Method and device for realizing automatic evaluation based on biological process large model and intelligent agent, processor and computer readable storage medium

Abstract

The invention relates to a method for realizing automatic evaluation based on a biological process large model and an intelligent agent, which comprises the following steps of: constructing a biological process knowledge base, performing automatic thinking chain processing by utilizing a large reasoning model LRM, and decomposing query into a reasoning chain sequence containing a plurality of sub-problems; related documents are retrieved from the biological process knowledge base, enhanced cue words are generated in combination with a resorter, and local answers are generated through a large language model; and aggregating all sub-questions, local answers and reasoning contexts, and inputting a large language model to generate a final global response. According to the method and device for realizing automatic evaluation based on the biological process large model and the intelligent agent, the processor and the computer readable storage medium, the answer accuracy of biological process questions is effectively improved by introducing the multi-level domain knowledge base, the influence of logic chaos is eliminated, high-quality reasoning response is generated, and the evaluation efficiency is improved. And powerful guidance is provided for monitoring, optimization and fault diagnosis of the biological manufacturing process.

Description

Methods, devices, processors, and computer-readable storage media for automated evaluation based on large-scale biological process models and intelligent agents. Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to the interdisciplinary field of bioengineering, specifically to a method, apparatus, processor, and computer-readable storage medium for automated evaluation based on a large biological process model and intelligent agents. Background Technology

[0002] Large language models have demonstrated remarkable application potential in scientific fields such as mathematics, bioinformatics, and drug discovery in recent years. Biomanufacturing engineering, an interdisciplinary field combining biology, chemistry, and systems engineering, focuses on producing high-value products using microorganisms or cell factories. This process is extremely complex, requiring not only an understanding of underlying biological mechanisms such as genomes and metabolic pathways, but also real-time monitoring of changes in physicochemical parameters and comprehensive decision-making based on long-term expert operational experience.

[0003] However, directly applying existing general-purpose language models to reasoning in biomanufacturing processes still faces significant challenges. First, the specialized knowledge in biomanufacturing is highly dispersed across various academic documents, technical patents, and expert experience. General-purpose models, lacking systematic domain knowledge input, are prone to generating factual illusions when answering specialized questions. Second, the analysis of biological processes often requires handling multi-step causal inferences and dynamic constraints. Existing model architectures lack structured reasoning mechanisms, making it difficult to organically integrate fragmented retrieved knowledge. This results in logically incoherent or context-irrelevant answers, i.e., generating relevance illusions.

[0004] Furthermore, existing technologies lack standardized evaluation benchmarks for complex reasoning tasks in biomanufacturing. Current evaluation datasets are mostly limited to general mathematical problems or simple biological fact-finding questions, failing to cover real-world application scenarios in biomanufacturing that require the integration of multi-level knowledge (mechanisms, states, experiences). Due to the lack of fine-grained evaluation standards and automated judging mechanisms, researchers struggle to accurately quantify the model's shortcomings in domain-specific reasoning and cannot effectively guide model optimization and iteration. Therefore, there is an urgent need for an enhancement method that integrates domain knowledge and structured reasoning, along with a corresponding automated evaluation framework. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, processor and computer-readable storage medium for automated evaluation based on a large biological process model and intelligent agent, which meets the requirements of accurate quantification, refinement and wide applicability.

[0006] To achieve the above objectives, the present invention provides a method, apparatus, processor, and computer-readable storage medium for automated evaluation based on a large biological process model and an intelligent agent as follows: The method for automated evaluation based on a large biological process model and an intelligent agent is characterized by the following steps: (1) Constructing a biological process knowledge base, obtaining user-input queries in the field of biomanufacturing, and using a large reasoning model (LRM) for automated thought chain CoT processing to decompose the query into a reasoning chain sequence containing several sub-problems; (2) For each sub-problem in the reasoning chain, using a retrieval enhancement generation (RAG) module to retrieve relevant documents from the biological process knowledge base, and combining a reorderer to generate enhanced prompts, and generating local answers through a large language model; (3) Aggregating all sub-problems, local answers, and reasoning context, inputting them into a large language model to generate the final global response, and using a large referee model to perform granular decomposition and multi-dimensional index evaluation of the global response.

[0007] Preferably, step (1) specifically includes the following steps: (1.1) collecting multi-source biomanufacturing data and organizing it into a three-level bioprocess knowledge base, BioBase, consisting of biological mechanisms, process state characteristics, and expert heuristic experience; (1.2) receiving complex bioprocess queries input by users. (1.3) Analyze queries using the large-scale inference model LRM. Generate reasoning chain ,in Indicates the first Each reasoning step or sub-problem This represents the total number of steps.

[0008] Preferably, step (2) specifically includes the following steps: (2.1) For each sub-problem Calculate its vector embedding and perform a maximum inner product search in the biological process knowledge base to retrieve the Top-K relevant document fragments; (2.2) Utilize the cross-coding reorderer The document fragments are retrieved and reordered to select the most relevant document set. (2.3) System prompt words The reordered collection of documents and sub-problems splicing to construct enhanced prompts (2.4) Input the enhanced prompt words into the large language model for answering. Generate a local answer to the subproblem. .

[0009] Preferably, the step (2.2) of retrieving and reordering candidate documents specifically involves retrieving and reordering candidate documents according to the following formula: ; ;in, Represents a collection of knowledge base documents. The embedding vector represents the query. Represents a document fragment The embedding vector.

[0010] Preferably, in step (2.4), a local answer to the subproblem is generated. Specifically, it involves generating partial answers to subproblems based on the following formula. : in, This represents the final generated global response. Indicates system role prompts. Indicates sequence concatenation. Indicates the first The subproblems of each step and their corresponding local solutions.

[0011] Preferably, step (3) specifically includes the following steps: (3.1) using a granularity decomposition algorithm to process the global response generated by the model. Divided into atomic semantic units (3.2) Using the large-scale referee model, based on the original query, standard reference answer and role definition, perform binary classification labeling on each atomic semantic unit and generate relevant labels respectively. and factual labels (3.3) Calculate the evaluation metrics, namely the relevance error rate R-Errors, the factual error rate F-Errors, and the overall accuracy.

[0012] Preferably, in step (3.3), the calculation of the relevance error rate R-Errors and the factual error rate F-Errors is specifically performed as follows: the relevance error rate R-Errors and the factual error rate F-Errors are calculated according to the following formulas: in, The total number of test samples, For the first The number of atomic units decomposed from each sample; Indicates the first The context of each unit is irrelevant. Indicates the first There are factual errors in one of the units.

[0013] The device for automated evaluation based on a large biological process model and intelligent agents is characterized in that it includes: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, which, when executed by the processor, implement the various steps of the aforementioned method for automated evaluation based on a large biological process model and intelligent agents.

[0014] The processor for automated evaluation based on a large biological process model and intelligent agents is characterized in that the processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the various steps of the aforementioned method for automated evaluation based on a large biological process model and intelligent agents.

[0015] The main feature of this computer-readable storage medium is that it stores a computer program thereon, which can be executed by a processor to implement the various steps of the above-described method for automated evaluation based on a large biological process model and intelligent agents.

[0016] This invention employs a method, apparatus, processor, and computer-readable storage medium for automated evaluation based on a large-scale biological process model and intelligent agents. By introducing a multi-level domain knowledge base, it effectively improves the accuracy of answers to biological process questions and reduces factual and relevance illusions. Automated thought chain decomposition and retrieval enhance generation, eliminating the influence of logical confusion. Furthermore, by combining expert evaluation mechanisms, the generated results are rigorously controlled, resulting in high-quality reasoning responses. This provides powerful guidance for the monitoring, optimization, and fault diagnosis of biomanufacturing processes. Attached Figure Description

[0017] Figure 1 is a schematic diagram of the overall architecture of the method for automated evaluation based on a large biological process model and intelligent agents according to the present invention.

[0018] Figure 2 is a flowchart of the knowledge base construction process of the method for automated evaluation based on a large biological process model and intelligent agents according to the present invention.

[0019] Figure 3 is a schematic diagram of the evaluation framework of the method for automated evaluation based on a large biological process model and intelligent agents according to the present invention.

[0020] Figure 4 is a comparison of ablation experiment results under different module missing conditions using the method of automated evaluation based on a large biological process model and intelligent agent according to the present invention.

[0021] Figure 5 is a comparison of the correlation error rate trends of the present invention and the benchmark model under different temperature parameters. Detailed Implementation

[0022] To more clearly describe the technical content of the present invention, the following description is provided in conjunction with specific embodiments.

[0023] The present invention provides a method for automated evaluation based on a large biological process model and an intelligent agent, which includes the following steps: (1) constructing a biological process knowledge base, obtaining user-input queries in the field of biomanufacturing, and using a large reasoning model (LRM) to perform automated thought chain CoT processing to decompose the query into a reasoning chain sequence containing several sub-problems; (2) for each sub-problem in the reasoning chain, using the retrieval enhancement generation (RAG) module to retrieve relevant documents from the biological process knowledge base, and combining it with a reorderer to generate enhanced prompt words, and generating local answers through a large language model; (3) aggregating all sub-problems, local answers, and reasoning context, inputting them into the large language model to generate the final global response, and using a large referee model to perform granular decomposition and multi-dimensional index evaluation of the global response.

[0024] As a preferred embodiment of the present invention, step (1) specifically includes the following steps: (1.1) collecting multi-source biomanufacturing data and organizing it into a three-level bioprocess knowledge base BioBase, consisting of biological mechanisms, process state characteristics, and expert heuristic experience; (1.2) receiving complex bioprocess queries input by users. (1.3) Analyze queries using the large-scale inference model LRM. Generate reasoning chain ,in Indicates the first Each reasoning step or sub-problem This represents the total number of steps.

[0025] As a preferred embodiment of the present invention, step (2) specifically includes the following steps: (2.1) For each sub-problem Calculate its vector embedding and perform a maximum inner product search in the biological process knowledge base to retrieve the Top-K relevant document fragments; (2.2) Utilize the cross-coding reorderer The document fragments are retrieved and reordered to select the most relevant document set. (2.3) System prompt words The reordered collection of documents and sub-problems splicing to construct enhanced prompts (2.4) Input the enhanced prompt words into the large language model for answering. Generate a local answer to the subproblem. .

[0026] In a preferred embodiment of the present invention, the step (2.2) of retrieving and reordering candidate documents specifically involves retrieving and reordering candidate documents according to the following formula: ; ;in, Represents a collection of knowledge base documents. The embedding vector represents the query. Represents a document fragment The embedding vector.

[0027] In a preferred embodiment of the present invention, the step (2.4) generates a partial answer to the subproblem. Specifically, it involves generating partial answers to subproblems based on the following formula. : in, This represents the final generated global response. Indicates system role prompts. Indicates sequence concatenation. Indicates the first The subproblems of each step and their corresponding local solutions.

[0028] As a preferred embodiment of the present invention, step (3) specifically includes the following steps: (3.1) using a granularity decomposition algorithm to process the global response generated by the model. Divided into atomic semantic units (3.2) Using the large-scale referee model, based on the original query, standard reference answer and role definition, perform binary classification labeling on each atomic semantic unit and generate relevant labels respectively. and factual labels (3.3) Calculate the evaluation metrics, namely the relevance error rate R-Errors, the factual error rate F-Errors, and the overall accuracy.

[0029] In a preferred embodiment of the present invention, the calculation of the relevance error rate R-Errors and the factual error rate F-Errors in step (3.3) is specifically as follows: the relevance error rate R-Errors and the factual error rate F-Errors are calculated according to the following formulas: in, The total number of test samples, For the first The number of atomic units decomposed from each sample; Indicates the first The context of each unit is irrelevant. Indicates the first There are factual errors in one of the units.

[0030] The apparatus of the present invention for automated evaluation based on a large biological process model and intelligent agent includes: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the various steps of the method for automated evaluation based on a large biological process model and intelligent agent are implemented.

[0031] The processor of the present invention, which is based on a large biological process model and an intelligent agent for automated evaluation, is configured to execute computer-executable instructions, which, when executed by the processor, implement the various steps of the above-described method for automated evaluation based on a large biological process model and an intelligent agent.

[0032] The computer-readable storage medium of the present invention stores a computer program that can be executed by a processor to implement the various steps of the above-described method for automated evaluation based on a large biological process model and an intelligent agent.

[0033] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, processor and computer-readable storage medium based on a large biological process model and intelligent agent with high accuracy, strong logic and a complete evaluation system. By combining enhanced retrieval with structured thinking chains, it solves the problems of knowledge gaps and logical reasoning confusion in the field of biomanufacturing of general large models, significantly reduces factual illusions and relevance illusions, and improves the accuracy and interpretability of biological process reasoning.

[0034] This invention proposes a method based on a large biological process model and intelligent agents, aiming to address the dual challenges of domain knowledge scarcity and confusing reasoning logic faced by general-purpose large language models in the field of biomanufacturing. This method achieves high-precision, low-illusion reasoning for complex biological process problems by constructing a multi-level knowledge base, introducing Automated Chain of Thought (CoT) reasoning, and combining retrieval-enhanced generation (RAG) technology.

[0035] Please refer to Figure 1. The method based on the large biological process model and intelligent agent includes the following steps: (1) constructing a biological process knowledge base, obtaining user input queries in the field of biological manufacturing, using a large reasoning model (LRM) to perform automated thought chain CoT processing, and decomposing the query into a reasoning chain sequence containing several sub-problems; (2) for each sub-problem in the reasoning chain, using the retrieval enhancement generation (RAG) module to retrieve relevant documents from the biological process knowledge base, and combining it with a reorderer to generate enhanced prompt words, and generating local answers through a large language model; (3) aggregating all sub-problems, local answers and reasoning context, inputting them into the large language model to generate the final global response, and using a referee large model to perform granular decomposition and multi-dimensional index evaluation of the global response.

[0036] As a preferred embodiment of the present invention, step (1) specifically includes the following steps: (1.1) collecting multi-source biomanufacturing data, including scientific literature, fermentation protocols and expert experience, and organizing them into a three-level bioprocess knowledge base BioBase, consisting of biological mechanisms, process state characteristics and expert heuristic experience; (1.2) receiving complex bioprocess queries input by users. (1.3) Analyze queries using a large-scale inference model (LRM) and generate inference chains. ,in Indicates the first Each reasoning step or sub-problem The total number of steps aims to decouple complex queries into subtasks such as causal inference, constraint handling, or action planning.

[0037] As a preferred embodiment of the present invention, step (2) specifically includes the following steps: (2.1) For each sub-problem Calculate its vector embedding and retrieve the Top-K relevant document fragments in the biological process knowledge base; (2.2) Utilize the cross-coding reorderer (2.3) Semantically align and score the retrieved document fragments to select the most relevant document set; (2.4) Construct enhanced prompt words containing system roles, retrieved document sets, and sub-questions; (2.5) Input the enhanced prompt words into the answer language model to generate a local answer to the sub-question. As a preferred embodiment of the present invention, relevant documents are retrieved in the following manner: in, This represents the initial set of document fragments retrieved. Represents a collection of knowledge base documents, where each It is a text paragraph; Indicates the use of a domain encoder Calculated subqueries The embedding vector; Represents a document fragment The embedding vector; preliminary retrieval results are obtained using maximum inner product search; documents are reordered as follows: in, This represents a collection of documents after reordering. This represents the subproblem computed by the cross-coding reorderer. With Documents The semantic relevance score is used to eliminate noisy documents that are semantically related but logically mismatched.

[0038] In a preferred embodiment of the present invention, the final global response is generated in the following manner:

[0039] in, This represents the final generated global response; Represents a large model of parameterized responses; Indicates system role prompts; Indicates sequence concatenation; Indicates the first By piecing together the reasoning context of all steps, the sub-problems of each step and their corresponding local answers are combined to ensure that the global response has a complete logical chain.

[0040] As a preferred embodiment of the present invention, the evaluation process in step (3) specifically includes the following steps: (3.1) using a granularity decomposition algorithm to divide the global response generated by the model into atomic semantic units Typically, it is based on sentences or logical clauses; (3.2) Utilizing the large-scale referee model, based on the original query The standard reference answer and role definition are used to perform binary classification annotation on each atomic semantic unit, and relevant labels are generated respectively. and factual labels (3.3) Calculate the correlation error rate R-Errors, the factual error rate F-Errors, and the overall accuracy.

[0041] As a preferred embodiment of the present invention, the evaluation index is calculated in the following manner: in, The total number of test samples; For the first The number of atomic units decomposed from each sample; Indicates the first In the nth sample The contexts of individual units are irrelevant, which creates an illusion of relevance.

[0042] in, Indicates the first In the nth sample The unit contains factual errors, that is, factual illusions exist.

[0043] in, This is an indicator function that indicates a sample is considered correct only if all atomic units in an answer contain neither relevance errors nor factual errors.

[0044] In a specific embodiment of this invention, the biomanufacturing process involves a multi-level knowledge system and complex reasoning logic. Due to limitations in training data, general-purpose large-scale models often lack in-depth knowledge of specific domains and are prone to errors when handling multi-step reasoning. This technical solution proposes a large-scale model bioprocess reasoning method (BioManner) and a corresponding evaluation framework (BioMeter) based on retrieval enhancement and thought chain analysis. First, by constructing a multi-level knowledge base, BioBase, the problem of knowledge scarcity is solved. Second, by utilizing automated thought chain analysis technology to decompose complex problems into ordered sub-tasks, combined with targeted retrieval enhancement generation, the problems of confused reasoning logic and lack of evidence are solved. Finally, automated evaluation ensures the reliability of the generated results.

[0045] The method based on the biological process big model and intelligent agent mainly includes the following three main stages: (1) having a multi-source data acquisition and knowledge base construction module; (2) having a reasoning generation module based on thought chain and retrieval enhancement; and (3) having an automated evaluation module based on granular decomposition.

[0046] 1. Multi-source data acquisition and knowledge base construction (BioBase) The data sources of this invention cover key information throughout the entire lifecycle of biomanufacturing, as shown in Figure 2. To address the problem of scattered and unstructured knowledge in biomanufacturing, we implemented a rigorous data acquisition and cleaning process.

[0047] The first step is data collection. Data sources include peer-reviewed scientific publications, industrial fermentation protocols, technical guidelines, and high-quality Q&A from expert forums. This data covers the entire process, from strain screening and fermentation process control to downstream separation and purification.

[0048] The second step is data cleaning and preprocessing. For scientific documents in PDF format, OCR and layout-aware parsing techniques are used to extract text content, mathematical formulas, and chemical symbols, while tables and images are removed to maintain text coherence. For forum Q&A data, a large-scale model cleaning pipeline is used to remove special characters and resolve issues with unclear pronoun references, and manual verification is performed to ensure factual accuracy.

[0049] The third step is knowledge layering and indexing. In order to reflect the complexity of real biological systems, we divide the knowledge base BioBase into three layers: (3.1) Biological mechanism layer: containing genomics, enzymology characteristics and metabolic pathway data, used to explain the intrinsic reasons for biological reactions.

[0050] (3.2) Process state characteristic layer: includes operating parameters (such as pH, DO, temperature), kinetic curves and quality control indicators, used to describe the physicochemical state of the system.

[0051] (3.3) Expert-inspired experience layer: includes fault diagnosis rules, amplification effect principles and best practices, which are used to guide actual operation.

[0052] The fourth step is vectorization and storage. This involves storing the processed document fragments. Encoder with fine-tuning of input domain Generate embedding vectors It is stored in a vector database to support efficient retrieval later.

[0053] 2. Reasoning Generation Based on Thinking Chain and Retrieval Enhancement (BioManner): Since biological process queries often implicitly involve multi-step reasoning logic (e.g., "How to adjust feeding strategy based on DOs" involves defining DOs, analyzing metabolic effects, and formulating strategies), direct retrieval often fails to yield complete answers. This invention designs a mechanism for the collaborative operation of automated thinking chain (CoT) and retrieval enhancement (RAG), as shown in Figure 3.

[0054] The first step is to automate the decomposition of the thought process. This involves receiving user queries. Generate reasoning chains using large-scale reasoning models Each This represents a specific reasoning step or subquery. For example, for the query "What is the mechanism by which dissolved oxygen concentration (DO) affects aerobic fermentation?", LRM breaks it down into: (1) What is aerobic fermentation? (2) How is dissolved oxygen concentration defined and measured? (3) What are the effects of dissolved oxygen concentration on microbial metabolism? The second step is step-by-step retrieval and reordering. For each subquery... Calculate its query vector Formula (1) is used to retrieve the Top-K relevant documents in BioBase. Subsequently, the cross-coding reorderer in formula (2) is used. The search results are finely scored, noisy documents are removed, and a high-quality evidence set is obtained. .

[0055] The third step is to generate partial answers. This involves assigning system roles... Collection of evidence Subquery Concatenate to enhance prompts Input the answer and the large model generates local answers. This step ensures that each link in the chain of reasoning is supported by facts.

[0056] Step 4: Global response synthesis. This involves combining the subqueries and local answers from all steps. Aggregate and use formula (3) to generate the final global response. The response not only accurately answered the user's question but also demonstrated a clear reasoning path.

[0057] 3. Automated evaluation based on granular decomposition (BioMeter) In order to objectively measure the performance of the model in biological process reasoning, this invention proposes an automated evaluation framework based on a large referee model, as shown in Figure 3.

[0058] The first step is granular decomposition. This involves processing the long text responses generated by the model. Decomposed into atomic semantic units For example, independent sentences or clauses. The second step is context-aware judgment. The judging model receives the original question, the standard reference answer, and the unit to be evaluated, and judges whether the unit contains factual errors (conflicts with the reference answer or violates common sense) or relevance errors (unrelated to the question's intent).

[0059] The third step is to calculate the indicators. R-Errors, F-Errors, and Accuracy are calculated using formulas (4), (5), and (6) to quantify the "illusion" level and inference accuracy of the model.

[0060] Experimental Verification: To verify the effectiveness of the proposed method based on a large biological process model and intelligent agent, we conducted comprehensive comparative experiments on the constructed BioMeter benchmark dataset. The experiments mainly examined the model's accuracy and "illusion" control ability when handling complex biological process reasoning tasks, including factual errors and relevance errors.

[0061] 1. Main Experimental Comparison Analysis: The method of this invention was compared with the current mainstream general-purpose large language models. To more intuitively demonstrate the performance of each model in suppressing hallucinations, the factual error rate (F-Errors) and relevance error rate (R-Errors) of each model are statistically shown in Table 1.

[0062] As can be seen from Table 1: (1) Due to the lack of domain knowledge, the factual error rate (F-Errors) of general large models (such as DeepSeek and Qwen) is as high as 93.0% and 79.0%, respectively.

[0063] (2) Although simple retrieval enhancement models (such as ChatGPT and ChatGLM) reduce factual errors, the relevance error rate (R-Errors) is still as high as 19.0% and 46.0% due to the lack of structured reasoning.

[0064] (3) The method of the present invention (BioManner) reduces the factual error rate to 13.0% and the relevance error rate to 0.0%, which shows that the answers generated by BioManner are not only factually accurate, but also logically sound and fully in line with the reasoning habits of biomanufacturing experts.

[0065] 2. Module Validity Analysis To further verify the necessity of the two core modules, "Retrieval Enhancement Generation (RAG)" and "Automated Thinking Chain (CoT)," in this invention, we conducted an ablation experiment, and the results are shown in Figure 4 of the specification.

[0066] (1) Basic model (Qwen, DeepSeek): The factual error rate is extremely high, reaching 79% and 93% respectively. Due to the lack of domain knowledge, the model fabricates a lot of facts. The relevance error rate is moderate, at 31% and 42% respectively. The relevance of the answers is acceptable, but there is still a lot of nonsense.

[0067] (2) Retrieval-only models (Qwen-RR, DeepSeek-RR): The factual error rate remains high. Although external knowledge is introduced, without a chain of thought (CoT) to organize reasoning, the model often cannot correctly utilize this knowledge, resulting in no significant decrease in the factual error rate. The relevance error rate still exists. The relevance error rate of DeepSeek-RR has decreased to 16%, but Qwen-RR remains at around 30%. This shows that knowledge alone (RAG) is insufficient to guarantee logical coherence.

[0068] (3) BioManner: F-Errors were significantly reduced to 13%, indicating that RAG provided accurate sources of knowledge. R-Errors were 0, indicating that the CoT thinking chain ensured that the reasoning process was closely related to the problem and avoided going off-topic or logical confusion.

[0069] 3. Method Stability Analysis: Biomanufacturing process control places extremely high demands on the stability of the AI ​​system. We tested the changes in the relevance error rate of each model's responses under different generation temperature settings, and the results are shown in Figure 5 of the instruction manual.

[0070] (1) The left figure of Figure 5 shows the change of factual error rate with temperature: as the temperature rises, the error rate of the comparative model (such as DeepSeek-RR) fluctuates greatly, while the method of the present invention (BioManner) can maintain a relatively low and stable factual error rate at different temperatures, indicating that structured reasoning helps to lock in key knowledge points.

[0071] (2) The right side of Figure 5 shows the change in correlation error rate with temperature: as the temperature parameter increases from 0.0 to 1.3, the correlation error rate of the comparative model shows a dramatic fluctuation and a significant upward trend, indicating that it is extremely prone to logical collapse when dealing with generation uncertainty. In contrast, the curve of the method of this invention remains stable and close to zero.

[0072] Case Study: Taking "Dissolved Oxygen (DO) Influence Mechanism" as an example, BioManner's answer is as follows: "Dissolved oxygen concentration is a key factor affecting aerobic fermentation efficiency and product synthesis. It determines the microbial respiration pattern, metabolic energy utilization rate, and the dynamic balance between cell growth and product accumulation. Through strategies such as real-time monitoring, staged regulation, and oxygen supply system optimization, coordinated control of dissolved oxygen and other process parameters can be achieved. This not only increases product yield and quality but also reduces energy consumption and environmental risks, thereby achieving efficient, stable, and sustainable development of the fermentation process." This answer accurately covers knowledge at three levels: mechanism (respiration pattern), state (real-time monitoring), and experience (staged regulation), and is logically coherent without any misleading content.

[0073] The model is evaluated as follows: Comparing the inference results generated by BioManner with the standard answers provided by experts reveals that this method not only represents known biological mechanisms but also demonstrates strong logical reasoning capabilities. The sub-problem paths decomposed by the automated thought chain are highly consistent with the expert's thought process. For example, defining key parameters before analyzing the metabolic effects aligns with the gradual nature of cognition. Overall, this method effectively utilizes domain knowledge bases to support each step of reasoning and can eliminate irrelevant information through its thought chain structure, demonstrating its effectiveness and significance.

[0074] Compared with other methods, this approach constructs a reasoning architecture based on a large-scale biological process model and intelligent agents. By building a multi-level knowledge base encompassing biological mechanisms, process states, and expert experience, it introduces domain-specific knowledge injection to eliminate knowledge gaps in the general large-scale model. Then, it utilizes Automated Chain of Thought (CoT) and Retrieval Enhancement Generation (RAG) to decompose complex problems and retrieve evidence step-by-step, eliminating invalid retrieval contexts and reducing the illusion of relevance, thus achieving high-precision reasoning for complex biological processes. A structured reasoning module is employed to eliminate the influence of logical breaks and erroneous attributions. The results can accurately represent the causal logical relationships between all known biological process variables. Furthermore, it can identify deep reasoning paths that current general models cannot handle, supplementing the reasoning capabilities of the large-scale model by combining analysis with existing scientific literature and expert experience. In summary, through comparative analysis with existing benchmark models, our method can effectively represent existing expert experience from data and reasonably supplement it. Therefore, the proposed method based on a large-scale biological process model and intelligent agents is effective and practically significant.

[0075] For the specific implementation scheme of this embodiment, please refer to the relevant descriptions in the above embodiments, which will not be repeated here.

[0076] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0077] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0078] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0079] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0080] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The corresponding program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0081] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0082] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0083] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0084] This invention employs a method, apparatus, processor, and computer-readable storage medium for automated evaluation based on a large-scale biological process model and intelligent agents. By introducing a multi-level domain knowledge base, it effectively improves the accuracy of answers to biological process questions and reduces factual and relevance illusions. Automated thought chain decomposition and retrieval enhance generation, eliminating the influence of logical confusion. Furthermore, by combining expert evaluation mechanisms, the generated results are rigorously controlled, resulting in high-quality reasoning responses. This provides powerful guidance for the monitoring, optimization, and fault diagnosis of biomanufacturing processes.

[0085] In this specification, the invention has been described with reference to specific embodiments thereof. However, it will be apparent that various modifications and variations can be made without departing from the spirit and scope of the invention. Therefore, the specification and drawings should be considered illustrative rather than restrictive.

Claims

1. A method for automated evaluation based on a large biological process model and intelligent agents, characterized in that, The method includes the following steps: (1) Constructing a bioprocess knowledge base, obtaining user-input biomanufacturing domain queries, using a large-scale reasoning model (LRM) for automated thought chain CoT processing, and decomposing the query into a reasoning chain sequence containing several sub-problems; (2) For each sub-problem in the reasoning chain, using the retrieval enhancement generation (RAG) module to retrieve relevant documents from the bioprocess knowledge base, and combining it with a reorderer to generate enhanced prompt words, and generating local answers through a large language model; (3) Aggregating all sub-problems, local answers and reasoning context, inputting them into a large language model to generate the final global response, and using a large-scale referee model to perform granular decomposition and multi-dimensional index evaluation of the global response.

2. The method for automated evaluation based on a large biological process model and intelligent agent as described in claim 1, characterized in that, Step (1) specifically includes the following steps: (1.1) Collecting multi-source biomanufacturing data and organizing it into a three-level bioprocess knowledge base, BioBase, consisting of biological mechanisms, process state characteristics, and expert heuristic experience; (1.2) Receiving complex bioprocess queries input by users. (1.3) Analyze queries using the large-scale inference model LRM. Generate reasoning chain ,in Indicates the first Each reasoning step or sub-problem This represents the total number of steps.

3. The method for automated evaluation based on a large biological process model and intelligent agent as described in claim 1, characterized in that, Step (2) specifically includes the following steps: (2.1) For each sub-problem Calculate its vector embedding and perform a maximum inner product search in the biological process knowledge base to retrieve the Top-K relevant document fragments; (2.2) Utilize the cross-coding reorderer The document fragments are retrieved and reordered to select the most relevant document set. (2.3) System prompt words The reordered collection of documents and sub-problems splicing to construct enhanced prompts (2.4) Input the enhanced prompt words into the large language model for answering. Generate a local answer to the subproblem. 。 4. The method for automated evaluation based on a large biological process model and intelligent agent as described in claim 3, characterized in that, The step (2.2) of retrieving and re-ranking candidate documents specifically involves retrieving and re-ranking candidate documents according to the following formula: ； ;in, Represents a collection of knowledge base documents. The embedding vector represents the query. Represents a document fragment The embedding vector.

5. The method for automated evaluation based on a large biological process model and intelligent agent as described in claim 3, characterized in that, In step (2.4), a local answer to the subproblem is generated. Specifically, it involves generating partial answers to subproblems based on the following formula. ： in, This represents the final generated global response. Indicates system role prompts. Indicates sequence concatenation. Indicates the first The subproblems of each step and their corresponding local solutions.

6. The method for automated evaluation based on a large biological process model and intelligent agent according to claim 1, characterized in that, The aforementioned step (3) specifically includes the following steps: (3.1) using a granularity decomposition algorithm to process the global response generated by the model. Divided into atomic semantic units (3.2) Using the large-scale referee model, based on the original query, standard reference answer and role definition, perform binary classification labeling on each atomic semantic unit and generate relevant labels respectively. and factual labels (3.3) Calculate the evaluation metrics, namely the relevance error rate R-Errors, the factual error rate F-Errors, and the overall accuracy.

7. The method for automated evaluation based on a large biological process model and intelligent agent as described in claim 6, characterized in that, In step (3.3), the relevance error rate R-Errors and the factual error rate F-Errors are calculated as follows: The relevance error rate R-Errors and the factual error rate F-Errors are calculated according to the following formulas: in, The total number of test samples, For the first The number of atomic units decomposed from each sample; Indicates the first The context of each unit is irrelevant. Indicates the first There are factual errors in one of the units.

8. A device for automated evaluation based on a large biological process model and intelligent agents, characterized in that, The apparatus includes: a processor configured to execute computer-executable instructions; and a memory storing one or more computer-executable instructions, which, when executed by the processor, implement the steps of the method for automated evaluation based on a large biological process model and intelligent agents as described in any one of claims 1 to 7.

9. A processor for automated evaluation based on a large biological process model and intelligent agents, characterized in that, The processor is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the method for automated evaluation based on a large biological process model and intelligent agent as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method for automated evaluation based on a large biological process model and intelligent agent as described in any one of claims 1 to 7.

Images