A biological mutation pathogenicity evaluation method based on multi-agent collaborative reasoning

Abstract

The application provides a biological variation pathogenicity evaluation method based on multi-agent collaborative reasoning, relates to the technical field of natural language processing, and comprises the following steps: obtaining a variation sample to be analyzed, performing standardization preprocessing on the variation sample, and obtaining a standard input; at least three functionally different agents are constructed; the standard input is obtained in each agent to obtain reasoning results and is uniformly converted into a structured intermediate evidence representation; a divergence score, a harmfulness score and a confidence score are obtained according to each intermediate evidence; and the target feature vector is obtained by fusing the basic features. The beneficial effect of the application is that the multi-functionally different agents are sequentially and collaboratively reasoned, the specialization division of labor from gene function analysis, calculation evidence integration to clinical risk judgment is realized, and the comprehensiveness and analysis depth of variation evaluation are improved.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and in particular relates to a method for assessing the pathogenicity of biological variations based on multi-agent collaborative reasoning. Background Technology

[0002] Assessing the pathogenicity of biological variants is a crucial step in the molecular diagnostics and genomic medicine practice for rare diseases. The analysis requires a comprehensive analysis of heterogeneous information from multiple sources, including the biological function of the gene, the genomic form of the variant, the pathogenesis of the disease, supporting evidence from existing literature and databases, patient phenotypic characteristics, and clinical experience, in order to make a reasonable judgment on the pathogenic risk of a specific variant.

[0003] Existing assessment techniques mainly include: computationally predictive scoring tools that provide quantitative scores by analyzing features such as sequence conservation and protein structural changes; rule-based expert knowledge systems that classify and judge based on preset classification criteria and evidence weights; and methods that generate assessment opinions using large-scale language models.

[0004] The aforementioned existing technologies have the following shortcomings: First, in terms of reasoning architecture, existing methods combine work from different professional dimensions, such as genetic mechanism analysis, computational evidence integration, and clinical judgment, into a single reasoning process. This limits the depth of analysis in each dimension and results in insufficient professional relevance of the evaluation results. Second, in terms of output format, evaluation results are mostly presented in natural language or as single numerical scores, lacking standardized intermediate representations for procedural processing, making it difficult to directly input structured features into downstream machine learning models or automated decision-making systems. Third, in terms of computational efficiency, when the same gene and variant type combination is repeatedly evaluated in different analysis batches, the entire reasoning or computation process must be executed from scratch each time, lacking means to reuse existing results, and computational resource consumption increases with the scale of evaluation. Fourth, in terms of quantifying discrepancies, different professional perspectives may lead to different risk assessments of the same variant, but existing methods lack a mechanism for numerically measuring such differences in cross-perspective assessments, making it difficult to prioritize reviews and assess the reliability of evidence accordingly. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention provides a method for assessing the pathogenicity of biological variants based on multi-agent collaborative reasoning. This method is particularly suitable for use in scenarios such as rare disease variant screening, large-scale genomic variant annotation, and clinical decision support. It can transform the results of multi-source expert reasoning into a computable structured feature representation for downstream risk modeling and prioritization.

[0006] The technical solution adopted in this invention is: to provide a method for assessing the pathogenicity of biological variations based on multi-agent collaborative reasoning, comprising the following steps: Obtain the variant sample to be analyzed, and perform standardized preprocessing on the variant sample to obtain standard input; the variant sample includes at least gene symbol, variant type and disease context information; Construct at least three intelligent agents with different functions, wherein the intelligent agents include at least a first intelligent agent for gene function and pathogenesis mechanism analysis, a second intelligent agent for computational evidence integration analysis, and a third intelligent agent for clinical evidence integration; The standard input is input into the first intelligent agent to generate the first structured intermediate evidence, which includes at least the gene function analysis results and the preliminary assessment results of the harmfulness of the variation. The first structured intermediate evidence and the standard input are input into the second agent, and the second structured intermediate evidence is generated based on the first structured intermediate evidence. The second structured intermediate evidence includes at least a risk level and a confidence score. The first structured intermediate evidence, the second structured intermediate evidence, and the standard input are input into the third intelligent agent. The third structured intermediate evidence is generated based on the first structured intermediate evidence and the second structured intermediate evidence. The third structured intermediate evidence includes at least a comprehensive risk judgment and a final evidence chain. The risk judgment results of each agent are obtained from the first structured intermediate evidence, the second structured intermediate evidence and the third structured intermediate evidence respectively. The risk judgment results are mapped to numerical values ​​and the dispersion is calculated as a divergence score. Harmfulness score and confidence score are extracted from the third structured intermediate evidence; The basic features generated based on the gene symbol and the mutation type are fused with the harmfulness score, confidence score and divergence score to obtain the target feature vector.

[0007] Furthermore, it also includes the following steps: inputting the target feature vector into the downstream risk assessment model and outputting a variation risk label or risk priority.

[0008] Furthermore, the variant samples are from a publicly available variant database, and the standardization preprocessing includes mapping the original records of different formats in the variant samples to uniform standard fields.

[0009] Furthermore, the first structured intermediate evidence also includes information on the mechanism of variation and the chain of evidence; the second structured intermediate evidence also includes key supporting evidence and evidence consistency judgment; and the third structured intermediate evidence also includes alternative interpretations and review suggestions.

[0010] Furthermore, the specific process of mapping each of the risk judgment results to numerical values ​​includes: calculating the dispersion through the standard deviation; converting the risk judgment result of the first agent into a risk level based on the initial assessment result of the variability harmfulness in the first structured intermediate evidence through a segmentation rule; and mapping the risk level of each agent to the corresponding numerical value.

[0011] Furthermore, the basic feature is generated by: performing one-hot encoding on the gene symbol, tag encoding on the mutation type, and concatenating the encoding results to form the basic feature.

[0012] Furthermore, when multiple variant samples need to be processed, a cache reuse step is also included: the generated structured intermediate evidence and scores are cached using the gene symbol and the variant type as identifiers; when the gene symbol and the variant type of a subsequent variant sample are the same as the identifier of an existing cached record, the target feature vector is generated directly using the data stored in the corresponding cached record.

[0013] Furthermore, the cached record stores the first structured intermediate evidence, the second structured intermediate evidence, the third structured intermediate evidence, and the corresponding harmfulness score, confidence score, and divergence score.

[0014] Furthermore, it also includes a rule completion step: for the variant samples for which no structured intermediate evidence has been generated, a default toxicity score and confidence score are generated using preset rules, and the samples are marked as samples to be reviewed.

[0015] Furthermore, it also includes a priority marking step: when the divergence score of the variant sample generated through the complete inference process exceeds a preset threshold, the corresponding variant sample is marked as a priority review sample.

[0016] The advantages and positive effects of this invention are as follows: By adopting the above-mentioned technical solution, multiple intelligent agents with different functions conduct sequential collaborative reasoning, realizing a professional division of labor from gene function analysis and computational evidence integration to clinical risk assessment, thereby improving the comprehensiveness and depth of variant assessment; the reasoning results of each intelligent agent are uniformly transformed into a structured intermediate evidence representation, which facilitates procedural parsing and direct use by downstream models, reducing the difficulty of system integration; the harmfulness score, confidence score, and divergence score are explicitly extracted, and the results of multiple rounds of reasoning are distilled into computable numerical features, enhancing the ability of downstream risk modeling. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for assessing the pathogenicity of biological variations according to an embodiment of the present invention. Detailed Implementation

[0018] The present disclosure will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments of the present disclosure. The technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative effort are within the scope of protection of the present disclosure.

[0019] like Figure 1 As shown, this invention provides a method for assessing the pathogenicity of biological variations based on multi-agent collaborative reasoning, comprising the following steps: S10. Obtain the variant sample to be analyzed, perform standardized preprocessing on the variant sample to obtain standard input; the variant sample should include at least gene symbol, variant type and disease context information. Specifically, the process involves acquiring the variant sample data to be analyzed. A variant sample refers to a record carrying gene variation information, containing at least three dimensions of information: gene symbol, i.e., the standard naming identifier of the gene where the variant is located (e.g., abbreviation or number of the gene name); variant type, i.e., the specific form of the variant at the genome level (e.g., missense mutation, nonsense mutation, splice site variation, insertion, deletion, etc.); and disease context information, i.e., a description of the disease background that may be related to the variant (e.g., disease name, known pathogenic mechanisms, related phenotypic characteristics, etc.). After acquiring this information, the sample undergoes standardized preprocessing, which involves uniformly mapping the text, terminology, and annotation formats from the original records from different sources into structured standard fields, forming a standardized input data that can be directly parsed by subsequent modules.

[0020] Gene symbols: These include, but are not limited to, standard gene nomenclature (such as the official HGNC symbol), gene identification numbers (such as Ensembl gene ID, NCBI Gene ID), or other naming methods that can uniquely identify the target gene. Gene symbols are used to determine the gene location of the variant and the corresponding gene functional context. Variance types: These include, but are not limited to, various genomic-level variations such as missense variants, nonsense variants, splice site variants, frameshift variants, synonymous variants, in-frame insertions, in-frame deletions, copy number variants, and regulatory region variants. Variance type determines the way and extent to which the variant may affect gene function. Disease context information: This includes, but is not limited to, the disease name, known descriptions of pathogenic mechanisms, related phenotypic characteristics, inheritance patterns (such as autosomal dominant, autosomal recessive, X-linked, etc.), known genotype-phenotype associations, necessary background knowledge, and other background information that helps in understanding the clinical significance of the variant.

[0021] S20. Construct at least three intelligent agents with different functions, including at least a first intelligent agent for gene function and pathogenesis mechanism analysis, a second intelligent agent for computational evidence integration analysis, and a third intelligent agent for clinical evidence integration. Specifically, at least three agent modules with different functional roles are constructed. An agent is a computational module with independent reasoning capabilities, capable of receiving structured input and generating structured output. The three agents perform the following roles: The first agent is specifically responsible for gene function and pathogenic mechanism analysis. Its task is to analyze the biological function of the target gene and the possible pathogenic mechanism caused by the mutation based on gene function knowledge, and to provide an initial assessment of the degree of harm of the mutation; the second agent is specifically responsible for computational evidence integration analysis. Its task is to receive the output of the first agent, combine computational evidence such as sequence conservation and protein structure prediction, and output a risk level and confidence score; the third agent is specifically responsible for clinical evidence integration. Its task is to synthesize the outputs of the first two agents and disease context information to complete a comprehensive risk assessment for clinical applications and generate a complete chain of evidence.

[0022] The implementation of intelligent agents includes, but is not limited to, the following types: reasoning modules based on large language models (defining roles and tasks through prompt word engineering and generating structured outputs using the reasoning capabilities of large language models), rule-based reasoning engines (performing reasoning based on predefined rules and knowledge bases), query and reasoning modules based on knowledge graphs (retrieving relevant information from structured knowledge graphs and performing reasoning), or a combination of the above methods.

[0023] S30. Input the standard input into the first intelligent agent to generate the first structured intermediate evidence. The first structured intermediate evidence includes at least the gene function analysis results and the preliminary assessment results of the mutation harmfulness. S40. Input the first structured intermediate evidence and the standard input into the second agent, and generate the second structured intermediate evidence based on the first structured intermediate evidence. The second structured intermediate evidence includes at least the risk level and confidence score. S50. Input the first structured intermediate evidence, the second structured intermediate evidence, and the standard input into the third intelligent agent, and generate the third structured intermediate evidence based on the first structured intermediate evidence and the second structured intermediate evidence. The third structured intermediate evidence includes at least a comprehensive risk judgment and a final evidence chain. S60. Obtain the risk judgment results of each agent from the first structured intermediate evidence, the second structured intermediate evidence and the third structured intermediate evidence respectively, map each risk judgment result to a numerical value and calculate the dispersion as a divergence score. Specifically, risk assessment results for each of the three agents are extracted from the structured intermediate evidence generated by each agent. The risk level or risk assessment information output by each agent is mapped to a calculable numerical value (e.g., low risk mapped to 1, medium risk to 2, and high risk to 3). The dispersion between these values ​​is then calculated, and this dispersion is used as a divergence score. The divergence score reflects the degree of consistency among different agents in their risk assessments: a smaller value indicates greater consistency, while a larger value indicates more significant divergence.

[0024] S70. Extract the harmfulness score and confidence score from the third structured intermediate evidence. The harmfulness score is a quantitative description of the pathogenicity risk of the variant, and the confidence score is a quantitative description of the reliability of the judgment. S80. The basic features generated based on gene symbols and mutation types are fused with the harmfulness score, confidence score and divergence score to obtain the target feature vector.

[0025] Specifically, basic features are generated by numerically encoding them based on two dimensions: gene symbol and mutation type. The encoded results are then concatenated into a feature vector. This basic feature vector is then fused and concatenated with harm score, confidence score, and divergence score to form the final target feature vector. This target feature vector condenses the structured results of multi-round agent reasoning into a set of numerical features that can be directly used for computation.

[0026] The beneficial effects of the above method are as follows: By having multiple agents with different functions conduct sequential collaborative reasoning, a specialized division of labor is achieved from gene function analysis and computational evidence integration to clinical risk assessment, improving the comprehensiveness and depth of variant assessment; the reasoning results of each agent are uniformly transformed into a structured intermediate evidence representation, which facilitates procedural parsing and direct use by downstream models, reducing the difficulty of system integration; the harmfulness score, confidence score, and divergence score are explicitly extracted, and the results of multiple rounds of reasoning are distilled into computable numerical features, enhancing the ability of downstream risk modeling; the divergence score can quantify the degree of consistency of judgments among different agents, providing an objective basis for subsequent review priority ranking.

[0027] In one embodiment, after obtaining the target feature vector, it is passed as input features to a downstream risk assessment model. The downstream model refers to a trained machine learning or statistical model capable of inference calculations based on the input features. Optional model types include gradient boosting trees, logistic regression, random forests, lightweight neural networks, or other models suitable for structured features. Based on the basic encoded features contained in the target feature vector and the harmfulness score, confidence score, and divergence score obtained through distillation, the downstream model, after inferring model parameters, outputs variant risk labels (e.g., classification results such as "high risk, medium risk, low risk") or risk priority ranking (i.e., arranging multiple variants according to their risk level to help determine the order of review or clinical intervention).

[0028] In one embodiment, the sources of variant samples include, but are not limited to, publicly available variant databases (such as ClinVar, dbSNP, gnomAD, etc.), variant reports generated by clinical testing institutions, variant annotation data in research projects, or other legal and compliant variant data sources. The specific operation of the standardization preprocessing is to uniformly map the text, terminology, and annotation information from raw variant records from different databases or in different formats to consistent standard fields. Standard fields include at least the gene symbol, variant type, disease context, candidate functional impact, and reference knowledge source. This mapping process eliminates parsing barriers caused by format differences between different data sources, ensuring that the subsequent agent can receive input data with a uniform format.

[0029] In one embodiment, the structured intermediate evidence output by each agent contains richer information dimensions. The first structured intermediate evidence, in addition to gene function analysis results and preliminary assessments of variant harmfulness, includes the variant's mechanism of action—a description of how the variant affects protein function or gene expression at the molecular level—and the chain of evidence information—an ordered record of the sources of evidence and reasoning paths supporting the above analytical conclusions. The second structured intermediate evidence, in addition to risk level and confidence score, includes key supporting evidence—evidence items that play a crucial supporting role in the risk level judgment—and a consistency assessment—an evaluation of whether the currently calculated evidence is consistent with the genetic analysis of the first agent. The third structured intermediate evidence, in addition to the comprehensive risk assessment and the final chain of evidence, includes alternative explanations—an explanation of other reasonable explanations for the pathogenicity of the variant—and review recommendations—further review directions or suggestions for uncertain parts of the analysis results.

[0030] In one embodiment, the specific process of mapping the risk assessment results of each agent to numerical values ​​and calculating the degree of divergence is as follows: First, the risk levels are uniformly mapped to numerical representations, where low risk is mapped to numerical value 1, medium risk to numerical value 2, and high risk to numerical value 3. For the first agent, since its direct output is the initial harm assessment result (a numerical score), it needs to be converted into the corresponding risk level through segmentation rules. The specific segmentation rules are: when the harm score is greater than or equal to 8, it is judged as high risk (corresponding to numerical value 3); when the harm score is greater than or equal to 5 and less than 8, it is judged as medium risk (corresponding to numerical value 2); when the harm score is less than 5, it is judged as low risk (corresponding to numerical value 1). The risk levels output by the second and third agents are directly converted into numerical values ​​according to the above mapping relationship. After obtaining the numerical values ​​corresponding to the three agents, the standard deviation of these three numerical values ​​is calculated as the degree of divergence score. The formula for calculating the standard deviation is the square root of the average of the squares of the differences between each numerical value and the mean. When the three agents' judgments are completely consistent, the standard deviation is 0; when the difference in judgment increases, the standard deviation increases accordingly.

[0031] In one embodiment, the basic features are generated as follows: For gene symbols, one-hot encoding is used, that is, each possible gene symbol is assigned a binary vector, where only the corresponding position is 1 and the rest are 0, thus converting the discrete gene symbols into a numerical vector representation. For mutation types, label encoding is used, that is, each mutation type is mapped to a unique integer number (e.g., missense mutation is encoded as 0, nonsense mutation is encoded as 1, and so on). The one-hot encoded gene symbol vector and the label-encoded mutation type numerical value are concatenated together in sequence to form a complete basic feature vector.

[0032] In one embodiment, when the system needs to process multiple variant samples in batches, a cache reuse mechanism is introduced to reduce the computational cost of repeated inference. Specifically, a combination of gene symbol and variant type fields is used as the cache identifier key. After each complete agent inference process (steps S30-S70) is executed, the structured intermediate evidence and related scoring results generated by that inference are stored in the cache using this identifier key as an index. When subsequent variant samples to be processed have the same combination of gene symbol and variant type, the system directly reads the previously generated inference results from the cache, skipping the repeated agent inference process and directly using them to generate the target feature vector. Only when the corresponding record does not exist in the cache is the complete three-agent sequential inference process executed, and the results are written to the cache for subsequent reuse after inference. The cache identifier key is composed of a combination of gene symbol and variant type fields. The definition of the cache key includes, but is not limited to, a tuple (gene_symbol, variant_type), a hash value of the concatenated string "gene_symbol|variant_type", or other methods that can uniquely identify the gene-variant type combination.

[0033] In one embodiment, the cache record stores not only the structured intermediate evidence generated by each of the three agents (first structured intermediate evidence, second structured intermediate evidence, and third structured intermediate evidence), but also the harmfulness score, confidence score, and divergence score extracted from the third structured intermediate evidence. In other words, the complete inference chain intermediate products and the final distillation result are persistently stored, and the entire set of data can be reused directly without any recalculation when the cache is hit.

[0034] In one embodiment, the rule completion step is used to handle mutated samples that are unsuitable for entering the complete multi-agent inference process (steps S30-S70). When some samples fail to undergo complete three-agent sequential inference due to incomplete data, abnormal format, or other technical limitations, the system uses preset fixed rules to generate default hazard and confidence scores for these samples. For example, the hazard score can be set to a neutral default value, and the confidence score can be set to a lower default value to reflect the uncertainty of the sample not having undergone complete inference. At the same time, these samples are marked as samples to be reviewed, indicating that subsequent manual intervention is required or the complete inference process should be re-executed when conditions permit.

[0035] In one embodiment, the priority marking step is performed after the complete inference process is completed. When the divergence score obtained by a variant sample through complete inference by three agents exceeds a preset threshold, it indicates that there is significant disagreement among the different agents regarding the sample's judgment, meaning that the uncertainty of its risk assessment result is high. At this time, the system marks the variant sample as a priority review sample, placing it in a higher priority position in subsequent manual review or clinical decision-making stages so that relevant professionals can conduct detailed reviews first. The preset threshold can be flexibly set and adjusted according to the actual application scenario and review resources.

[0036] In one embodiment, the prompts for the three agents predefine their respective roles, task scopes, and output constraints. The prompt for the genetics expert agent specifies four output fields: gene_function, pathogenic_mechanism, harm_score, and clinical_significance, and requires output in plain JSON format. The prompt for the bioinformatics analysis agent specifies three output fields: risk_level, key_features, and confidence. The prompt for the clinical expert agent specifies four output fields: top_disease, differential_diagnosis, final_risk_assessment, and evidence_chain. Each prompt directly defines the field names, ensuring consistent output structure across different agents and automatic parsing by downstream modules.

[0037] In one embodiment, ClinVar labels (i.e., pathogenicity classification labels already existing in the ClinVar database) are not included in the prompt word input of the large language model, nor are they part of the cache key. ClinVar labels are only used to evaluate the predictive performance of the downstream model during the external validation phase. Specifically, ClinVar significance is mapped to three classification labels: High, Medium, and Low, which serve as the target variable for external validation.

[0038] In one embodiment, in an experiment based on an HLH (hemophagocytic lymphohistiocytosis) scenario, 33 HLH core genes covered 170 core evidence units. These core genes generated fine-grained structured intermediate evidence through a complete three-agent reasoning process. For non-core genes not within the scope of the core genes, rule-based filling was used as a coverage layer. That is, conservative default scores were generated through preset rules, rather than performing complete agent reasoning. This hierarchical processing strategy ensures the accuracy of core gene evaluation while achieving basic coverage of variations across the entire genome.

[0039] In one embodiment, internal validation uses the risk_level generated by a bioinformatics analysis agent as the target variable to evaluate the model's fit to known inference results. External validation maps ClinVar significance (clinical significance classification) to three-category labels: High, Medium, and Low, to evaluate the model's generalization ability on independent external data. This dual validation strategy ensures both internal consistency and external reliability of the model.

[0040] To better realize a biological variant pathogenicity assessment method based on multi-agent collaborative reasoning, this application also provides a biological variant pathogenicity assessment system based on multi-agent collaborative reasoning, comprising: The data acquisition and preprocessing module is used to acquire variant samples to be analyzed from external data sources and to perform standardized preprocessing on the samples. This module includes a data source interface submodule (for connecting to data sources of different formats), a terminology mapping submodule (for unifying terms from different sources into standard expressions), and a field standardization submodule (for mapping raw records into a unified standard field structure). This module outputs standardized input data in a regulated format and passes it to subsequent modules.

[0041] The Agent Construction and Management module is used to build and manage at least three agents with different functions. This module includes an Agent Configuration submodule (for loading prompts, role definitions, and task constraints for each agent), an Agent Instantiation submodule (for creating running instances of agents, supporting different implementations based on large language models, rule engines, or knowledge graphs), and an Agent Lifecycle Management submodule (for managing the initialization, operation, and destruction of agents). This module ensures that each agent runs according to predefined functional roles and output format requirements.

[0042] The sequential reasoning execution module is used to call each agent in a fixed order to execute the complete reasoning process. This module internally includes a reasoning scheduling submodule (for controlling the order in which agents are called and data transfer), an input assembly submodule (for assembling the output of the previous agent and the standard input into the input of the next agent), and a reasoning state management submodule (for recording intermediate states during the reasoning process). This module outputs the structured intermediate evidence generated by each of the three agents.

[0043] The evidence distillation and feature extraction module is used to extract hazard scores, confidence scores, and divergence scores from structured intermediate evidence. This module internally includes a score extraction submodule (for parsing hazard and confidence scores from third-party structured intermediate evidence), a risk level mapping submodule (for converting the risk judgment results of each agent into a unified numerical representation through segmentation rules and numerical mapping), and a divergence calculation submodule (for calculating the dispersion as the divergence score based on the mapped numerical values). The module outputs a distillation feature vector containing the hazard score, confidence score, and divergence score.

[0044] The basic feature generation and feature fusion module is used to generate basic features and perform feature fusion. This module contains an encoding submodule (for one-hot or equivalent encoding of gene symbols and label encoding or equivalent encoding of variant types) and a fusion submodule (for concatenating the basic feature vector and the distilled feature vector into the target feature vector). This module outputs the final target feature vector.

[0045] The cache management module manages the caching of structured intermediate evidence and scoring results. Internally, this module includes a cache key generation submodule (for generating cache identifier keys based on gene symbols and variant types), a cache query submodule (for determining whether the current sample hits an existing cached record), and a cache write submodule (for writing newly generated inference results to the cache). When a cache hit occurs, this module directly provides the cached data for downstream use, skipping the sequential inference execution module. Cache storage media include, but are not limited to, memory, local files, databases, or distributed caching services.

[0046] The rule completion module is used to complete rules for variant samples that have not entered the complete inference process. When it is detected that some samples have failed to complete the complete inference due to data anomalies or other technical limitations, this module generates default harm scores and confidence scores using preset rules, and adds a "to be reviewed" mark to such samples.

[0047] The priority marking module is used to check whether the divergence score exceeds a preset threshold after the complete inference is completed. When the divergence score exceeds the threshold, the corresponding sample is marked as a priority review sample.

[0048] The downstream model interface module is used to pass the target feature vector to the downstream risk assessment model and receive the risk label or risk priority output by the model. This module supports integration with different types of downstream models, providing a unified feature input interface and result output interface.

[0049] This embodiment provides a preferred implementation method, and the specific parameters, model selection and implementation details of this preferred implementation method are described in detail below.

[0050] Samples to be analyzed are obtained from public variant resources. These samples contain gene symbols (gene_symbol), variant types (variant_type), and disease context information. The input is standardized to form a standard input (xi). Supplementary lightweight external knowledge feature sources include HPO (Human Phenotypic Ontology), ClinGen (Clinical Genome Resources Consortium), and Orphadata (Rare Disease Database).

[0051] The experiment used a GLM-4.7-based large language model and implemented sequential collaboration among three agents via API calls to the ZhipuAI interface. The task definitions for the three agents were pre-written using pre-defined prompts. Each agent directly specified its output as pure JSON and field names within the prompts, ensuring structured and consistent output. These agents corresponded to the following roles: a genetics expert agent (first agent), a bioinformatics analysis agent (second agent), and a clinical expert agent (third agent).

[0052] Inputting the standard input xi into the genetics expert agent outputs the first structured intermediate evidence ziG, which includes at least gene function explanation, variation mechanism, initial harmfulness assessment, and supporting evidence chain. Inputting ziG and xi into the bioinformatics analysis agent outputs the second structured intermediate evidence ziB, which includes at least risk level (low, medium, and high), key characteristics, confidence score, and evidence consistency assessment. Inputting ziG, ziB, and xi into the clinical expert agent outputs the third structured intermediate evidence ziC, which includes at least a comprehensive risk assessment, alternative explanations, final evidence chain, and review recommendations. All three agents' outputs use a unified JSON format.

[0053] Harmfulness score extracted from third-structured intermediate evidence (ziC) and confidence score The hazard rating in the prompt is defined as an integer subjective hazard score from 0 to 10, where 10 represents the strongest pathogenicity. This score is not a clinically calibrated continuous score, but rather a discrete risk intensity description given by the agent based on structured evidence.

[0054] The process of calculating the degree of divergence is as follows: The risk level represented by the genetic expert is converted from the harmfulness score through a segmentation rule. When riG is greater than or equal to 8, riG = 3 (high risk). When riG is greater than or equal to 5 and less than 8, riG = 2 (medium risk). When the risk level is less than 5, riG = 1 (low risk). The risk levels of bioinformatics analysis agents and clinical expert agents are denoted as riB and riC, respectively. After mapping low, medium, and high to 1, 2, and 3, respectively, the dissonance score is defined as... =Std(riG, riB, riC), which represents the standard deviation of the three values. When the three agents agree... =0; when the difference in judgment increases The corresponding increase.

[0055] The basic features are generated as follows: `pd.get_dummies` is used to generate one-hot encodings (dummy variables) for `gene_symbol`, and `LabelEncoder` is used to encode the `variant_type`. These are then concatenated with other numerical features to form the model input. The basic features are represented as follows: Distillation characteristics The final feature after fusion is represented as .

[0056] A cache key ki=(gi, vi) is established for combinations of identical gene symbols and variant types. The cache key consists only of two dimensions: gene symbol and variant type. ClinVar tags and disease context are not included in the cache key. The caching strategy is a tuple-level cache reuse, aiming to reuse generated structured evidence for recurring gene-variant tuples. When a cache key is matched, the corresponding structured intermediate evidence and scoring results are directly read; when a cache key is not matched, the complete three-agent sequential reasoning is executed, and the generated results are written to the cache.

[0057] The downstream modeling master model uses LightGBM, and GroupKFold is used for grouped cross-validation based on gene_symbol. The goal of the internal task is to analyze the risk_level generated by the bioinformatics agent; in the external validation, ClinVar significance is mapped to three-class labels (High, Medium, Low) for evaluation.

[0058] The experimental data included 18,499 variant records, 353 unique evidence units, and 170 core evidence units covered by 33 HLH core genes. The training strategy can be summarized as follows: LightGBM was used as the downstream model, gene_symbol was used for grouping and validation, ClinVar three-class classification mapping was used as the external label, and three-agent distillation features were used as augmented inputs.

[0059] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A method for assessing the pathogenicity of biological variations based on multi-agent collaborative reasoning, characterized in that, Includes the following steps: Obtain the variant sample to be analyzed, and perform standardized preprocessing on the variant sample to obtain standard input; the variant sample includes at least gene symbol, variant type and disease context information; Construct at least three intelligent agents with different functions, wherein the intelligent agents include at least a first intelligent agent for gene function and pathogenesis mechanism analysis, a second intelligent agent for computational evidence integration analysis, and a third intelligent agent for clinical evidence integration; The standard input is input into the first intelligent agent to generate the first structured intermediate evidence, which includes at least the gene function analysis results and the preliminary assessment results of the harmfulness of the variation. The first structured intermediate evidence and the standard input are input into the second agent, and the second structured intermediate evidence is generated based on the first structured intermediate evidence. The second structured intermediate evidence includes at least a risk level and a confidence score. The first structured intermediate evidence, the second structured intermediate evidence, and the standard input are input into the third intelligent agent. The third structured intermediate evidence is generated based on the first structured intermediate evidence and the second structured intermediate evidence. The third structured intermediate evidence includes at least a comprehensive risk judgment and a final evidence chain. The risk judgment results of each agent are obtained from the first structured intermediate evidence, the second structured intermediate evidence and the third structured intermediate evidence respectively. The risk judgment results are mapped to numerical values ​​and the dispersion is calculated as a divergence score. Harmfulness score and confidence score are extracted from the third structured intermediate evidence; The basic features generated based on the gene symbol and the mutation type are fused with the harmfulness score, confidence score and divergence score to obtain the target feature vector.

2. The method according to claim 1, characterized in that, It also includes the following steps: The target feature vector is input into the downstream risk assessment model, which outputs a variation risk label or risk priority.

3. The method according to claim 1 or 2, characterized in that: The variant samples are from a public variant database, and the standardization preprocessing includes mapping the original records of different formats in the variant samples to uniform standard fields.

4. The method according to claim 1 or 2, characterized in that: The first structured intermediate evidence also includes information on the mechanism of variation and the chain of evidence; the second structured intermediate evidence also includes key supporting evidence and evidence consistency judgment; the third structured intermediate evidence also includes alternative interpretations and review suggestions.

5. The method according to claim 1 or 2, characterized in that, The specific process of mapping each of the risk assessment results to numerical values ​​includes: calculating the dispersion through the standard deviation; converting the risk assessment results of the first agent into risk levels based on the initial assessment results of the harmfulness of the variation in the first structured intermediate evidence through segmentation rules; and mapping the risk levels of each agent to corresponding numerical values.

6. The method according to claim 1 or 2, characterized in that, The basic feature is generated by: performing one-hot encoding on the gene symbol, tag encoding on the mutation type, and concatenating the encoding results to form the basic feature.

7. The method according to claim 1 or 2, characterized in that, When multiple variant samples need to be processed, a cache reuse step is also included: the generated structured intermediate evidence and scores are cached using the gene symbol and the variant type as identifiers; when the gene symbol and the variant type of a subsequent variant sample are the same as the identifier of an existing cached record, the target feature vector is generated directly using the data stored in the corresponding cached record.

8. The method according to claim 7, characterized in that, The cached record stores the first structured intermediate evidence, the second structured intermediate evidence, the third structured intermediate evidence, and the corresponding harmfulness score, confidence score, and divergence score.

9. The method according to claim 1 or 2, characterized in that, It also includes a rule completion step: for the variant samples for which no structured intermediate evidence has been generated, a default hazard score and confidence score are generated using preset rules, and the samples are marked as samples to be reviewed.

10. The method according to claim 9, characterized in that, It also includes a priority marking step: when the divergence score of the variant sample generated through the complete inference process exceeds a preset threshold, the corresponding variant sample is marked as a priority verification sample.

Images