A method and system for generating biological mechanism hypotheses

Abstract

This application discloses a method and system for generating biological mechanism hypotheses, relating to the interdisciplinary fields of bioinformatics and artificial intelligence. The method includes: acquiring multi-scale biological entity data and constructing a weighted evidence graph; mapping and generating candidate mechanism hypothesis chains based on a skeleton mining algorithm using pre-set templates and global topology optimization; establishing a full-link traceability index including an evidence hierarchy reduction mechanism; initiating a retrieval-enhanced generation process based on the traceability index; utilizing decoding space hard constraint technology to numerically correct the generation probability and combining it with a consistency circuit breaker mechanism to construct a constrained generation space and output a natural language report; and finally updating the graph weights and skeleton template library based on human feedback reinforcement learning. For cross-institutional data, a federated privacy computing architecture is used for secure graph construction. This application addresses the problems of low data utilization and the existence of illusions and lack of traceability in content generated by large language models in biomedical research and development.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of bioinformatics and artificial intelligence, and in particular to a method and system for generating biological mechanism hypotheses. Background Technology

[0002] With the rapid development of high-throughput sequencing, biomedical imaging, and clinical digitization technologies, the biomedical field has accumulated massive amounts of multimodal data, including genomics, transcriptomics, metabolomics, medical imaging, and electronic medical records. In scenarios such as drug development, disease pathogenesis analysis, and personalized diagnosis and treatment, researchers need to extract potential biological mechanisms from this vast amount of data and formulate verifiable scientific hypotheses.

[0003] Although existing technologies attempt to utilize knowledge graphs or deep learning models for assisted analysis, the following technical problems still urgently need to be solved when handling high-precision, high-reliability pharmaceutical R&D tasks: 1. Existing generative models lack "hard fact constraints" and safety fallback mechanisms in the decoding layer, resulting in insufficient credibility.

[0004] Large Language Models (LLMs) based on the Transformer architecture, when generating biomedical reports, still essentially rely on the probability distribution of next-word predictions. While existing Retrieval-Augmented Generation (RAG) techniques use external knowledge bases as context to assist generation, their constraints remain limited to the input stage. During the decoding phase at the model's output, the model may still generate seemingly coherent but biologically inaccurate content due to biases in the training data or probability sampling, ignoring retrieved evidence—a phenomenon known as "hallucination." Currently, there is a lack of a comprehensive control mechanism that can penetrate deep into the model's decoding layer, employing a combination of soft and hard constraints and triggering circuit breakers when evidence is lost.

[0005] 2. Traditional knowledge graph reasoning rules are rigid and difficult to automatically adapt to new knowledge and explore new topologies.

[0006] Traditional biomedical knowledge graph construction often relies on predefined, fixed patterns or simple path-search algorithms. These methods struggle to handle complex, nonlinear biological regulatory networks, such as long chains of "upstream genes - metabolic mediators - downstream phenotypes." Furthermore, existing knowledge graph reasoning frameworks are typically static, unable to automatically evolve new logical topologies based on new experimental data or user feedback. This makes it difficult for the system to discover novel mechanisms beyond the boundaries of human cognition, and it lacks the adaptive evolutionary capability of its logical structure.

[0007] 3. Conflicting evidence exists across multiple scales, and there is a lack of quantifiable and context-aware hierarchical regulation mechanisms.

[0008] Biological data is characterized by its multimodality and multiple sources. Data from different sources (such as CRISPR knockout experiments and pure text mining results) often exhibit contradictions in direction or nature. Existing automatic generation systems often simply aggregate information, lacking algorithms for automatic conflict resolution and confidence correction based on evidence credibility levels and omics context matching. This results in poor internal logical consistency within the generated hypothesis chain, making it difficult to directly guide subsequent wet experimental verification.

[0009] 4. Data privacy barriers and cross-modal feature heterogeneity limit the efficiency of global graph construction.

[0010] High-quality clinical phenotypic data and patient genetic data are scattered across different medical institutions, and due to privacy regulations, the raw data cannot leave their domains. Existing centralized training models are difficult to implement, resulting in a severe "data silo" effect. A key challenge in improving model generalization ability is how to construct a globally weighted directed evidence graph that includes cross-center statistical effect sizes without transmitting raw privacy data and overcoming differences in cross-scale feature dimensions.

[0011] In summary, there is an urgent need for a biological mechanism hypothesis generation method that can effectively integrate multi-scale biological data, possesses decoding layer factual hard and soft constraints and circuit breaker capabilities, supports conflict resolution based on contextual evidence levels and cross-institutional privacy computing, and can achieve adaptive evolution of the logical skeleton as data accumulates. Summary of the Invention

[0012] The purpose of this application is to provide a method and system for generating biological mechanism hypotheses. By constructing a multi-scale evidence map, using soft and hard constraints and circuit breaking technology in the decoding layer, the generation model is forced to perform fact-anchored decoding within the retrieved evidence loop. Through the adaptive evolution of the logical skeleton and the reduction of the evidence hierarchy, a logically rigorous, traceable, and highly confident biological mechanism hypothesis is output.

[0013] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for generating biological mechanism hypotheses, comprising: acquiring multimodal data, clinical phenotypes, and environmental exposure data to be analyzed; defining multi-scale biological entities; calculating statistical association evidence between the multi-scale biological entities; constructing or incrementally updating a weighted directed evidence graph; accessing a logical skeleton template library to obtain a logical skeleton template; the logical skeleton template is selected from at least one of the following: a preset expert logical template, an adaptive template generated by a skeleton self-evolution module from historical high-confidence hypotheses; performing constrained path search on the weighted directed evidence graph according to the topological structure of the logical skeleton template to generate multiple candidate mechanism hypothesis chains, and mapping the multi-scale biological entities in the weighted directed evidence graph to variable nodes of the logical skeleton template; generating a full-link source index for each candidate mechanism hypothesis chain, and based on preset evidence pyramid rules and omics context matching degree, analyzing the... The process involves performing hierarchical reduction and confidence correction on conflicting evidence within the candidate mechanism hypothesis chain. Based on the candidate mechanism hypothesis chain after hierarchical reduction and confidence correction, potential counter-evidence entries are identified from a pre-configured database, and penalty weights are applied to calculate a multi-dimensional confidence score vector for the candidate mechanism hypothesis chain. A pre-configured scoring model is then used to output selected candidate hypotheses based on the multi-dimensional confidence score vector. A structured context is constructed based on the end-to-end source index to initiate the retrieval enhancement generation process. Simultaneously, a hard-constrained decoding space is constructed. Numerical corrections, including soft and hard penalties, are applied to the word generation probability of the large language model, combined with a consistency circuit breaker mechanism, to output a natural language report containing sentence-by-sentence evidence anchors. User interaction feedback on the natural language report is obtained. Based on this feedback and using a reinforcement learning algorithm, the scoring model is updated, triggering an adaptive update of the logical skeleton template library.

[0014] Secondly, this application provides a biological mechanism hypothesis generation system comprising: a graph construction module configured to acquire multi-scale data, perform cross-modal feature dimensionality reduction and covariate correction operations, and calculate statistical correlation evidence in a unified metric space to construct or incrementally update a weighted directed evidence graph; an adaptive skeleton management module configured to store preset logical skeleton templates and mine exploratory topological structures based on topological information entropy to update the templates; a graph mining engine configured to perform constrained path search and biological entity parameter mapping based on the topological structure; and a source tracing, reduction, and scoring engine configured to generate a full-link source tracing index for candidate mechanism hypothesis chains, and match it based on preset evidence pyramid rules and omics context. The system performs hierarchical reduction and confidence correction on conflicting evidence, and calculates a multidimensional confidence score vector by identifying potential counter-evidence items from a pre-set database and applying penalty weights based on the corrected chain. Then, it uses a pre-configured scoring model to output selected candidate hypotheses based on the multidimensional confidence score vector. The retrieval enhancement generation and hard constraint module is configured to initiate retrieval enhancement generation based on the full-link tracing index, and uses a decoding space hard constraint controller to perform numerical correction, including soft and hard penalties, on the generation probability at the decoding layer to output a report. It also includes a consistency circuit breaker to force a rollback when evidence is lost. The reinforcement feedback loop module is configured to train a reward model based on user interaction data and update system parameters.

[0015] Compared with the prior art, this application has the following beneficial effects: 1. Hard constraints and circuit breaker mechanism in the decoding layer achieve fact anchoring: This application constructs a hard constraint decoding space, applies differentiated soft and hard constraints on the lexical generation probability of the large language model, and triggers a consistency circuit breaker mechanism when the model generation deviates from the basis, fundamentally solving the "illusion" problem of generative large language models in rigorous scientific fields, and ensuring that every sentence in the output report is verifiable.

[0016] 2. Adaptive evolution and exploration capabilities of the logical skeleton: Unlike traditional graph reasoning that relies on fixed expert rules, this application performs constrained path search based on the topological structure of the logical skeleton template, and automatically summarizes new skeletons from historical successful cases and abnormal topological information entropy. This gives the system the ability to learn continuously, enabling it to discover complex biological regulation patterns that have not yet been defined by human experts as data accumulates.

[0017] 3. Full-link fusion and reduction of multi-scale data: The cross-modal feature dimensionality reduction, full-link tracing and context-aware evidence pyramid rule mechanism proposed in this application effectively solves the semantic gap and evidence conflict between multi-source heterogeneous data (such as high-throughput sequencing and clinical phenotypes), and ensures that the generated hypotheses are highly consistent in terms of statistical significance and biological rationality. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for generating biological mechanism hypotheses provided in Embodiment 1 of this application; Figure 2 This is a schematic diagram of data flow in Embodiment 1 of this application, which illustrates the construction of a global weighted directed evidence graph based on a federal privacy computing architecture. Figure 3 This is a schematic diagram of the logic skeleton self-evolution and template acquisition process in Embodiment 1 of this application; Figure 4 This is a schematic diagram illustrating the Retrieval Enhancement Generation (RAG) principle based on full-link traceability indexing and Logit masking technology in Embodiment 1 of this application; Figure 5 This is a functional module structure diagram of a biological mechanism hypothesis generation system provided in Embodiment 2 of this application; Figure 6 This is a schematic diagram of the hardware structure of an electronic device for performing the method of Embodiment 1, provided in Embodiment 4 of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 See Figure 1 This embodiment provides a method for generating biological mechanism hypotheses. This method combines the rigor of symbolic logic reasoning with the generative capabilities of Large Language Models (LLMs). The specific process is as follows: It should be noted that the method provided in this application is a computer data processing method, and its output results are only for scientific research and auxiliary analysis reference, and do not include clinical diagnostic and treatment steps that directly affect the human body.

[0023] Step S1: Multi-scale entity feature calculation and evidence map construction.

[0024] In this embodiment, the system first acquires multimodal data sources, macro-level clinical phenotypes, and environmental exposure data. Multimodal data sources include genomics, transcriptomics (such as RNA-Seq), and proteomics data; clinical phenotypes include electronic health records (EHRs). Then, multi-scale biological entities are defined, statistical association evidence between these entities is calculated, and a weighted directed evidence map is constructed or incrementally updated.

[0025] Multi-scale biological entities encompass the following three levels of biological entity nodes: The microscopic layer includes genes, proteins, metabolites, and single nucleotide polymorphism (SNP) sites.

[0026] Mesoscopic layer: includes cell types, microbiome, and histopathological patches, including microbiome such as gut microbiota.

[0027] Macro level: including clinical phenotype, physiological indicators, environmental factors, and drug intervention nodes.

[0028] In particular, such as Figure 2 As shown, for sensitive data involving patient privacy (such as clinical phenotypes), this step can use a federated privacy computing architecture to construct a weighted directed evidence graph, specifically including the following steps: 1. Local Computation: Within each medical institution (local data node), cross-modal feature reduction (e.g., using principal component analysis (PCA) or autoencoders) and covariate correction operations (e.g., regression correction for batch effects, age, gender, and other confounding factors) are first performed on the high-dimensional micro-level data and low-dimensional macro-level data, mapping them to a unified metric space. Subsequently, within this unified metric space, statistical effect sizes and covariance matrices between entities are calculated using local privacy data. For example, the Pearson correlation coefficient between a drug and a specific clinical indicator is calculated.

[0029] 2. Encrypted transmission: The covariance matrix and statistical effect size calculated locally are used as statistical intermediate variables. Differential privacy Laplace noise is added to the statistical intermediate variables, or homomorphic encryption technology is used for encryption.

[0030] 3. Central Aggregation: The encrypted statistical intermediate variables are sent to the central server to aggregate and construct a global weighted directed evidence graph. The weights of the edges in the global weighted directed evidence graph represent the comprehensive statistical strength across centers, such as the p-value after meta-analysis.

[0031] Therefore, by supporting the federated learning architecture, high-quality evidence maps can be constructed using distributed data from multiple centers without compromising the privacy of the original patients, significantly improving the mining efficiency and generalization ability in rare disease or small sample data scenarios.

[0032] Step S2: Obtain the adaptive logic skeleton template.

[0033] The system accesses the logic skeleton template library to obtain logic skeleton templates for guiding reasoning. Depending on the configuration of the actual application scenario, the logic skeleton template is selected from at least one of the following: Pre-defined expert logic templates include classic causal chains such as "drug → target → pathway → disease" or "environmental exposure → epigenetic changes → gene expression → phenotype".

[0034] The adaptive template generated by the skeleton self-evolution module is generated through the following steps: Figure 3 As shown, specifically: the skeleton self-evolution module periodically scans multiple selected hypothesis chains (Top-N) in the historical database that have been "accepted" by users and subsequently verified by wet experiments. Then, using the Maximum Common Subgraph Isomorphism algorithm, it extracts the common topological structure of these hypothesis chains. Simultaneously, the system calculates the topological entropy of the weighted directed evidence graph, identifying anomalously high-weight subgraph structures that are topologically connected but have a historical call frequency below a preset threshold, as potential exploratory topological structures. Finally, the common topological structure and the manually reviewed exploratory topological structures are abstracted into new logical skeleton templates (e.g., discovering a new "metabolite-mediated immune regulation" pattern) and stored in the logical skeleton template library for subsequent use.

[0035] Therefore, unlike traditional graph reasoning that relies on fixed expert rules, this embodiment uses a subgraph isomorphism algorithm to automatically summarize new skeletons from historical success cases. This gives the system the ability to learn continuously, enabling it to discover complex biological regulatory patterns that have not yet been defined by human experts as data accumulates.

[0036] Step S3: Skeleton mining and parameter mapping.

[0037] On the constructed weighted directed evidence graph, the system performs constrained path search based on the topological structure of the selected logical skeleton template (e.g., A→B→C), generating multiple candidate mechanism hypothesis chains. To ensure the robustness and accuracy of the search results, this step employs at least one of the following algorithmic strategies: A subgraph extraction algorithm based on global topology optimization aims to find the subgraph structure that connects key entities (i.e., core nodes in the graph with degree centrality or betweenness centrality higher than a preset threshold) and has the largest total weight. In a specific implementation, a prize-collecting Steiner Tree algorithm can be used as a concrete implementation of this strategy.

[0038] The correlation propagation algorithm based on node affinity aims to calculate the strength of topological associations between entities. In practice, a random walk with restart algorithm can be used as a specific implementation of this strategy.

[0039] During the constrained path search process, the algorithm maps specific biological entities in the graph to variable nodes of the template and performs directional logic checks to forcibly eliminate paths that are connected but whose directions contradict the causal logic of the template.

[0040] Step S4: End-to-end tracing and evidence hierarchy specification.

[0041] For each generated candidate mechanism hypothesis chain, a full-link traceability index is established. When conflicting evidence appears in the chain, evidence hierarchy reduction and confidence correction are performed based on the preset evidence pyramid rules.

[0042] In this embodiment, the preset evidence pyramid rule can specifically be: the system pre-stores a metadata weight parameter table to define the numerical priority of different data sources. In a specific preferred configuration, this parameter table sets the following weight numerical relationship based on the evidence pyramid rule: Level 1 (high weighting): CRISPR / RNAi knockout experiments, randomized controlled trials (RCTs).

[0043] Level 2 (medium weighted value): Statistical association in cohort studies and case-control studies.

[0044] Level 3 (low weight value): high-throughput screening of cell lines, animal models.

[0045] Level 4 (Basic Weight Value): Purely calculated prediction, text mining evidence.

[0046] For the candidate mechanism hypothesis chain, when the processor identifies competing evidence edges with directional or attribute conflicts between the same pair of nodes: first, it reads the source metadata tag corresponding to the competing evidence edge and calculates the omics context matching degree between the experimental environment or tissue type represented by the source metadata tag and the current clinical phenotype data to be analyzed; then, based on the omics context matching degree, it performs dynamic scaling on the basic numerical priority weights to obtain corrected weight values; finally, it performs a numerical reduction operation, retains the direction indicated by the evidence edge with the higher corrected weight value, and deducts the confidence score of the retained evidence edge based on the strength value of the conflicting evidence edge with the lower weight value through a preset decay function to achieve confidence correction.

[0047] Therefore, the end-to-end tracing and evidence pyramid reduction mechanism proposed in this embodiment effectively solves the semantic gap and evidence conflict between multi-source heterogeneous data (such as high-throughput sequencing and clinical phenotypes), and ensures that the generated hypotheses are highly consistent in terms of statistical significance and biological rationality.

[0048] In one specific implementation, each evidence edge e_i is represented as: e_i=<v_s,v_t,p_i,d_i,source_i,level_i,context_i,effect_i,pvalue_i,confidence_i> .

[0049] Where v_s and v_t represent the source node and target node, respectively, p_i represents the predicate type, d_i represents the direction, source_i represents the source of evidence, level_i represents the level of evidence, context_i represents the organization, disease, omics type or experimental condition label, effect_i represents the effect size, pvalue_i represents the statistical significance, and confidence_i represents the initial confidence level.

[0050] For the i-th evidence edge, the system calculates the correction weight W'_i according to the following formula: W'_i=W_level(i)×M_context(i)×Q_source(i).

[0051] Where W_level(i) is the basic weight of the evidence level, M_context(i) is the context matching degree, and Q_source(i) is the data source quality coefficient.

[0052] When multiple evidence edges with opposite directions or opposite predicate attributes exist between the same pair of nodes, the system identifies these multiple evidence edges as a set of conflicting evidence edges, E_conflict. For each evidence edge e_i in E_conflict, the system calculates its correction weight W'_i and selects the evidence edge with the largest correction weight as the primary evidence edge. e*=argmax_{e_i∈E_conflict}W'_i.

[0053] Where e* represents the main evidence edge that is retained. argmax_ is the evidence edge with the largest correction weight selected as the main evidence edge.

[0054] After determining the primary evidence edge e*, the system treats all evidence edges in E_conflict that conflict with e* in terms of direction or predicate attribute, excluding e*, as conflict-deducting edges, and performs attenuation correction on the confidence of the primary evidence edge according to the following formula: C'=C_main×exp(-λ×Σ_{e_j∈E_conflict,e_j≠e*}W'_j).

[0055] Where C_main is the initial confidence level of the main evidence edge, C' is the confidence level of the main evidence edge after attenuation correction, W'_j is the correction weight of the j-th conflict deduction edge, λ is the attenuation coefficient, and e_j is the j-th evidence edge.

[0056] Optionally, a conflict-deducted edge is included in the confidence decay calculation only if its correction weight is greater than a preset threshold τ. The specification results include the retained main evidence edges, evidence edges that have been downweighted or marked as conflicting, the confidence values ​​before and after decay, and the corresponding evidence sources, and are written into the full-link traceability index.

[0057] Step S5: Multidimensional scoring and rebuttal penalty.

[0058] The multidimensional confidence score vector for the candidate mechanism hypothesis chain is calculated, including: evidence strength score, which is an aggregate score based on statistical p-value and effect size; logical fit, which characterizes the degree of topological matching between the candidate mechanism hypothesis chain and the logical skeleton template; cross-modal consistency, which characterizes the directional consistency of evidence at different omics levels; and rebuttal risk score, which is a negative score based on the penalty weights applied to the identified conflicting evidence and potential rebuttal items.

[0059] At the same time, if a clear counter-evidence is found, such as a toxicology database showing that a certain drug is highly toxic to the target organ, a high penalty weight is applied, and the final output is a screened candidate hypothesis. The screened candidate hypothesis can be understood as a pre-defined top-N candidate hypothesis with the highest scores.

[0060] In one specific implementation, the overall score Score(H) of the candidate mechanism hypothesis chain H is calculated according to the following formula: Score(H)=αS_e+βS_l+γS_m-δS_r.

[0061] Wherein, S_e represents the strength of evidence score, which is obtained by aggregating the corrected weights, effect sizes, and statistical significance of each evidence edge in the chain; S_l represents the logical fit score, which is calculated based on the degree of matching between the node type, edge direction, and predicate type in the chain and the logical skeleton template; S_m represents the cross-modal consistency score, which is calculated based on the consistency of evidence direction and effect trend at different omics levels; S_r represents the risk of rebuttal score, which is calculated based on the retrieved conflicting evidence, toxicology records, side effect records, or contraindication records; α, β, γ, and δ are preset weights or weight parameters trained from historical feedback data.

[0062] Step S6: Report generation based on retrieval enhancement generation (RAG) and hard constraint decoding.

[0063] The system employs the Retrieval Enhancement Generation (RAG) principle based on full-link source indexing and Logit masking technology to prevent generation illusions, such as... Figure 4 As shown, the specific process is as follows: Constructing a dynamic decoding constraint set: Based on the full-link source index of the top-ranked hypothesis, the system extracts all relevant entity names, valid logical connectors, and a pre-defined high-frequency vocabulary of general grammar (such as commonly used prepositions, pronouns, and auxiliary words) to construct a restricted candidate set. In a preferred embodiment, this set is constructed as a dynamic trie.

[0064] Perform numerical correction operation: At the time step when the large language model generates the next token, the system calculates the legal subset of the next hop token by combining the current context.

[0065] Numerical Correction and Soft / Hard Constraint Mechanisms: At the time step when the large language model generates the next token, a differentiated numerical correction operation is performed on the probability distribution of the output layer. For key biological entities and causal predicate tokens, the generation probability of all tokens outside the subset of legal candidate tokens is set to zero or a very low value close to zero (i.e., Logit mask hard constraint). For non-key decorative tokens, a preset decay penalty coefficient (soft constraint) is applied to preserve the coherence of natural language, thereby forcing the model to decode and generate the natural language report within the factual boundary defined by the source index. Furthermore, a real-time monitoring and consistency circuit breaker mechanism is introduced: The generated text stream of the natural language report is monitored in real time. When a preset number of consecutive tokens cannot find corresponding evidence in the full-link source index, it is determined that the model has experienced a phantom slippage. At this time, the consistency circuit breaker mechanism is triggered, forcibly terminating the current generation process and re-triggering the step of obtaining the logical skeleton template. Thus, by combining hardware and software decoding layer constraints and circuit breaker fallback mechanisms, the "illusion" problem of generative AI in rigorous scientific fields is fundamentally solved, ensuring that the output reports are both natural and fluent, and that every sentence is verifiable.

[0066] In one specific implementation, let V be the complete vocabulary of the large language model, A_t be the subset of legal candidate lexical units obtained at the t-th decoding time step based on the dynamic decoding constraint set, and K be the set of key biological entity lexical units and causal predicate lexical units. The lexical set K is extracted from multi-scale biological entity data by mapping the variable nodes in the logical skeleton template described in step S2 through a full-link source tracing index.

[0067] For any lexical v∈V, the system corrects its logit value according to the following formula: logit'_t(v) = logit_t(v), when ; logit'_t(v) = -M, when and ; logit'_t(v)=logit_t(v)- ,when and ; Where M is the hard blocking penalty constant, This is the soft penalty coefficient. The corrected logit'_t(v) is normalized and used for the next token sampling or greedy decoding.

[0068] Furthermore, the system performs anchor consistency checks on the generated text. For each generated sentence, the sentence-level anchor rate (AnchorRate) is calculated: AnchorRate = N_anchor / N_key.

[0069] Where N_anchor represents the number of key entity words and causal predicate words in the current sentence that can be mapped to the full-link tracing index, and N_key represents the total number of key entity words and causal predicate words in the current sentence. When AnchorRate is lower than a preset threshold θ (the value of which is preset by the experimenter according to the need for report rigor), or when k consecutive generated words cannot be mapped to the full-link tracing index, dynamic decoding constraint set, or general grammar word set, a consistency circuit breaker mechanism is triggered, terminating the current generation process and reconstructing the decoding constraint set based on the target mechanism hypothesis chain before regenerating.

[0070] Step S7: Feedback-based self-evolutionary closed loop.

[0071] As a further optimization step, the system obtains user (typically researchers) ratings or revision feedback on the report. For example, direct user ratings can be used as positive reward signals, or reward values ​​can be quantified by calculating the overlap between the revised text and the core entities or causal directions in the original report. This interactive data is used to train the reward model, and the rating parameters are fine-tuned using reinforcement learning algorithms such as Proximal Policy Optimization (PPO). Simultaneously, successfully validated hypotheses are sent back to step S2, triggering an update to the logical skeleton template library.

[0072] Example 2 See Figure 5 This embodiment provides a biological mechanism hypothesis generation system. The system includes multiple functional modules, each of which can be stored as a computer program entity in a memory and read and executed by a processor. Specifically, the system includes: The graph construction module is configured to be responsible for the extraction, transformation, and loading (ETL) of multi-source data, cross-modal feature dimensionality reduction and covariate correction operations, statistical computation in a unified space, and updating of the evidence graph; preferably, it supports federated statistical computation in a multi-center environment.

[0073] Adaptive skeleton management module: Configured to store pre-set logical skeleton templates and perform logical skeleton template self-evolution mining based on historical data and exploratory topological structure mining based on topological information entropy, specifically using the maximum common subgraph isomorphism algorithm.

[0074] Graph mining engine: configured to perform constrained path search and biological entity parameter mapping based on topology.

[0075] The source tracing, reduction, and scoring engine is configured to generate a full-link source tracing index for the candidate mechanism hypothesis chain. Based on the preset evidence pyramid rules and omics context matching degree, it performs hierarchical reduction and confidence correction on conflicting evidence. Based on the corrected chain, it calculates a multi-dimensional confidence score vector by identifying potential counter-evidence items and applying penalty weights, and then outputs the screened candidate hypotheses.

[0076] The retrieval enhancement generation and hard constraint module integrates a decoding space hard constraint controller (Logit mask controller), configured to perform numerical correction on the generation probability at the decoding layer, including hard blocking of keyword elements and soft penalty of modifier elements, to block the generation of non-source content; the retrieval enhancement generation and hard constraint module also includes a consistency circuit breaker, configured to monitor the generated text stream of the natural language report in real time; when a preset number of tokens cannot be found in the full-link source index, the current generation process is forcibly terminated, and the step of obtaining the logical skeleton template is retried.

[0077] Reinforcement feedback closed-loop module: Configured to collect user feedback, perform reinforcement learning from human feedback (RLHF) to train the reward model and update system parameters.

[0078] Example 3 This embodiment uses the example of "finding potential existing drugs for treating idiopathic pulmonary fibrosis (IPF)" to illustrate the specific implementation process of the method described in this application in a drug repositioning scenario. In this embodiment, the disease phenotype is IPF, the candidate drug library is an approved drug library, and the node types in the weighted directed evidence graph include drug nodes, target nodes, transcription factor nodes, pathway nodes, and phenotype nodes. The logical skeleton template is configured as follows: .

[0079] in," "→" indicates a suppressive relationship, while "→" indicates a promoting, related, or causative relationship.

[0080] S1: Obtain the evidence graph and logical skeleton template, and search for candidate mechanism hypothesis chains.

[0081] The system acquires a pre-constructed or incrementally updated weighted directed evidence graph. The weighted directed evidence graph includes multi-scale biological entity nodes, evidence edges between entities, evidence source metadata, edge direction information, and edge weight information.

[0082] In this embodiment, the system calls a weighted directed evidence graph that includes approved drugs, drug targets, transcription factors, pathways, IPF-related phenotypes, public transcriptome data, drug action databases, and literature evidence. Each evidence edge records the following information: <Source node, target node, predicate type, direction, source of evidence, level of evidence, effect size, statistical significance, context label, initial confidence level>.

[0083] For example, an evidence graph can include edges of the following types: Drug → Target.

[0084] drug Target point.

[0085] Target → Transcription factor.

[0086] Transcription factors → fibrotic phenotype.

[0087] Pathway → IPF phenotype.

[0088] The system further obtains the logical skeleton template for the drug relocation scenario: .

[0089] Subsequently, the system performs a constraint path search in the weighted directed evidence graph, generates a chain of candidate mechanism hypotheses that satisfies the logical skeleton template, and eliminates paths whose directions are inconsistent with the template.

[0090] For example, if the following path exists in the graph: Vertiporfen YAP1 → fibrotic phenotype.

[0091] If the node type, edge direction, and predicate type in the path all match the logical skeleton template, then the path is retained as a candidate mechanism hypothesis chain.

[0092] Conversely, if a path connects the drug, target, and phenotypic nodes, but the direction is: Fibrosis phenotype → YAP1 → Drug.

[0093] The path is then eliminated because it does not meet the template direction constraint.

[0094] S2: Generate a full-link traceability index for the candidate mechanism hypothesis chain.

[0095] For each candidate mechanism hypothesis chain obtained from the search, the system generates a full-link source index. The full-link source index is used to record each node, each edge, and its source of evidence in the chain.

[0096] In this embodiment, regarding the candidate mechanism hypothesis chain: Vertiporfen YAP1 → fibrotic phenotype.

[0097] The system-generated end-to-end traceability index includes at least: Hypothesis chain numbering; Drug entity: virtepofen; Target or transcription factor entity: YAP1; Disease phenotypic entities: IPF or fibrotic phenotype; Predicate type of drug-target edge: inhibition; Predicate type of target-phenotype edge: facilitating or positively correlated; The source of evidence for each piece of evidence; The evidence level of each piece of evidence; The directional relationship of each piece of evidence; The effect size of each evidence edge; Statistical significance of each evidence edge; Context label for each evidence edge; The initial confidence level of each evidence edge; Each piece of evidence corresponds to an evidence number or evidence anchor point.

[0098] The end-to-end traceability index is used not only for subsequent evidence reduction and scoring, but also for building a dynamic decoding constraint set during the report generation stage.

[0099] S3: Perform evidence specification and confidence correction based on evidence level and context matching degree.

[0100] For evidence edges in the candidate mechanism hypothesis chain, the system identifies whether there are competing evidence edges with conflicting directions or conflicting predicate attributes.

[0101] In this embodiment, regarding the relationship between "YAP1 and fibrosis phenotype", the following conflicting evidence may exist: Evidence edge e1: YAP1 → fibrotic phenotype.

[0102] Evidence edge e2: YAP1 Fibrosis phenotype.

[0103] e1 may come from transcriptome data of lung tissue from IPF patients or CRISPR knockout experiments, while e2 may come from other tissue backgrounds or early literature reports.

[0104] The system assigns basic weights to different sources of evidence to determine their levels of evidence. For example: Level 1: CRISPR / RNAi intervention experiments, randomized controlled trials; Level 2: Cohort studies, case-control studies; Level 3: Cell line screening, animal models; Level 4: Computational prediction and text mining evidence.

[0105] For the i-th evidence edge, the system calculates its correction weight: W'_i=W_level(i)×M_context(i)×Q_source(i).

[0106] in: W_level(i): The basic weight of the evidence level for the i-th evidence edge; M_context(i): The context matching degree between the i-th evidence edge and the current IPF scenario; Q_source(i): The data source quality coefficient of the i-th evidence edge.

[0107] When there is a conflict among multiple evidence edges, the system selects the evidence edge with the largest correction weight as the main evidence edge: e*=argmax W'_i.

[0108] And perform attenuation correction on the confidence level of the main evidence edge: C'=C_main×exp(-λ×ΣW'_conflict).

[0109] in: C': Corrected confidence level of the principal evidence side; C_main: Initial confidence level of the main evidence edge; λ: Attenuation coefficient, used to adjust the intensity of the negative impact of conflicting evidence on the final confidence level. Its value range is preset by the experimenters based on historical verification data. ΣW'_conflict: The sum of correction weights for other evidence edges that conflict with the main evidence edge.

[0110] For example, if CRISPR evidence in the context of IPF lung tissue shows that YAP1 promotes fibrosis, while earlier, lower-ranking literature suggests that YAP1 has a protective effect in other tissues, then the system should be preferentially preserved: YAP1 → fibrotic phenotype.

[0111] At the same time, conflict evidence is written into the full-link traceability index as evidence of de-weighting, instead of being simply deleted.

[0112] S4: Calculate multidimensional confidence scores and screen target mechanism hypothesis chains.

[0113] After completing the evidence specification, the system calculates a multidimensional confidence score for each candidate mechanism hypothesis chain.

[0114] In this embodiment, the comprehensive score of candidate mechanism hypothesis chain H is calculated according to the following formula: Score(H)=αS_e+βS_l+γS_m-δS_r.

[0115] in: S_e: Strength of evidence score; S_l: Logical fit; S_m: Cross-modal consistency score; S_r: Risk score for proof by contradiction; α, β, γ, δ: Preset weights or weight parameters obtained through historical feedback training.

[0116] Specifically: S_e is obtained by aggregating the corrected weights, effect sizes, and statistical significance of drug-target edges, target-phenotype edges, or transcription factor-phenotype edges. S_l, based on the candidate mechanism hypothesis chain and Drug The Target / Transcription Factor → Phenotype logical skeleton template is calculated based on the degree of matching in node type, edge direction, and predicate type. S_m is calculated based on the consistency of evidence in the direction of drug action, transcriptome, pathway and phenotypic evidence. S_r is calculated based on counter-evidence entries in drug side effects databases, toxicology databases, contraindication records, or contrary experimental results that conflict with the IPF treatment goals.

[0117] If a candidate mechanism chain simultaneously possesses strong drug-target evidence, high IPF lung tissue context matching, good cross-modal directional consistency, and no high-level counter-evidence is found, then the chain's score (H) is high.

[0118] The system selects the target mechanism hypothesis chain from multiple candidate mechanism hypothesis chains based on a comprehensive score. For example: Vertiporfen YAP1 → fibrotic phenotype.

[0119] It can be selected as a target mechanism hypothesis chain.

[0120] S5: Construct a dynamic decoding constraint set based on the target mechanism hypothesis chain.

[0121] The system constructs a dynamic decoding constraint set based on the full-link traceability index of the target mechanism hypothesis chain. Among them, the key biological entity term set K is extracted from the graph data by mapping the variable nodes in the drug relocation logic skeleton template described in step S1 (corresponding to the overall process S2) through the full-link traceability index.

[0122] In this embodiment, the system extracts the allowed content from the full-link traceability index, including: Drug entity words: virteporfin; Target or transcription factor term: YAP1; Disease phenotype terms: IPF, idiopathic pulmonary fibrosis, fibrotic phenotype; Causal predicate words: inhibit, promote, upregulate, downregulate, correlate, possibly through; Evidence anchor terms: evidence number, database number, document number; Common grammatical terms: due to, therefore, possibly, prompt, with, related to, through, thus, etc.

[0123] The system constructs the above lexical set into a Trie or a finite state automaton, which is used to provide a subset of legal candidate lexical units at each decoding time step in the large language model's report generation process.

[0124] For example, when the generation context is: Vertepofen may pass.

[0125] The system can restrict the next legal candidate terms to: Suppress, downregulate, regulate, and match entities and predicates within the full-link tracing index, without allowing the model to freely generate mechanistic relationships that do not appear in the index.

[0126] S6: Perform Logit value correction and generate a report during the decoding phase.

[0127] When calling a large language model to generate a natural language report, the system performs numerical correction on the word probabilities or logit values ​​of the model output layer at each decoding time step based on the dynamic decoding constraint set.

[0128] set up: V: The complete vocabulary of a large language model; A_t: The subset of legal candidate lexical units at the t-th decoding time step; K: Set of key biological entity lexical units and causal predicate lexical units; logit_t(v): The model's original logit value for word v; logit'_t(v): The corrected logit value.

[0129] The system will make corrections according to the following rules: logit'_t(v)=logit_t(v), when v∈A_t; logit'_t(v) = -M, when And v∈K; logit'_t(v)=logit_t(v)- ,when and .

[0130] Where M is the hard blocking penalty constant, This is a soft penalty coefficient.

[0131] In other words, for key biological entities, causal predicates, or mechanism relationships that do not appear in the full-link traceability index, the system performs hard blocking; for non-key modifying terms, it performs soft penalties to balance factual constraints and language fluency.

[0132] For example, when the model attempts to generate: Verteporfen can directly degrade collagen.

[0133] When the evidence edge "virtoporfen → collagen degradation" does not exist in the full-link source tracing index, key predicates and entity terms related to "collagen degradation" will be hard blocked or significantly penalized. The system will instead generate terms within the range of legitimate candidate terms. Vertepofen may reduce the risk of profibrotic phenotypes by inhibiting YAP1-related signaling.

[0134] And add corresponding evidence anchors at the end of the sentence.

[0135] S7: Perform anchor consistency monitoring, circuit breaker rollback, and output a traceable report.

[0136] During the generation process, the system monitors the anchoring consistency between the generated text and the full-link traceability index in real time.

[0137] For each generated sentence, the system calculates the sentence-level anchoring rate: AnchorRate = N_anchor / N_key.

[0138] in: N_anchor: The number of key entity terms and causal predicate terms in the current sentence that can be mapped to the full-link source index; N_key: The total number of key entity words and causal predicate words in the current sentence.

[0139] The system triggers a consistency circuit breaker mechanism (i.e., a generation process termination and rollback logic based on real-time consistency monitoring) when any of the following conditions occur: The AnchorRate of the current sentence is lower than the preset threshold θ (the value of which is preset by the experimenters according to the need for the rigor of the report). k consecutive generated lexical units cannot be mapped to the full-link source index, dynamic decoding constraint set, or general syntax lexical set; The generated text contains biological entities, causal predicates, or evidence numbers that are not included in the full-link tracing index; The causal direction in the generated text conflicts with the edge direction in the chain of the target mechanism hypothesis.

[0140] After the circuit breaker is triggered, the system terminates the current generation process and reconstructs the dynamic decoding constraint set based on the target mechanism hypothesis chain or re-executes the restricted decoding.

[0141] Finally, the system outputs a report on the drug relocation mechanism hypothesis, containing sentence-by-sentence evidence anchors. For example: The candidate drug virtepofen may affect the IPF-related fibrosis phenotype by inhibiting YAP1-related regulatory activity [Anchor of Evidence 1]. Within the target mechanism chain, there is high-level evidence supporting a consistent direction between YAP1 and the pro-fibrotic phenotype [Anchor of Evidence 2]. Untraceable mechanisms such as "direct collagen degradation" were not included in the report because this relationship is not present in the end-to-end traceability index.

[0142] Thus, this embodiment achieves a complete closed loop from the generation of candidate mechanism chains for drug repositioning, evidence specification, mechanism chain scoring, dynamic decoding constraints, restricted report generation to consistency circuit breaker rollback.

[0143] Example 4 This embodiment provides an electronic device, the internal structure of which can be shown in the following diagram. Figure 6 As shown. The computer device includes a processor, memory, system bus, and communication interface. The memory stores a computer program, which, when executed by the processor, implements the steps described in Embodiment 1. This electronic device can be a high-performance server, workstation, or a cloud-based virtual computing node.

[0144] Example 5 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0145] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Furthermore, any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory.

[0147] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating biological mechanism hypotheses, characterized in that, include: Acquire multimodal data, clinical phenotypes, and environmental exposure data to be analyzed; define multi-scale biological entities; calculate statistical association evidence between the multi-scale biological entities; and construct or incrementally update a weighted directed evidence map. Access the logic skeleton template library to obtain a logic skeleton template; the logic skeleton template is selected from at least one of the following: a preset expert logic template, an adaptive template generated by the skeleton self-evolution module from historical high-confidence hypotheses; On the weighted directed evidence graph, a constraint path search is performed based on the topological structure of the logical skeleton template to generate multiple candidate mechanism hypothesis chains, and the multi-scale biological entities in the weighted directed evidence graph are mapped to the variable nodes of the logical skeleton template. A full-link source index is generated for each candidate mechanism hypothesis chain, and hierarchical reduction and confidence correction are performed on conflicting evidence in the candidate mechanism hypothesis chain based on the preset evidence pyramid rules and omics context matching degree. Based on the candidate mechanism hypothesis chain after the hierarchical reduction and confidence correction, potential counter-evidence entries are identified from a pre-set database and penalty weights are applied to calculate the multidimensional confidence score vector of the candidate mechanism hypothesis chain. Then, a pre-configured scoring model is used to output the selected candidate hypotheses based on the multidimensional confidence score vector. Based on the full-link source tracing index, a structured context is constructed to initiate the retrieval enhancement generation process, and a hard-constrained decoding space is constructed simultaneously. By applying a numerical correction including soft and hard penalties to the word generation probability of the large language model and combining it with a consistency circuit breaker mechanism, a natural language report containing sentence-by-sentence evidence anchors is output. Obtain user interaction feedback on the natural language report, update the scoring model based on the interaction feedback and using a reinforcement learning algorithm, and trigger an adaptive update of the logical skeleton template library.

2. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The multi-scale biological entity comprises a microscopic layer, a mesoscopic layer, and a macroscopic layer; wherein, the nodes of the microscopic layer include genes, proteins, metabolites, and single nucleotide polymorphism sites; the nodes of the mesoscopic layer include cell subpopulations, microbial communities, and histopathological feature blocks; and the nodes of the macroscopic layer include clinical phenotypes, physiological indicators, environmental exposure factors, and drug intervention nodes.

3. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The construction or incremental update of the weighted directed evidence graph specifically adopts a federated privacy computing architecture, including: At multiple local data nodes, cross-modal feature dimensionality reduction and covariate correction operations are performed on high-dimensional micro-level data and low-dimensional macro-level data to map them to a unified metric space; within the unified metric space, the covariance matrix and statistical effect size among the multi-scale biological entities are independently calculated based on their respective privacy data. The covariance matrix and statistical effect size are transmitted to the central server as statistical intermediate variables. Before transmission, the statistical intermediate variables are subjected to security processing by adding noise using differential privacy technology or by at least one of homomorphic encryption technology. In the central server, statistical intermediate variables from each of the local data nodes are aggregated to construct a global weighted directed evidence graph; wherein, the weight of the edge in the global weighted directed evidence graph represents the comprehensive statistical strength across data nodes.

4. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The adaptive template, generated by the skeleton self-evolution module from historical high-confidence fallacies, specifically includes: Periodically extract filtered hypothesis chains that have been accepted by users and verified by experiments from historical interaction feedback; The common topological structure of multiple candidate hypothesis chains is calculated using the maximum common subgraph isomorphism algorithm. Calculate the topological information entropy of the weighted directed evidence graph, and mine abnormally high weighted subgraph structures that are topologically connected but have a historical call frequency lower than a preset threshold as potential exploratory topological structures; The shared topology and the exploratory topology after manual review are abstracted into a new logical skeleton template and stored in the logical skeleton template library.

5. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The algorithm for performing constraint path search on the weighted directed evidence graph based on the topological structure of the logical skeleton template includes at least one of the following: A subgraph extraction algorithm based on global topology optimization is used to find subgraph structures in the weighted directed evidence graph that connect entities with topological centrality higher than a preset threshold and maximize aggregate weight. A correlation propagation algorithm based on node affinity is used to calculate the correlation strength of the multi-scale biological entities in the topology of the weighted directed evidence graph. The constraint path search further includes: forcibly eliminating paths with mismatched directions based on the causal definition of the logical skeleton template.

6. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The method based on the preset evidence pyramid rule performs hierarchical reduction and confidence correction on conflicting evidence in the candidate mechanism hypothesis chain, specifically including: Configure a metadata weight parameter table associated with the data source type, where different experiment types or data sources are assigned different basic numerical priority weights; For the candidate mechanism hypothesis chain, identify competing evidence edges where there is a conflict in direction or attribute between the same pair of nodes; Read the source metadata tags corresponding to the competing evidence edge, and calculate the omics context matching degree between the experimental environment or organization type represented by the source metadata tags and the current clinical phenotype data to be analyzed; Based on the omics context matching degree, the basic numerical priority weights are dynamically scaled to obtain the corrected weight values. Perform a numerical reduction operation, retain the direction indicated by the evidence edge with higher weight value, and deduct the confidence score of the retained evidence edge based on the strength value of the conflict evidence edge with lower weight value through a preset decay function.

7. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, Based on the full-link source tracing index, a structured context is constructed to initiate the retrieval enhancement generation process, and a hard-constraint decoding space is simultaneously constructed to output a natural language report containing sentence-by-sentence evidence anchors, specifically including: Based on the entity names and predefined logical connectors and general grammatical units in the full-link tracing index, a dynamic decoding constraint set is constructed. At each time step of the retrieval enhancement generation process, the subset of legal candidate lemmas for the current moment is determined from the dynamic decoding constraint set, in conjunction with the retrieved context information. The probability distribution of the large language model called by the retrieval enhancement generation process at the output layer is numerically corrected. For key biological entities and causal predicates, the generation probability of all words outside the subset of legal candidate words is set to a preset value. For non-key modifying words, a preset attenuation penalty coefficient is applied. Decoding is performed within the fact boundary defined by the full-link tracing index to generate a natural language report. The generated text stream of the natural language report is monitored in real time; when a preset number of consecutive tokens cannot be found in the full-link traceability index, a consistency circuit breaker mechanism is triggered to forcibly terminate the current generation process and re-trigger the step of obtaining the logical skeleton template.

8. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The multidimensional confidence score vector includes the following dimensions: The strength of evidence score is a composite score based on the statistical p-value and effect size. Logical fit, characterizing the degree of topological matching between the candidate mechanism hypothesis chain and the selected logical skeleton template; Cross-modal consistency represents the directional consistency of evidence across different omics levels; The risk score for rebuttal is a negative score based on the penalty weights applied to the identified conflicting evidence and potential rebuttal items.

9. The method for generating biological mechanism hypotheses according to claim 1, characterized in that, The method is configured for use in drug repositioning scenarios, wherein: The logical skeleton template is configured as a ternary structure representing the association between drug, target, and phenotype; The identification of potential counter-evidence entries is specifically configured to identify records in the drug side effects database that conflict with the target treatment phenotype.

10. A biological mechanism hypothesis generation system, characterized in that, include: The graph construction module is configured to acquire multi-scale data, perform cross-modal feature dimensionality reduction and covariate correction operations, and calculate statistical association evidence in a unified metric space to construct or incrementally update a weighted directed evidence graph. An adaptive skeleton management module is configured to store preset logical skeleton templates and mine exploratory topological structures based on topological information entropy to update the templates; The graph mining engine is configured to perform constrained path search and biological entity parameter mapping based on topology. The source tracing, reduction, and scoring engine is configured to generate a full-link source tracing index for the candidate mechanism hypothesis chain, perform hierarchical reduction and confidence correction on conflicting evidence based on preset evidence pyramid rules and omics context matching degree, and calculate a multi-dimensional confidence score vector by identifying potential counter-evidence items from a preset database and applying penalty weights based on the corrected chain. Then, the pre-configured scoring model outputs the screened candidate hypotheses according to the multi-dimensional confidence score vector. The retrieval enhancement generation and hard constraint module is configured to start retrieval enhancement generation based on the full-link traceability index, and use the decoding space hard constraint controller to perform numerical correction of the generation probability with soft and hard penalties at the decoding layer to output a report, and includes a consistency circuit breaker to force a rollback when the evidence is lost. The feedback loop module is enhanced and configured to train the reward model based on user interaction data and update system parameters.

Images