segment of the original text within the input window, and null must be output for any field that cannot be determined within the input window. After extraction, the validation module performs rule-based validation on numerical range, unit mapping, negative words, and saliency statements. It also aggregates multiple results with the same assay_type / readout from the same document, marking conflict=true if conflicts exist and retaining their respective source_spans.

[0057] The dynamic implementation process of multi-agent game:

[0058] like Figure 2As shown, the core reasoning process unfolds in the multi-agent game module 300. This module instantiates three mutually exclusive agent roles: proposer agent 301, refuter agent 302, and decision-maker agent 303. To implement these mutually exclusive roles, the system configures a specific system prompt for each agent. For example, the proposer agent is configured as follows: "You are a radical geneticist, and your goal is to find as much evidence as possible to support pathogenicity, even if the evidence is weak"; the refuter agent is configured as follows: "You are a conservative reviewer, and your task is to attack the proposer's logical flaws, find benign evidence, and rigorously examine the statistical significance of the literature"; the decision-maker agent is configured as follows: "You are an impartial judge, please weigh the arguments of both sides based on the ACMG criteria, do not generate new arguments, and only make a weighted decision based on existing evidence."

[0059] First round of the game (S100):

[0060] Proposer Action: Proposer agent 301 receives the retrieved evidence and is compelled to execute the following actions: Figure 3 The thought process is illustrated below. First, it anchors two "pathogenic" submissions in ClinVar, matching them against the PP3 (Computational Predictive Support) and PS4 (Patient Case Statistics) entries in the ACMG criteria. Next, it assesses the strength of the evidence, concluding that multiple independent cases support the conclusion, and ultimately determines it as "pathogenic."

[0061] The rebuttal agent's action: The rebuttal agent 302 is immediately activated. It scans the gnomAD database and finds that the frequency of this variant in the East Asian population is 0.02% (low but present). The rebuttal agent 302 immediately constructs a rebuttal vector, pointing out that the proposer has overlooked the possibility that BS1 (population frequency is higher than expected) and that some cases cited by the proposer lack pedigree cosegregation data (insufficient evidence for PP1).

[0062] Decision-maker Action: Decision-maker agent 303 receives arguments from both sides. It uses a Bayesian weighting algorithm to weight conflicting evidence. Since crowd frequency evidence (BS1) typically has a strong veto power, decision-maker agent 303 lowers the current confidence score to 0.65 and downgrades the provisional rating to "VUS" (Various Uncertainties). Specifically, the decision-maker agent is configured to output the final conclusion in JSON format, containing the key-value pair 'confidence_score', which the inference control module parses to obtain the value.

[0063] The "Bayesian weighting algorithm" is implemented by multiplying the likelihood ratios in the dominance ratio space: Let the prior pathogenicity probability be P0, and the prior dominance ratio O0 = P0 / (1-P0). For each piece of evidence ei, its ACMG strength is mapped to a likelihood ratio LRi, and the overall dominance ratio O = O0 × ∏LRi is calculated, thus obtaining the posterior pathogenicity probability P = O / (1+O). The confidence score is defined as confidence_score = max(P, 1-P). When the population frequency evidence BS1 is valid, it is mapped as "strong benign evidence" to LRB S1 = 1 / LRstrong, so that it strongly inhibits the pathogenic direction in the product term, thereby reflecting a "strong veto power". When P≈0.35 is calculated, confidence_score = 1-P≈0.65, and the system sets the provisional rating to VUS accordingly. The decision-maker agent outputs the final conclusion in JSON format, which includes at least the key-value pair "confidence_score". The inference control module parses this field and executes a dynamic truncation rule: if confidence_score ≥ 0.95 or two consecutive provisional ratings are consistent, the subsequent rounds are immediately terminated and the final rating is output; otherwise, the next round of argumentation begins.

[0064] Iteration and Convergence (S200):

[0065] The inference control module 400 monitors the decision-maker's output in real time. Since the confidence level of 0.65 is lower than the preset threshold of 0.95, and the current round (Round 1) has not reached the maximum number of rounds (set to 3 rounds), the system triggers the next round of the game.

[0066] Second round of the game (S300):

[0067] Recursive Correction: In the thought chain generated in the second round, the proposer agent 301 must include a response to the rebuttal in the first round. It re-invokes the data retrieval module 200 to conduct in-depth analysis on the "crowd frequency" question raised by the rebuttal.

[0068] Key Finding: Proposer agent 301 successfully retrieved a PubMed paper published in 2023, which describes how this variant leads to homologous recombination repair deficiency (HRD) in a functional study. This constitutes strong evidence for PS3. Based on this, proposer agent 301 revised its argument: although the population frequency is slightly higher, the functional study confirms its disruptive nature, and this frequency may belong to a founder effect in a specific subpopulation.

[0069] The refutator compromises: Refutator agent 302 verifies the authenticity of the document (via DOI link verification) and confirms the validity of the PS3 evidence. Unable to find evidence refuting the functional experiment, it switches to "partial acceptance" in the second round, only notifying the user of the need to pay attention to the penetrance issue.

[0070] Final ruling: Decision-maker agent 303 synthesizes PS3 (strong pathogenicity), PP3 (supportive), PS4 (strong pathogenicity), and weakened BS1 evidence. According to the ACMG scoring matrix, two strong pieces of evidence plus one supporting piece of evidence are sufficient to determine "pathogenicity." At this point, the confidence level of decision-maker agent 303's output surges to 0.98.

[0071] The inference control module 400 detects a confidence level of 0.98 > 0.95, satisfying the dynamic truncation condition, and immediately terminates the iteration. The state memory module 600 encapsulates the complete dialogue history of the two rounds of game, the cited document IDs, and the thought chain derivation process into a JSON object and stores it in a vector database for auditing. Finally, the system generates a PDF report containing a complete chain of evidence.

[0072] In the validation of 500 "time-cutoff" test samples, the system of this embodiment performed as follows:

[0073] Accuracy: 92.4% (consistent with the conclusions of the human expert panel).

[0074] Hallucination rate: 0.8% (minor deviations only occurred in non-critical descriptive texts; no fabricated documents were found).

[0075] Average processing time: 145.2 seconds per mutation.

[0076] Token consumption: The average consumption per processing session is 12,500 tokens.

[0077] Convergence: 85% of the mutations converged within 2 rounds, and only 15% of the complex mutations triggered a third round of debate.

[0078] Compared to traditional rule-oriented tools (such as InterVar), this system successfully activated 68% of the PS3 (functional experiment) rules, while traditional tools only achieved 15%. This demonstrates the significant advantages of this system in understanding unstructured experimental data and eliminating logical conflicts through game theory.

[0079] Example 2

[0080] This embodiment describes a localized implementation method based on an open-source large model that emphasizes data privacy, suitable for hospital intranet environments with strict compliance requirements for patient genetic data.

[0081] The main difference from Example 1 is that the agent in this example is built based on an open-source large language model (such as Llama-3-70B-Instruct) with 70 billion parameters. Since the individual inference capability of the open-source model is slightly weaker than that of the closed-source top-level model, the parameters of the inference control module 400 are adaptively adjusted in this example.

[0082] Parameter adjustment and fault tolerance mechanism:

[0083] Temperature coefficient: set to 0.2 (0.1 in Example 1). Slightly increasing the temperature is to increase the diversity of the model in generating rebuttal arguments and prevent the open-source model from getting trapped in local optima too early.

[0084] Number of iteration rounds: The maximum number of iteration rounds is set to 5. This is to give the agent more opportunities to interact and correct logical flaws.

[0085] Dynamic truncation condition: In addition to the confidence threshold, "consistent rating conclusions in two consecutive rounds" is added as a strong termination condition. For example, if the decision-makers in both round 4 and round 5 determine it as "possibly pathogenic," the system will force convergence even if the confidence level is only 0.85, to prevent the model from endlessly looping on minor details.

[0086] When dealing with a missense mutation involving the MYH7 gene, the proposer agent initially gave a "VUS" rating in the first round because the open-source model missed crucial pedigree data during the initial search. However, the refutation agent, in the second round, triggered a recall of an older article through broader keyword fuzzy matching, pointing to potential PM2 (low population frequency) evidence. After four rounds of back-and-forth, both agents gradually pieced together the evidence, finally reaching a consensus of "potentially pathogenic" in the fifth round.

[0087] Although the processing time per cycle in this embodiment increases to 280 seconds and the token consumption rises to 22,000 tokens, its accuracy still reaches 88.5%. More importantly, all data processing is completed within the video memory of the local server 100, physically cutting off data interaction with the external network, fully meeting the data security compliance requirements of a top-tier hospital.

[0088] Example 3

[0089] This embodiment is an extension of Embodiment 1, and is designed to address undiagnosed diseases where the phenotype and genotype are not clearly related.

[0090] The multi-agent game module 300 adds two new expert roles: Phenotype Agent and Structural Biology Agent.

[0091] Phenotype expert agent: Equipped with an HPO (Human Phenotype Ontology) interface. It calculates the semantic similarity between a patient's clinical symptoms (such as "retinitis pigmentosa") and gene-related disease phenotypes. When the similarity score is >0.8, it submits PP4 strong supporting evidence to the decision-maker.

[0092] The structural biology expert agent is equipped with an AlphaFold prediction database interface. It analyzes whether the mutation site is located in the core region or active site of the protein's three-dimensional structure. If the mutation is predicted to disrupt the hydrogen bond network, it will generate supporting evidence for PM1.

[0093] The decision-making agent 303 employs a weighted voting mechanism. The base weight for proposers and detractors is 1.0, the weight for phenotype experts is set at 1.2 (because clinical phenotype matching has high specificity in diagnosis), and the weight for structural biology experts is 0.8.

[0094] In handling a complex case presenting with atypical cardiomyopathy, the conventional system classified it as VUS due to a lack of functional experimental evidence. In this embodiment, a phenotypic expert agent identified subtle features of the patient (such as mild hearing loss) that highly matched the specific syndrome, while a structural biology expert agent discovered mutations located in a highly conserved α-helix region. These two additional dimensions of evidence prompted the decision-maker to upgrade the rating to "probably pathogenic," ultimately guiding the physician in making a diagnosis. This embodiment improved the diagnostic positivity rate in the difficult case cohort by 8.2% compared to Embodiment 1.

[0095] Comparative Example 1

[0096] This comparative study aims to verify the necessity of the technical feature that "the multi-agent game module is configured with three mutually exclusive roles: proposer, refuterator, and decision-maker".

[0097] In this comparative example, the hardware environment (8x NVIDIA A100), data retrieval module (RAG+ClinVar / PubMed), and underlying large model (GPT-4 Turbo) are kept completely consistent with those in Example 1. The only variable is the removal of the multi-agent game architecture, which degenerates into a single-agent mode. This single agent is configured to execute the same thought chain (CoT) steps (retrieval-matching-evaluation-determination), but only performs "self-reflection" during inference. That is, the same model instance is responsible for both generating evidence and self-checking in the same context, and there is no independent "rebuttal" role to launch adversarial attacks.

[0098] The system received a request for the BRCA2 c.406C>T variant site mentioned in Example 1. The agent initiated a search and successfully retrieved a 2023 PubMed article containing the description "Homologous Recombination Repair Deficiency (HRD)". In the "Evaluating the Strength of Evidence" step of the thought chain, the agent established a prior belief that "this variant is pathogenic" based on the initially retrieved ClinVar "pathogenic" record. Subsequently, it attempted to find evidence in the literature to support this hypothesis. Due to the lack of an independent "refutationist" to introduce a negative feedback mechanism, the agent fell into a serious confirmation bias logical trap. Specifically, when reading the literature, the agent's attention mechanism over-focused on the conclusive statement "HRD positive," while selectively ignoring the limiting description in the literature's discussion section regarding "the experiment used an overexpression plasmid system, which may cause non-physiological interference." In the absence of an external adversarial perspective, the probability distribution within the model irreversibly collapsed towards "pathogenicity." More seriously, the model exhibited hallucination in an attempt to improve the coherence of the logical chain. In generating the final report, the agent fabricated a non-existent statistical parameter—"p < 0.001"—in an attempt to strengthen the weight of the PS3 evidence. This hallucination was not random but rather a flawed optimization by the model to minimize the perplexity between the "pathogenic hypothesis" and the "chain of evidence." Ultimately, Example 1 output a "pathogenic" rating with a confidence level as high as 0.99. However, this conclusion is logically fragile because it ignores the limitations of the experimental method (which was successfully pointed out and downgraded by the opposing agent in Example 1).

[0099] This failure process conversely demonstrates that relying solely on CoT and high-performance models cannot solve the problems of endogenous illusions and biases in large models. The "adversarial game of mutually exclusive roles" is not simply functional redundancy, but a necessary technical means to break the illusion of model self-consistency and forcibly introduce an orthogonal verification perspective. Without this feature, when processing complex evidence, the system's accuracy plummeted from 92.4% in Example 1 to 81.5%, and the illusion rate surged to 5.2%, completely compromising its safety for clinical application. Figure 4 As shown, the data points in Comparative Example 1 are in the high false alarm rate range, which cannot meet the stringent requirements of medical-grade diagnosis.

[0100] Comparative Example 2

[0101] The "rule-oriented automated tool" mentioned in the background of this comparative simulation aims to verify the significant progress of this invention in using "intelligent agent game theory to understand unstructured data" compared to traditional rule matching techniques.

[0102] This comparison example constructs an automated rating system based on hard-coded logic in Python:

[0103] Database: Consistent with Example 1, connecting ClinVar, gnomAD, and the preprocessed structured database.

[0104] Core logic: Employs strict "If-Then-Else" code logic. For example, for the BS1 rule (crowd frequency), it is set to "activate BS1 if gnomAD_AF>0.01".

[0105] Text processing: It lacks the semantic understanding capabilities of a large language model and is only equipped with a text scanner based on regular expressions and keyword matching to search for specific words (such as "pathogenic", "damaging", "activity<10%)" in PubMed summaries.

[0106] Hardware: Runs in the same server environment, but only uses CPU computing resources (since no GPU inference is required).

[0107] The system addressed the MYH7 gene missense mutation (leading to atypical cardiomyopathy) mentioned in Example 3. This mutation has an extremely low frequency in gnomAD (does not trigger BS1) and is not recorded in ClinVar. The system scan found a relevant article that described: "The mutant protein exhibited an aggregated distribution pattern in cardiomyocytes, disrupting the sarcomere structure."

[0108] Comparative Example 2's keyword scanner searches for standard terms like "loss of function" or "reduced activity" in the text. However, the literature uses the description "aggregated distribution," which biologically corresponds to "dominant negative effect," representing strong evidence for PM4 or PS3. Because the rule base doesn't pre-define a mapping between "aggregated distribution" and "PM4," and the system lacks semantic generalization ability, this crucial evidence is directly filtered out. The system detects a low population frequency (triggering PM2, moderate pathogenicity evidence), but due to the inability to extract functional experimental evidence (PS3 missing) and phenotypic matching evidence (PP4 missing), the system falls into a logical deadlock of insufficient evidence. According to the ACMG scoring matrix, PM2 alone cannot determine pathogenicity. Ultimately, Comparative Example 2 outputs a "VUS" rating. This is a typical false negative error. This failure profoundly reveals the inherent flaws of traditional technical approaches: the "limitations of pre-defined rules" cannot cover the "infinite diversity of biological descriptions."

[0109] exist Figure 4 In Comparative Example 2, the activation rate for the PS3 rule was only 15% (compared to 68% in Example 1), and the activation rate for the PP4 rule was close to 0%. This "semantic blind spot" resulted in extremely low reclassification ability for VUS. The "LLM-based agent" of this invention is not a simple text reader, but rather acts as a "semantic translator," transforming natural language experimental descriptions into structured ACMG evidence codes—a process that rule-guided systems cannot achieve at the physical level. Therefore, the feature of this invention regarding "identifying unstructured experimental data in documents" is crucial.

[0110] Comparative Example 3

[0111] This comparative example aims to verify the necessity of the "dynamic truncation condition" and demonstrate the waste of resources and performance degradation caused by excessive debate.

[0112] Keeping all configurations of Example 1 unchanged, only the dynamic truncation logic in the inference control module is removed. The system is forced to perform a full 10 rounds of proposal-refutation iterations regardless of confidence level.

[0113] In handling a well-documented benign variant, this comparative example demonstrates how redundant iterations directly undermine logical purity. Unlike Example 1, where the variant terminates precisely in the second round due to the establishment of BS1 (high frequency) and BA1 (very high frequency) evidence (taking only 45 seconds), in Comparative Example 3, after removing the truncated logic, although the system has substantially reached a consensus in the third round, it is forced to continue into a meaningless game state. Starting from the fourth round, the rebuttal agent, in order to fulfill its "rebuttal" duty, begins to launch invalid attacks on the details of the literature cited by the proposer. For example, it questions "the statistical software version used in the literature is too old" or "the latitude and longitude of the sample collection location are not marked," causing the entire reasoning chain to fall into a serious logical dead end.

[0114] As the number of iterations was forcibly increased, the focus of the dialogue gradually shifted from core biological evidence (such as enzyme activity and pedigrees) to irrelevant methodological details. This "debating for the sake of debating" introduced a large amount of noise. After receiving a large amount of conflicting information about minor details, the decision-making agent's internal attention weights were scattered, causing the confidence score, which was originally as high as 0.99, to drop to 0.85 in the 8th round. The token consumption per processing session surged to 120,000 tokens, and the processing time increased to 900 seconds. In concurrent testing, this led to server memory overflow (OOM) and an 85% drop in system throughput.

[0115] Figure 4 Data shows that the accuracy of Comparative Example 3 did not increase with the number of rounds; in fact, it decreased slightly (92.0%) due to the introduction of noise, but its resource consumption was six times that of Example 1. This conversely proves that the "dynamic truncation mechanism" is not only a means of engineering optimization, but also the core control logic for maintaining the system's "signal-to-noise ratio" in the optimal range.

[0116] Comprehensive analysis of technical effects and data correlation

[0117] Combination Figure 4Based on the experimental data and the detailed analysis of the above comparative examples, the "Automatic Rating System for Pathogenicity of Gene Mutations Based on Multi-Agent Game Theory and Thought Chain Reasoning" proposed in this invention can achieve a non-linear technological leap in solving the two major technical obstacles that have long existed in the industry: "difficulty in extracting unstructured evidence" and "uncontrollable AI illusions".

[0118] Comparing the data curves of Example 1 and Comparative Example 1 (monolithic model), it can be seen that after introducing the adversarial mechanism of "proposer-refuter," the illusion rate of the system plummeted from 5.2% to 0.8%. Mechanistically, the monolithic model (Comparative Example 1), limited by the probabilistic characteristics of the autoregressive generative model, tends to generate sycophancy text with a "high probability but potentially erroneous" nature. This invention constructs a "Nash equilibrium approximation process" at the algorithmic level through mutually exclusive roles. The existence of the refuter agent forces the proposer to provide a highly robust chain of evidence (such as specific experimental values ​​and explicit statistical parameters) to survive in the game. This mechanism transforms "accuracy" from an implicit goal of the model into an explicit game-winning condition, thereby fundamentally suppressing the generation of illusions. Furthermore, the recursive context injection mechanism forces the model to process preceding 'negative information' when generating new text. This mechanism suppresses illusionary paths that conflict with preceding rebuttal evidence in terms of probability distribution, making it impossible for the model to ignore unfavorable evidence and maintain its own narrative. This effect is not a simple superposition of models, but a qualitative change brought about by the alteration of the system's topology.

[0119] Comparing the data from Example 1 and Comparative Example 2 (rule-oriented), this invention achieves a 4.5-fold improvement in the activation rate of PS3 (functional experiment) rules (15% -> 68%). The failure of Comparative Example 2 demonstrates that keyword matching alone cannot bridge the gap between natural language descriptions and structured rules. This invention utilizes the semantic generalization capability of LLM, combined with mandatory thought chains, to precisely map ambiguous biological descriptions (such as "clustered distribution" and "structural distortion") into ACMG evidence codes. This enables this invention to address a large number of VUS cases missed by traditional tools, significantly improving the positive rate of clinical diagnosis.

[0120] Comparative example 3 shows that blindly increasing the number of inference rounds leads to negative marginal utility. The dynamic truncation mechanism proposed in this invention actually constructs an adaptive feedback loop of "confidence-computing power". For example... Figure 4 As shown, the data points in Example 1 are located in the optimal upper left corner interval of "high accuracy - low time consumption", successfully solving the computing power bottleneck problem in the application of large models.

[0121] Regarding the impact of model parameters, experimental data show that when the temperature coefficient is set in the range of 0 to 0.5, the system performance exhibits a specific trend: as the temperature coefficient increases from 0 to 0.5, the system's evidence recall gradually improves, making it suitable for challenging cases requiring extensive exploration; however, when it exceeds 0.5, the hallucination rate increases exponentially. Therefore, this system limits the temperature coefficient to 0 to 0.5 to achieve a balance between creativity and accuracy.

[0122] In summary, this invention is not a simple patchwork of existing technologies (Agent concept + rule base). Through specific adversarial game logic and dynamic control strategies, it solves the problems of "evidence conflict resolution" and "unstructured data understanding" that existing technologies cannot address. Its technical effectiveness (92.4% accuracy, 0.8% illusion rate) far exceeds what those skilled in the art could expect based on existing technologies, demonstrating outstanding substantive features and significant progress.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automated rating method for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, characterized in that, Includes the following steps: Data retrieval steps: Retrieve data related to the target variant site; Multi-agent game steps: Based on the relevant data, the proposer agent, the refutator agent, and the decision-maker agent conduct multiple rounds of iterative rating of the pathogenicity of gene mutations. Each of the proposer agent, the refutator agent, and the decision-maker agent is configured with a unified thought chain structure, which includes the following steps executed sequentially: retrieving evidence, matching ACMG evidence entries, evaluating the strength of the evidence, and making a comprehensive judgment. Each round of the game includes: Based on the currently retrieved chain of evidence, the proposing agent generates a first argument demonstrating that the target variant site is pathogenic. The rebuttal agent searches for benign evidence and / or points out logical flaws in the first argument based on the relevant data, and generates a second argument refuting the pathogenicity of the target mutation site. The decision-maker agent performs a weighted calculation by combining the first argument and the second argument, and outputs the rating conclusion and corresponding confidence score for the current round; wherein, in the N+1 round of the game, the first argument generated by the proposer agent includes the response to the second argument in the Nth round, so as to form a self-correcting feedback loop, where N is a positive integer; Reasoning control steps: Monitor the rating conclusion and confidence score output by the decision-maker agent in real time. When the preset dynamic truncation condition is met, terminate the multi-agent game step and output the final gene mutation pathogenicity rating conclusion. The preset dynamic truncation condition includes: the confidence score in the current round is higher than a preset threshold, and / or the rating conclusions in a consecutive preset number of rounds are consistent.

2. The method according to claim 1, characterized in that, The data retrieval step further includes: identifying unstructured experimental data in the retrieved relevant data; parsing the natural language description of the unstructured experimental data using a large language model, and converting it into structured ACMG evidence fragments for use in the multi-agent game step.

3. The method according to claim 1, characterized in that, The first argument includes evidence entries supporting the pathogenicity of ACMG found by the proposing agent; the second argument includes evidence entries supporting the benign nature of ACMG found by the rebutting agent, and / or, a review of the statistical significance of the literature cited in the first argument.

4. The method according to claim 1, characterized in that, The multi-agent game step also introduces phenotypic expert agents and structural biology expert agents to participate in the game; wherein, the phenotypic expert agent is used to calculate the semantic similarity between the patient's clinical symptoms and the phenotype of the gene-related disease, and submits phenotypic supporting evidence to the decision-maker agent when the similarity is higher than a preset value; the structural biology expert agent is used to analyze whether the target mutation site destroys the three-dimensional structural core region or active site of the protein, and submits structural supporting evidence to the decision-maker agent when destruction is confirmed.

5. The method according to claim 4, characterized in that, The decision-maker agent uses a weighted voting mechanism or a Bayesian weighting algorithm to assign different weights to the arguments and evidence output by the proposer agent, the refuter agent, the phenotypic expert agent, and the structural biology expert agent, and to comprehensively calculate the rating conclusion and corresponding confidence score for the current round.

6. The method according to claim 1, characterized in that, It also includes: State memory step: using a vector database to store the intermediate reasoning state, dialogue history, cited literature identifiers and thought chain derivation process of each round of the game, and generating an automated rating report containing a complete chain of evidence based on the stored data.

7. An automated rating system for the pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning, characterized in that, include: The data retrieval module is configured to perform the data retrieval steps as described in any one of claims 1 to 6; A multi-agent game module, configured to execute the multi-agent game steps as described in any one of claims 1 to 6; The inference control module is configured to perform the inference control steps as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the automated rating method for pathogenicity of gene mutations based on multi-agent game theory and thought chain reasoning as described in any one of claims 1 to 6.

Images