A biomedical knowledge discovery method and system
By integrating multiple XAI methods and a large language model multi-agent architecture, combined with the Boda counting method and expert agent evaluation, the problem of low efficiency in traditional biomedical knowledge discovery systems is solved, achieving efficient and reliable knowledge discovery and hypothesis generation.
Patent Information
- Application Number
- CN202511726987.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-11-06
- Estimated Expiration
- 2045-11-24
AI Technical Summary
Existing biomedical knowledge discovery systems rely on traditional bioinformatics methods and pipeline architectures, which suffer from inefficiency and insufficient knowledge integration, failing to fully utilize the potential of explainable artificial intelligence (XAI) and large language models (LLM).
This study employs a combination of various interpretable artificial intelligence methods and the Porta counting fusion method to rank biomedical triples. It utilizes a large language model multi-agent system to generate and optimize candidate hypotheses, and constructs a closed-loop mechanism of interpretation-generation-review-optimization through multi-dimensional scoring and expert agent evaluation.
It has enabled efficient and reliable biomedical knowledge discovery, enhanced feature-level explanatory power and the scientific rationality of hypotheses, and significantly improved the efficiency and reliability of knowledge discovery.
Smart Images
Figure CN121393570B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a biomedical knowledge discovery method, a biomedical knowledge discovery system, an electronic device, and a computer-readable medium. Background Technology
[0002] With the rapid growth of bioinformatics and biomedical data, knowledge discovery is becoming a crucial pathway for advancing scientific research and clinical applications. Traditional knowledge discovery methods, relying on manual organization and statistical approaches, struggle to efficiently analyze complex omics data and literature information. Against this backdrop, the rise of artificial intelligence technologies, particularly Explainable Artificial Intelligence (XAI) and Large Language Models (LLM), has provided new tools for knowledge discovery. XAI aims to improve the transparency of model prediction and reasoning processes, enabling researchers to understand how models arrive at conclusions. In biomedical settings, XAI methods (such as SHAP, LIME, and Integrated Gradients) are widely used for feature importance ranking and visualization, helping to identify key genes, proteins, or pathways. LLM possesses powerful natural language understanding and generation capabilities, capable of extracting information from massive amounts of literature and databases, answering questions, generating reviews, and even proposing initial hypotheses.
[0003] Current autonomous research frameworks (such as DREAM and Proteus) explore automated research processes, but they primarily rely on traditional bioinformatics methods combined with LLM for knowledge discovery, failing to deeply integrate XAI's structured interpretation with LLM's reasoning and generation mechanisms. Furthermore, DREAM and PROTEUS depend on pre-defined bioinformatics pipelines, whose linear architecture and insufficient knowledge integration result in shortcomings in efficiency and versatility for pipeline-based automated knowledge discovery systems. These issues are precisely what XAI addresses through feature-level interpretation and prior knowledge embedding. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a biomedical knowledge discovery method and a corresponding biomedical knowledge discovery system, an electronic device, and a computer-readable medium to overcome or at least partially solve the above problems.
[0005] This invention discloses a biomedical knowledge discovery method, the method comprising: By employing various interpretable artificial intelligence methods and the Porta count fusion method, biomedical triples are sorted to obtain a candidate set of biomedical triples. A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. The biomedical candidate hypotheses are then grouped, evaluated, and optimized to select the biomedical champion hypothesis for each group of biomedical candidate hypotheses. Each biomedical champion hypothesis is quantitatively scored in multiple dimensions to generate an overall score for each biomedical champion hypothesis, and all biomedical champion hypotheses are ranked based on their overall scores.
[0006] Optionally, biomedical triples are ranked using various interpretable artificial intelligence methods and the Borda count fusion method to obtain a candidate set of biomedical triples, including: Biomedical data is taken as input and a hypergraph model is used as the model to be interpreted, wherein nodes in the hypergraph model represent biomedical features and hyperedges represent clusters of biomedical features. The average expression level of all biomedical features in the training dataset was used as the baseline value to normalize the explanation score; On the test dataset, various interpretable artificial intelligence methods were invoked to calculate the attribution score for each node and each hyperedge. The interpretable artificial intelligence methods include DeepLIFT (Deep Learning Importance Propagation) and Integral Gradient (IG) methods. For each explainable AI method, nodes and hyperedges are ranked according to their attribution scores; The ranking results of different interpretable artificial intelligence methods are integrated using the Porta count fusion method to generate a comprehensive ranking score for nodes and hyperedges; Calculate the weighted attribution score for each biomedical triple based on the combined ranking scores of nodes and hyperedges; The weighted attribution scores are normalized, and the biomedical triples are ranked by importance based on the normalized values to obtain a candidate set of biomedical triples.
[0007] Optionally, a biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set, and to evaluate and optimize the biomedical candidate hypotheses to select the biomedical champion hypothesis, including: A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. Each of the biomedical candidate hypotheses is mapped to a high-dimensional semantic space using a pre-trained text embedding model. The cosine similarity between each pair of biomedical candidate hypotheses is calculated, and the hypotheses are paired according to the maximum cosine similarity to group the biomedical candidate hypotheses into multiple hypothesis groups. For each of the biomedical candidate hypotheses, the top-k most relevant literature abstracts are retrieved from the biomedical literature database based on semantic retrieval methods, and a large language model agent is used to evaluate whether the biomedical candidate hypothesis provides a new causal explanation or whether it has been supported or refuted by existing literature. Each of the biomedical candidate hypotheses is independently evaluated using expert agents from multiple fields, including clinical expert agents, omics expert agents, and cell biology expert agents. The evaluation dimensions include originality, feasibility, logical consistency, and scientific rigor. By combining the evaluation results of the large language model agent with the independent evaluation results of the expert agents in the multiple domains, a structured meta-analysis result is generated for each of the biomedical candidate hypotheses. The structured meta-analysis result is a summary of the advantages and disadvantages of the biomedical candidate hypotheses. Within each hypothesis group, based on the structured meta-analysis, a ranking agent is used to compare and eliminate biomedical candidate hypotheses within the group, selecting the winning hypothesis for each group; the comparison dimensions of the ranking agent include novelty, feasibility, testability, and security. The winning hypothesis is rewritten and optimized based on the review feedback in the structured meta-analysis using a large language model agent. The optimized winning hypothesis is used as a new round of candidate hypotheses, and the cycle of grouping, evaluation, competition and optimization is repeated until each group generates a unique biomedical champion hypothesis.
[0008] Optionally, each biomedical champion hypothesis is quantitatively scored across multiple dimensions to generate an overall score for each biomedical champion hypothesis, and all biomedical champion hypotheses are ranked based on their overall scores, including: Each of the aforementioned biomedical champion hypotheses is scored for feasibility, resulting in a feasibility score for each biomedical champion hypothesis. Each of the aforementioned biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis. The overall score for each biomedical champion hypothesis is calculated by combining the feasibility score and the innovativeness score of each hypothesis. All biomedical champion hypotheses are ranked based on their overall scores.
[0009] Optionally, a feasibility score is assigned to each of the biomedical champion hypotheses to obtain a feasibility score for each biomedical champion hypothesis, including: The biomedical champion hypothesis is decomposed into multiple fine-grained independent statements; each independent statement describes the relationship between two biomedical entities. The mcp-server-pubtator3 is called to automatically extract medical subject term descriptions of two biomedical entities and construct a retrieval query without relational terms for each biomedical entity in the statement; Using the aforementioned search query, a search is performed in the biomedical literature database via mcp-server-pubtator3 to identify co-occurring literature that simultaneously contains the two biomedical entities; Evidence supporting the relationship between the two biomedical entities was filtered from the co-occurring documents using mcp-server-pubtator3; The feasibility score of the biomedical champion hypothesis is calculated based on the ratio of the number of evidential documents to the number of co-occurring documents retrieved.
[0010] Optionally, each of the biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis, including: The biomedical literature database constructed for current research will be divided into a historical corpus and a contemporary corpus according to the publication time. Calculate the degree of difference of each biomedical champion hypothesis relative to the historical corpus to obtain the historical degree of difference of each biomedical champion hypothesis; The contemporary dissimilarity of each biomedical champion hypothesis is obtained by calculating the dissimilarity of each biomedical champion hypothesis relative to the contemporary corpus. The average number of citations of the nearest neighbor literature in the contemporary corpus associated with each biomedical champion hypothesis was calculated to characterize the potential academic impact of the research topic. The innovation score for each biomedical champion hypothesis is calculated by combining its historical dissimilarity, contemporary dissimilarity, and average number of citations.
[0011] This invention also discloses a biomedical knowledge discovery system, the system comprising: The biomedical triplet ranking module is used to rank biomedical triples using a variety of interpretable artificial intelligence methods and the Borda count fusion method to obtain a candidate set of biomedical triples. The biomedical hypothesis generation and screening module is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set using a biomedical hypothesis generation system based on a large language model multi-agent system, and to group, evaluate and optimize the biomedical candidate hypotheses, and screen out the biomedical champion hypothesis of each group of biomedical candidate hypotheses. The Biomedical Hypothesis Quantitative Scoring and Ranking Module is used to quantitatively score each Biomedical Champion Hypothesis in multiple dimensions, generate an overall score for each Biomedical Champion Hypothesis, and rank all Biomedical Champion Hypotheses based on the overall scores of each Biomedical Champion Hypothesis.
[0012] The present invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor, when executing a program stored in the memory, implements the biomedical knowledge discovery method as described in this invention.
[0013] The present invention also discloses a computer-readable medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the biomedical knowledge discovery method as described in the present invention.
[0014] This invention has the following advantages: This invention discloses a biomedical knowledge discovery method that employs multiple interpretable artificial intelligence (XAI) methods and a Porta counting fusion approach to rank biomedical triples, obtaining a candidate set of biomedical triples. Then, a biomedical hypothesis generation system based on a large language model (LLM) multi-agent framework generates biomedical candidate hypotheses from this candidate set. These candidate hypotheses are grouped, evaluated, and optimized, and a champion biomedical hypothesis is selected from each group. Each champion hypothesis is quantitatively scored across multiple dimensions, generating an overall score for each. All champion hypotheses are then ranked based on their overall scores. This invention achieves stable interpretation by integrating multiple XAI methods and employing Porta counting, combines an LLM multi-agent architecture for hypothesis generation, evaluation, and optimization, and utilizes RAG technology to evaluate feasibility and novelty through literature review, constructing a closed-loop mechanism of interpretation-generation-review-optimization-evaluation. Compared to existing technologies that rely on pipelines, this invention offers advantages in integrating prior knowledge, efficiency, and reliability. Attached Figure Description
[0015] Figure 1 This is a flowchart of the steps of a biomedical knowledge discovery method provided in an embodiment of the present invention; Figure 2 This is a flowchart of the triple importance ranking provided in an embodiment of the present invention; Figure 3 This is a flowchart of hypothesis generation and tournament provided in an embodiment of the present invention; Figure 4This invention provides a flowchart for quantifying and ranking hypotheses. Figure 5 This is a structural block diagram of a biomedical knowledge discovery system provided in an embodiment of the present invention; Figure 6 This is a block diagram of an electronic device provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of a computer-readable medium provided in an embodiment of the present invention. Detailed Implementation
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0017] Reference Figure 1 The diagram illustrates a flowchart of a biomedical knowledge discovery method provided in an embodiment of the present invention, which may specifically include the following steps: Step 101: The biomedical triples are sorted using a variety of interpretable artificial intelligence methods and the Borda count fusion method to obtain a candidate set of biomedical triples. Step 102: A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. The biomedical candidate hypotheses are then grouped, evaluated, and optimized to select the biomedical champion hypothesis for each group of biomedical candidate hypotheses. Step 103: Quantitatively score each biomedical champion hypothesis in multiple dimensions to generate an overall score for each biomedical champion hypothesis, and rank all biomedical champion hypotheses based on the overall scores of each biomedical champion hypothesis.
[0018] This invention belongs to the field of artificial intelligence, specifically explaining artificial intelligence, large language models, multi-agent systems, and AI4S. Its aim is to provide a knowledge discovery method that combines explainable artificial intelligence (XAI) with a large language model (LLM) multi-agent architecture, overcoming the problems of lengthy and inefficient traditional pipelines. By integrating multiple XAI methods and employing a Borda Score fusion strategy, it establishes an importance ranking of features and triples, achieving feature-level interpretability. While ensuring interpretability, it improves the system's efficiency, reliability, and transferability, providing a new technical approach for knowledge discovery in bioinformatics and biomedicine.
[0019] The knowledge discovery method proposed in this invention combines multiple interpretable artificial intelligence (XAI) methods with a large language model (LLM) multi-agent architecture. Specifically, it includes: integrating multiple XAI methods and employing the BordaScore fusion mechanism to perform feature ranking of gene / protein and functional / structural modules, obtaining a stable candidate set of triples (gene / protein, function / structure, disease); based on this, introducing an LLM-based multi-agent system to generate, evaluate, and optimize the triples, and combining retrieval-enhanced generation (RAG) technology to evaluate feasibility and novelty through literature evidence, forming a complete closed loop of interpretation-generation-review-optimization-evaluation.
[0020] Unlike traditional technologies such as DREAM and PROTEUS, which rely on pre-defined bioinformatics pipelines combined with LLM for task execution, this invention abandons cumbersome serial analysis processes and directly utilizes model interpretation results to achieve feature-level quantitative interpretation and stable ranking. Traditional technologies focus on process automation and result organization, but lack interpretability of feature contributions and verification of hypothesis novelty. This invention, through the deep integration of XAI and LLM, achieves breakthroughs in interpretability, knowledge integration, and hypothesis testing capabilities, significantly distinguishing it from traditional solutions.
[0021] In an optional embodiment of the present invention, biomedical triples are ranked using a variety of interpretable artificial intelligence methods and employing the Porta counting fusion method to obtain a candidate set of biomedical triples, including: Biomedical data is taken as input and a hypergraph model is used as the model to be interpreted, wherein nodes in the hypergraph model represent biomedical features and hyperedges represent clusters of biomedical features. The average expression level of all biomedical features in the training dataset was used as the baseline value to normalize the explanation score; On the test dataset, various interpretable artificial intelligence methods were invoked to calculate the attribution score for each node and each hyperedge. The interpretable artificial intelligence methods include DeepLIFT (Deep Learning Importance Propagation) and Integral Gradient (IG) methods. For each explainable AI method, nodes and hyperedges are ranked according to their attribution scores; The ranking results of different interpretable artificial intelligence methods are integrated using the Porta count fusion method to generate a comprehensive ranking score for nodes and hyperedges; Calculate the weighted attribution score for each biomedical triple based on the combined ranking scores of nodes and hyperedges; The weighted attribution scores are normalized, and the biomedical triples are ranked by importance based on the normalized values to obtain a candidate set of biomedical triples.
[0022] In biomedical applications, model interpretability (post-hoc interpretation) plays a crucial role in clinical credibility and biomarker discovery. This embodiment proposes a post-hoc interpretation workflow that combines multiple interpretable artificial intelligence methods to interpret high-order hypergraph models on proteomic data and achieve importance ranking of the protein-pathway-cancer triplet.
[0023] Reference Figure 2 The process of ranking the importance of triples is as follows: 1. Proteomics data from a mixed-type hepatocellular carcinoma cohort were selected as input, and a hypergraph model was used as the model to be explained.
[0024] 2. The average expression level of all proteins in the training set was used as the baseline value to normalize the explanatory score.
[0025] 3. On the reserved test set, call the DeepLIFT (Deep Learning Important Feasibility) and IG (Integrated Gradients) methods respectively to calculate the attribution score of each node (protein, V) and each hyperedge (protein cluster, E).
[0026] 4. Perform Porta count fusion on the ranking results obtained from each interpretation method. For the i-th entity, its score is calculated as follows:
[0027] in, Indicates the total number of ranked entities. This represents the rank of entity i after sorting by attribution value in descending order. The overall score of an entity is defined as:
[0028] Finally, a unique score S(i) is obtained for each node or superedge, and its impact on the prediction result is quantified.
[0029] 5. For the protein-pathway-cancer triplet, the weighted attribution score is defined as:
[0030] Where S(V) is the Borda score of the protein node, and S(E) is the Borda score of the protein cluster hyperedge. and These represent the total number of proteins and pathways, respectively.
[0031] 6. The weighted scores above are normalized using Min-Max, and the importance of all triples is finally ranked according to the normalized values.
[0032] In an optional embodiment of the present invention, a biomedical hypothesis generation system based on a large language model multi-agent is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set, and to evaluate and optimize the biomedical candidate hypotheses to select the biomedical champion hypothesis, including: A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. Each of the biomedical candidate hypotheses is mapped to a high-dimensional semantic space using a pre-trained text embedding model. The cosine similarity between each pair of biomedical candidate hypotheses is calculated, and the hypotheses are paired according to the maximum cosine similarity to group the biomedical candidate hypotheses into multiple hypothesis groups. For each of the biomedical candidate hypotheses, the top-k most relevant literature abstracts are retrieved from the biomedical literature database based on semantic retrieval methods, and a large language model agent is used to evaluate whether the biomedical candidate hypothesis provides a new causal explanation or whether it has been supported or refuted by existing literature. Each of the biomedical candidate hypotheses is independently evaluated using expert agents from multiple fields, including clinical expert agents, omics expert agents, and cell biology expert agents. The evaluation dimensions include originality, feasibility, logical consistency, and scientific rigor. By combining the evaluation results of the large language model agent with the independent evaluation results of the expert agents in the multiple domains, a structured meta-analysis result is generated for each of the biomedical candidate hypotheses. The structured meta-analysis result is a summary of the advantages and disadvantages of the biomedical candidate hypotheses. Within each hypothesis group, based on the structured meta-analysis, a ranking agent is used to compare and eliminate biomedical candidate hypotheses within the group, selecting the winning hypothesis for each group; the comparison dimensions of the ranking agent include novelty, feasibility, testability, and security. The winning hypothesis is rewritten and optimized based on the review feedback in the structured meta-analysis using a large language model agent. The optimized winning hypothesis is used as a new round of candidate hypotheses, and the cycle of grouping, evaluation, competition and optimization is repeated until each group generates a unique biomedical champion hypothesis.
[0033] To systematically generate, evaluate, and iteratively optimize triples derived from interpretable AI reasoning, this embodiment proposes a tournament-based hypothesis generation framework. This framework employs a multi-stage, multi-agent collaborative competition mechanism to progressively screen and evolve biomedical hypotheses while ensuring diversity, thereby enhancing their scientific rationality, innovativeness, and clinical usability.
[0034] Reference Figure 3 Assume the tournament process is as follows: 1. When the triples generated by XAI inference are input into the tournament hypothesis generation system, N candidate biological hypotheses are automatically generated using a large language model.
[0035] 2. Each candidate hypothesis is mapped to a high-dimensional semantic space using a pre-trained text embedding model. Cosine similarity is calculated between each pair of hypotheses, and they are paired and grouped according to the maximum similarity (if the number of candidates is odd, a single hypothesis group is formed). This grouping strategy ensures that similar hypotheses directly compete, thus promoting diversity in the next round.
[0036] 3. For each hypothesis, extract the top-k most relevant literature abstracts from the PubMed database using a semantic retrieval method (vector similarity retrieval). An LLM agent then evaluates whether the hypothesis provides a new causal explanation, or whether it has been supported or refuted by existing literature.
[0037] 4. Each hypothesis is further independently evaluated by LLM agents comprised of experts from three domains: clinical experts, omics experts, and cell biology experts. This evaluation aims to assess the significance of the hypothesis from an academic perspective, and the evaluation dimensions include: originality, feasibility, logical consistency, and scientific rigor. In the specific evaluation process, the expert LLM agent will summarize each candidate biological hypothesis and systematically review each evaluation dimension in 1-2 paragraphs.
[0038] 5. The above independent review results and causal explanations are combined into a structured meta-analysis using a summarizing agent to summarize the advantages and disadvantages of each hypothesis and output a clear overall evaluation.
[0039] 6. Within each group, competition is assumed to be conducted through a player-kill (PK) process. A dedicated ranking agent, based on meta-analysis results, compares hypotheses from a broader perspective, considering dimensions including: innovativeness, feasibility, testability, and security. Unlike expert agents that focus on academic perspectives, ranking agents emphasize striking a balance between innovative potential and implementability. In the one-on-one competition phase, the ranking agent module sequentially evaluates candidate hypotheses against the four evaluation dimensions mentioned above and selects the winner based on the comprehensive judgment. The winning hypothesis enters the evolution phase, where the LLM agent rewrites and optimizes it based on reviewer feedback, thereby improving its evidentiary consistency and scientific rigor.
[0040] 7. The winning hypothesis from each group advances to the next round, and the cycle of "grouping-evaluation-competition-evolution" is repeated until a unique "champion hypothesis" is finally produced.
[0041] In an optional embodiment of the present invention, each biomedical champion hypothesis is quantitatively scored in multiple dimensions to generate an overall score for each biomedical champion hypothesis, and all biomedical champion hypotheses are ranked based on the overall scores of each biomedical champion hypothesis, including: Each of the aforementioned biomedical champion hypotheses is scored for feasibility, resulting in a feasibility score for each biomedical champion hypothesis. Each of the aforementioned biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis. The overall score for each biomedical champion hypothesis is calculated by combining the feasibility score and the innovativeness score of each hypothesis. All biomedical champion hypotheses are ranked based on their overall scores.
[0042] In an optional embodiment of the present invention, a feasibility score is assigned to each of the biomedical champion hypotheses to obtain a feasibility score for each biomedical champion hypothesis, including: The biomedical champion hypothesis is decomposed into multiple fine-grained independent statements; each independent statement describes the relationship between two biomedical entities. The mcp-server-pubtator3 is called to automatically extract medical subject term descriptions of two biomedical entities and construct a retrieval query without relational terms for each biomedical entity in the statement; Using the aforementioned search query, a search is performed in the biomedical literature database via mcp-server-pubtator3 to identify co-occurring literature that simultaneously contains the two biomedical entities; Evidence supporting the relationship between the two biomedical entities was filtered from the co-occurring documents using mcp-server-pubtator3; The feasibility score of the biomedical champion hypothesis is calculated based on the ratio of the number of evidential documents to the number of co-occurring documents retrieved.
[0043] In an optional embodiment of the present invention, each of the biomedical champion hypotheses is scored for its innovativeness to obtain an innovativeness score for each biomedical champion hypothesis, including: The biomedical literature database constructed for current research will be divided into a historical corpus and a contemporary corpus according to the publication time. Calculate the degree of difference of each biomedical champion hypothesis relative to the historical corpus to obtain the historical degree of difference of each biomedical champion hypothesis; The contemporary dissimilarity of each biomedical champion hypothesis is obtained by calculating the dissimilarity of each biomedical champion hypothesis relative to the contemporary corpus. The average number of citations of the nearest neighbor literature in the contemporary corpus associated with each biomedical champion hypothesis was calculated to characterize the potential academic impact of the research topic. The innovation score for each biomedical champion hypothesis is calculated by combining its historical dissimilarity, contemporary dissimilarity, and average number of citations.
[0044] To systematically evaluate the biological hypotheses generated by the tournament, this embodiment proposes an automated scoring pipeline that quantitatively evaluates hypotheses from two dimensions: literature support and novelty in time stratification. Furthermore, it synthesizes an overall scoring index, thereby providing a scientific basis for the screening and ranking of hypotheses.
[0045] Reference Figure 4 The process of quantitatively scoring and ranking the hypotheses is as follows: 1. Feasibility score (1) The final ranking of the "champion" hypothesis is decomposed into several fine-grained independent statements, each of which describes the relationship between two biological entities. The decomposed agent is represented as entity-entity edges.
[0046] (2) Call mcp-server-pubtator3 (a service based on the PubTator3 API) to automatically extract the MeSH (Medical Subject Headings) descriptions of the two entities and construct a search query that does not contain relational terms in order to expand the search coverage.
[0047] (3) Search PubMed abstracts to identify co-occurring documents containing two entities.
[0048] (4) If a potential supporting abstract is found, the open-access full text is downloaded for confirmation.
[0049] (5) For each document supporting the relationship, record its PubMed ID, title, and the specific sentence containing the evidence. The system outputs a Boolean decision (support / no support).
[0050] (6) The Overall Plausibility (OP) score is defined as follows:
[0051] OP ranges from (0,1) and represents the proportion of support for the hypothetical viewpoint in the literature.
[0052] 2. Scoring based on innovation (1) The PubMed abstract vector database of liver cancer related literature constructed in this study was divided into two time periods: historical corpus (2010–2021) and contemporary corpus (2022–2025).
[0053] (2) Calculation of three novelty indices ①Historical Dissimilarity (HD): Calculates the mean Euclidean distance between the hypothesis vector and its five nearest neighbors in the historical corpus. The higher the value, the greater the discrepancy with historical research.
[0054] ② Contemporary Dissimilarity (CD): Calculates the mean Euclidean distance between the hypothesis vector and its five nearest neighbors in the contemporary corpus. The smaller the value, the closer it is to the forefront of research.
[0055] ③ Calculate the average number of citations of the nearest neighbor documents in the contemporary corpus to characterize the potential academic influence of the research topic.
[0056] (3) Normalize each indicator by the annual mean to obtain annual comparability. The final overall novelty score (ON) is defined as:
[0057] 3. Overall Score (1) Feasibility score The original novelty score ON has a large range. Min-max normalization is performed on the ON scores of all hypotheses to obtain... .
[0058] (2) Define the overall score (OS) as:
[0059] (3) OS takes into account both feasibility and novelty, and suppresses extreme values of a single dimension by means of harmonic averaging.
[0060] (4) All candidate triplet hypotheses are reordered based on OS scores to prioritize subsequent biological validation and clinical translation.
[0061] The technical advantages of this invention are as follows: This invention integrates multiple XAI methods and employs Porta counting to achieve stable interpretation. It combines an LLM multi-agent architecture for hypothesis generation, evaluation, and optimization, and utilizes RAG technology to evaluate feasibility and novelty based on literature evidence, constructing a closed-loop mechanism of interpretation-generation-review-optimization-evaluation. Compared to existing technologies that rely on pipelines, this invention has advantages in integrating prior knowledge, efficiency, and reliability. This invention effectively bridges the gap between interpretable artificial intelligence and practical applications, significantly improving the efficiency and reliability of knowledge discovery by automating hypothesis generation and verification through multi-agent systems.
[0062] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.
[0063] Reference Figure 5 The diagram illustrates a structural block diagram of a biomedical knowledge discovery system provided in an embodiment of the present invention, which may specifically include the following modules: The biomedical triplet sorting module 501 is used to sort biomedical triples using a variety of interpretable artificial intelligence methods and the Borda count fusion method to obtain a candidate set of biomedical triples. The biomedical hypothesis generation and screening module 502 is used to generate biomedical candidate hypotheses based on the biomedical triplet candidate set using a biomedical hypothesis generation system based on a large language model multi-agent system, and to group, evaluate and optimize the biomedical candidate hypotheses, and screen out the biomedical champion hypothesis of each group of biomedical candidate hypotheses. The Biomedical Hypothesis Quantitative Scoring and Ranking Module 503 is used to quantitatively score each biomedical champion hypothesis in multiple dimensions, generate an overall score for each biomedical champion hypothesis, and rank all biomedical champion hypotheses based on the overall score of each biomedical champion hypothesis.
[0064] In an optional embodiment of the present invention, the biomedical triplet sorting module is further configured to: Biomedical data is taken as input and a hypergraph model is used as the model to be interpreted, wherein nodes in the hypergraph model represent biomedical features and hyperedges represent clusters of biomedical features. The average expression level of all biomedical features in the training dataset was used as the baseline value to normalize the explanation score; On the test dataset, various interpretable artificial intelligence methods were invoked to calculate the attribution score for each node and each hyperedge. The interpretable artificial intelligence methods include DeepLIFT (Deep Learning Importance Propagation) and Integral Gradient (IG) methods. For each explainable AI method, nodes and hyperedges are ranked according to their attribution scores; The ranking results of different interpretable artificial intelligence methods are integrated using the Porta count fusion method to generate a comprehensive ranking score for nodes and hyperedges; Calculate the weighted attribution score for each biomedical triple based on the combined ranking scores of nodes and hyperedges; The weighted attribution scores are normalized, and the biomedical triples are ranked by importance based on the normalized values to obtain a candidate set of biomedical triples.
[0065] In an optional embodiment of the present invention, the biomedical hypothesis generation and screening module is further configured to: A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. Each of the biomedical candidate hypotheses is mapped to a high-dimensional semantic space using a pre-trained text embedding model. The cosine similarity between each pair of biomedical candidate hypotheses is calculated, and the hypotheses are paired according to the maximum cosine similarity to group the biomedical candidate hypotheses into multiple hypothesis groups. For each of the biomedical candidate hypotheses, the top-k most relevant literature abstracts are retrieved from the biomedical literature database based on semantic retrieval methods, and a large language model agent is used to evaluate whether the biomedical candidate hypothesis provides a new causal explanation or whether it has been supported or refuted by existing literature. Each of the biomedical candidate hypotheses is independently evaluated using expert agents from multiple fields, including clinical expert agents, omics expert agents, and cell biology expert agents. The evaluation dimensions include originality, feasibility, logical consistency, and scientific rigor. By combining the evaluation results of the large language model agent with the independent evaluation results of the expert agents in the multiple domains, a structured meta-analysis result is generated for each of the biomedical candidate hypotheses. The structured meta-analysis result is a summary of the advantages and disadvantages of the biomedical candidate hypotheses. Within each hypothesis group, based on the structured meta-analysis, a ranking agent is used to compare and eliminate biomedical candidate hypotheses within the group, selecting the winning hypothesis for each group; the comparison dimensions of the ranking agent include novelty, feasibility, testability, and security. The winning hypothesis is rewritten and optimized based on the review feedback in the structured meta-analysis using a large language model agent. The optimized winning hypothesis is used as a new round of candidate hypotheses, and the cycle of grouping, evaluation, competition and optimization is repeated until each group generates a unique biomedical champion hypothesis.
[0066] In an optional embodiment of the present invention, the biomedical hypothesis quantitative scoring and ranking module is further configured to: Each of the aforementioned biomedical champion hypotheses is scored for feasibility, resulting in a feasibility score for each biomedical champion hypothesis. Each of the aforementioned biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis. The overall score for each biomedical champion hypothesis is calculated by combining the feasibility score and the innovativeness score of each hypothesis. All biomedical champion hypotheses are ranked based on their overall scores.
[0067] In an optional embodiment of the present invention, the biomedical hypothesis quantitative scoring and ranking module is further configured to: The biomedical champion hypothesis is decomposed into multiple fine-grained independent statements; each independent statement describes the relationship between two biomedical entities. The mcp-server-pubtator3 is called to automatically extract medical subject term descriptions of two biomedical entities and construct a retrieval query without relational terms for each biomedical entity in the statement; Using the aforementioned search query, a search is performed in the biomedical literature database via mcp-server-pubtator3 to identify co-occurring literature that simultaneously contains the two biomedical entities; Evidence supporting the relationship between the two biomedical entities was filtered from the co-occurring documents using mcp-server-pubtator3; The feasibility score of the biomedical champion hypothesis is calculated based on the ratio of the number of evidential documents to the number of co-occurring documents retrieved.
[0068] In an optional embodiment of the present invention, the biomedical hypothesis quantitative scoring and ranking module is further configured to: The biomedical literature database constructed for current research will be divided into a historical corpus and a contemporary corpus according to the publication time. Calculate the degree of difference of each biomedical champion hypothesis relative to the historical corpus to obtain the historical degree of difference of each biomedical champion hypothesis; The contemporary dissimilarity of each biomedical champion hypothesis is obtained by calculating the dissimilarity of each biomedical champion hypothesis relative to the contemporary corpus. The average number of citations of the nearest neighbor literature in the contemporary corpus associated with each biomedical champion hypothesis was calculated to characterize the potential academic impact of the research topic. The innovation score for each biomedical champion hypothesis is calculated by combining its historical dissimilarity, contemporary dissimilarity, and average number of citations.
[0069] As the system implementation is basically similar to the method implementation, it is described in a relatively simple way. For relevant details, please refer to the description of the method implementation.
[0070] In addition, embodiments of the present invention also provide an electronic device, such as... Figure 6 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. Memory 603 is used to store computer programs; The processor 601, when executing the program stored in the memory 603, implements the biomedical knowledge discovery method as described in the above embodiments.
[0071] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.
[0072] The communication interface is used for communication between the aforementioned terminal and other devices.
[0073] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0074] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0075] like Figure 7 As shown, in another embodiment of the present invention, a computer-readable storage medium 701 is also provided, which stores instructions that, when executed on a computer, cause the computer to perform the biomedical knowledge discovery method described in the above embodiments.
[0076] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the biomedical knowledge discovery method described in the above embodiments.
[0077] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).
[0078] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0079] The various embodiments in this specification are described in a related manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A method for discovering biomedical knowledge, characterized in that, The method includes: By employing various interpretable artificial intelligence methods and the Porta count fusion method, biomedical triples are sorted to obtain a candidate set of biomedical triples. A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. The biomedical candidate hypotheses are then grouped, evaluated, and optimized to select the biomedical champion hypothesis for each group of biomedical candidate hypotheses. Each biomedical champion hypothesis is quantitatively scored in multiple dimensions to generate an overall score for each biomedical champion hypothesis, and all biomedical champion hypotheses are ranked based on the overall scores of each biomedical champion hypothesis. By employing various interpretable artificial intelligence methods and the Porta counting fusion approach, biomedical triples are ranked to obtain a candidate set of biomedical triples, including: Biomedical data is taken as input and a hypergraph model is used as the model to be interpreted, wherein nodes in the hypergraph model represent biomedical features and hyperedges represent clusters of biomedical features. The average expression level of all biomedical features in the training dataset was used as the baseline value to normalize the explanation score; On the test dataset, various interpretable artificial intelligence methods were invoked to calculate the attribution score for each node and each hyperedge. The interpretable artificial intelligence methods include DeepLIFT (Deep Learning Importance Propagation) and Integral Gradient (IG) methods. For each explainable AI method, nodes and hyperedges are ranked according to their attribution scores; The ranking results of different interpretable artificial intelligence methods are integrated using the Porta count fusion method to generate a comprehensive ranking score for nodes and hyperedges; Calculate the weighted attribution score for each biomedical triple based on the combined ranking scores of nodes and hyperedges; The weighted attribution scores are normalized, and the biomedical triples are ranked by importance based on the normalized values to obtain a candidate set of biomedical triples.
2. The method according to claim 1, characterized in that, A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. These candidate hypotheses are then evaluated and optimized to select the champion biomedical hypothesis, including: A biomedical hypothesis generation system based on a large language model and multiple agents is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set. Each of the biomedical candidate hypotheses is mapped to a high-dimensional semantic space using a pre-trained text embedding model. The cosine similarity between each pair of biomedical candidate hypotheses is calculated, and the hypotheses are paired according to the maximum cosine similarity to group the biomedical candidate hypotheses into multiple hypothesis groups. For each of the biomedical candidate hypotheses, the top-k most relevant literature abstracts are retrieved from the biomedical literature database based on semantic retrieval methods, and a large language model agent is used to evaluate whether the biomedical candidate hypothesis provides a new causal explanation or whether it has been supported or refuted by existing literature. Each of the biomedical candidate hypotheses is independently evaluated using expert agents from multiple fields, including clinical expert agents, omics expert agents, and cell biology expert agents. The evaluation dimensions include originality, feasibility, logical consistency, and scientific rigor. By combining the evaluation results of the large language model agent with the independent evaluation results of the expert agents in the multiple domains, a structured meta-analysis result is generated for each of the biomedical candidate hypotheses. The structured meta-analysis result is a summary of the advantages and disadvantages of the biomedical candidate hypotheses. Within each hypothesis group, based on the structured meta-analysis, a ranking agent is used to compare and eliminate biomedical candidate hypotheses within the group, selecting the winning hypothesis for each group; the comparison dimensions of the ranking agent include novelty, feasibility, testability, and security. The winning hypothesis is rewritten and optimized based on the review feedback in the structured meta-analysis using a large language model agent. The optimized winning hypothesis is used as a new round of candidate hypotheses, and the cycle of grouping, evaluation, competition and optimization is repeated until each group generates a unique biomedical champion hypothesis.
3. The method according to claim 1, characterized in that, Each biomedical champion hypothesis is quantitatively scored across multiple dimensions to generate an overall score for each hypothesis. All biomedical champion hypotheses are then ranked based on their overall scores, including: Each of the aforementioned biomedical champion hypotheses is scored for feasibility, resulting in a feasibility score for each biomedical champion hypothesis. Each of the aforementioned biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis. The overall score for each biomedical champion hypothesis is calculated by combining the feasibility score and the innovativeness score of each hypothesis. All biomedical champion hypotheses are ranked based on their overall scores.
4. The method according to claim 3, characterized in that, Each of the aforementioned biomedical champion hypotheses is scored for feasibility, resulting in a feasibility score for each biomedical champion hypothesis, including: The biomedical champion hypothesis is decomposed into multiple fine-grained independent statements; each independent statement describes the relationship between two biomedical entities. The mcp-server-pubtator3 is called to automatically extract medical subject term descriptions of two biomedical entities and construct a retrieval query without relational terms for each biomedical entity in the statement; Using the aforementioned search query, a search is performed in the biomedical literature database via mcp-server-pubtator3 to identify co-occurring literature that simultaneously contains the two biomedical entities; Evidence supporting the relationship between the two biomedical entities was filtered from the co-occurring documents using mcp-server-pubtator3; The feasibility score of the biomedical champion hypothesis is calculated based on the ratio of the number of evidential documents to the number of co-occurring documents retrieved.
5. The method according to claim 3, characterized in that, Each of the aforementioned biomedical champion hypotheses is scored for its innovativeness, resulting in an innovativeness score for each biomedical champion hypothesis, including: The biomedical literature database constructed for current research will be divided into a historical corpus and a contemporary corpus according to the publication time. Calculate the degree of difference of each biomedical champion hypothesis relative to the historical corpus to obtain the historical degree of difference of each biomedical champion hypothesis; The contemporary dissimilarity of each biomedical champion hypothesis is obtained by calculating the dissimilarity of each biomedical champion hypothesis relative to the contemporary corpus. The average number of citations of the nearest neighbor literature in the contemporary corpus associated with each biomedical champion hypothesis was calculated to characterize the potential academic impact of the research topic. The innovation score for each biomedical champion hypothesis is calculated by combining its historical dissimilarity, contemporary dissimilarity, and average number of citations.
6. A biomedical knowledge discovery system, characterized in that, The system includes: The biomedical triplet ranking module is used to rank biomedical triples using a variety of interpretable artificial intelligence methods and the Borda count fusion method to obtain a candidate set of biomedical triples. The biomedical hypothesis generation and screening module is used to generate biomedical candidate hypotheses based on a biomedical triplet candidate set using a biomedical hypothesis generation system based on a large language model multi-agent system, and to group, evaluate and optimize the biomedical candidate hypotheses, and screen out the biomedical champion hypothesis of each group of biomedical candidate hypotheses. The Biomedical Hypothesis Quantitative Scoring and Ranking Module is used to quantitatively score each Biomedical Champion Hypothesis in multiple dimensions, generate an overall score for each Biomedical Champion Hypothesis, and rank all Biomedical Champion Hypotheses based on the overall scores of each Biomedical Champion Hypothesis. The biomedical triplet sorting module is also used for: Biomedical data is taken as input and a hypergraph model is used as the model to be interpreted, wherein nodes in the hypergraph model represent biomedical features and hyperedges represent clusters of biomedical features. The average expression level of all biomedical features in the training dataset was used as the baseline value to normalize the explanation score; On the test dataset, various interpretable artificial intelligence methods were invoked to calculate the attribution score for each node and each hyperedge. The interpretable artificial intelligence methods include DeepLIFT (Deep Learning Importance Propagation) and Integral Gradient (IG) methods. For each explainable AI method, nodes and hyperedges are ranked according to their attribution scores; The ranking results of different interpretable artificial intelligence methods are integrated using the Porta count fusion method to generate a comprehensive ranking score for nodes and hyperedges; Calculate the weighted attribution score for each biomedical triple based on the combined ranking scores of nodes and hyperedges; The weighted attribution scores are normalized, and the biomedical triples are ranked by importance based on the normalized values to obtain a candidate set of biomedical triples.
7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the biomedical knowledge discovery method as described in any one of claims 1-5.
8. A computer-readable medium, characterized in that, It stores instructions that, when executed by one or more processors, cause the processors to perform a biomedical knowledge discovery method as described in any one of claims 1-5.