Intelligent retrieval reasoning method and system for strain development

Abstract

The invention belongs to the cross technical field of artificial intelligence and synthetic biology, and discloses an intelligent retrieval reasoning method and system oriented to strain development, and the method comprises the steps: collecting a document in a portable document format from an open acquisition database, carrying out document analysis processing to generate a lightweight markup language file, carrying out semantic splitting to obtain text paragraphs, and storing the text paragraphs in the open acquisition database; forming a data set; taking the split text paragraphs as input, generating a sparse vector and a dense vector for each text paragraph by utilizing a retrieval plug-in, establishing a vector mixed index framework, and respectively storing the vector mixed index framework in a database; an application interface layer is built, and a data interface used for full-text retrieval and paragraph precise reading is provided. According to the method, a general technology and a synthetic biological strain development whole process are deeply fused, and strain development key information such as gene modification, metabolism regulation and control, culture medium setting and fermentation conditions is accurately matched; and the DBTL whole-process operation efficiency of a biological manufacturing enterprise and an invention mechanism in strain development is obviously optimized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and synthetic biology, and in particular to an intelligent retrieval and reasoning method and system for bacterial strain development. Background Technology

[0002] Synthetic biology aims to achieve sustainable biomanufacturing through engineered microbial cell factories, but its development is limited by the complexity of metabolic engineering and the time-consuming nature of DBTL cycles.

[0003] Traditional constraint modeling methods rely on manual intervention, resulting in low efficiency. While existing large-scale language models (such as GPT-4) possess knowledge extraction capabilities, their general corpora are insufficient to cover cutting-edge knowledge in synthetic biology and are prone to "illusion errors." Retrieval-enhanced generation (RAG) technology alleviates some of these problems through dynamic knowledge bases, but early versions based on vector retrieval (such as SynBioGPT v1.0) still suffer from the defect of retrieving semantically similar but context-independent documents, affecting accuracy.

[0004] At the same time, because the RAG retrieval process is a "static retrieval" that is completed in one go, it does not have the ability to dynamically adjust the overall retrieval process based on the answers to sub-questions.

[0005] Therefore, how to provide an intelligent retrieval and reasoning method and system for strain development is an urgent problem to be solved. Summary of the Invention

[0006] This invention provides an intelligent retrieval and reasoning method and system for strain development, in order to solve the aforementioned technical problems existing in the prior art.

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0008] According to a first aspect of the present invention, an intelligent retrieval and reasoning method for strain development is provided.

[0009] In one embodiment, the intelligent retrieval and reasoning method for strain development includes:

[0010] Portable Document Format (PDF) documents are collected from open access databases, processed through document parsing to generate lightweight markup language files, and then semantically segmented into text paragraphs to form a dataset.

[0011] Using the split text segments as input, a retrieval plugin is used to generate sparse and dense vectors for each text segment, a vector hybrid index framework is established, and the vectors are stored separately in the database.

[0012] An application interface layer is built to provide data interfaces for full-text search and paragraph in-depth reading. Based on user questions, external tools are dynamically invoked to write prompt words for input into the language interaction model. The language interaction model is then used to autonomously perform query decomposition and content reasoning to generate the final answer to the user's input query.

[0013] In one embodiment, the process of collecting portable document format documents from an open access database, parsing them to generate lightweight markup language files, and semantically splitting them into text paragraphs to form a dataset includes:

[0014] Collect open-access portable document format documents, along with the unique digital object identifiers for each portable document format, and convert them into a standardized dataset format;

[0015] Use open-source document parsing tools to convert portable document format documents into lightweight markup language files;

[0016] The paragraph structure of lightweight markup language files is extracted using a document layout processing library, and references, sensitive information, headers and footers are removed to generate a refined lightweight markup language column.

[0017] Extract the directory from the lightweight markup language column and store the directory information in a separate directory column.

[0018] In one embodiment, the open-source document parsing tool disables its built-in image processing functions and enables optical character recognition and table enhancement functions during operation.

[0019] In one embodiment, the retrieval plugin includes a sparse vector plugin based on the best matching word frequency algorithm and a dense vector plugin based on the fine-tuning encoder model; the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

[0020] In one embodiment, the step of dynamically calling external tools to write prompts based on user questions and inputting them into a language interaction model, and then using the language interaction model to autonomously perform query decomposition and content reasoning to generate the final answer to the user's input query includes:

[0021] The code for initializing the lightweight AI agent framework generates intelligent agents, preventing redundant resource allocation.

[0022] The tool leverages data applications to quickly build tools that facilitate user interaction, acquires interaction sessions and chat history data, stores it in a local database, and integrates with the business development center's single sign-on service through open authorization to enable third-party account access and authentication.

[0023] Extract user input queries from historical chat logs, break them down into multiple sub-questions, and convert them into keyword search instructions. Use a language interaction model to retrieve the answers to each sub-question. When the total number of search rounds reaches the limit or all sub-questions are answered, the search ends, and all search results are summarized to generate the final answer.

[0024] According to a second aspect of the present invention, an intelligent retrieval and reasoning system for strain development is provided.

[0025] In one embodiment, the intelligent retrieval and reasoning system developed for bacterial strains includes:

[0026] The data acquisition and cleaning module is used to collect portable document format documents from open access databases, process the documents through parsing, generate lightweight markup language files, and semantically split them into text paragraphs to form a dataset.

[0027] The index creation and storage module is used to take the split text segments as input, use the retrieval plugin to generate sparse and dense vectors for each text segment, establish a vector hybrid index framework, and store them separately in the database.

[0028] The service query reasoning module is used to build the application interface layer, providing data interfaces for full-text search and paragraph in-depth reading. It also dynamically calls external tools to write prompt words based on user questions and inputs them into the language interaction model. The language interaction model then autonomously performs query decomposition and content reasoning to generate the final answer to the user's input query.

[0029] In one embodiment, the data acquisition and cleaning module includes: a document acquisition module, a format conversion module, a document simplification module, and a directory storage module, wherein,

[0030] The document acquisition module is used to acquire open-access portable document format documents and the unique digital object identifier of each portable document format document, and convert them into a standardized dataset format;

[0031] The format conversion module is used to convert portable document format documents into lightweight markup language files using open-source document parsing tools;

[0032] The document simplification module is used to extract the paragraph structure of lightweight markup language files using a document layout processing library, and remove citations, sensitive information, headers and footers to generate a refined lightweight markup language column.

[0033] The directory storage module is used to extract the directory from the lightweight markup language column and store the directory information in a separate directory column.

[0034] In one embodiment, the open-source document parsing tool disables its built-in image processing functions and enables optical character recognition and table enhancement functions during operation.

[0035] In one embodiment, the retrieval plugin includes a sparse vector plugin based on the best matching word frequency algorithm and a dense vector plugin based on the fine-tuning encoder model; the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

[0036] In one embodiment, the service query inference module includes: an initialization module, a service deployment module, and an inference output module, wherein,

[0037] The initialization module is used to initialize the code-generating intelligent agent of the lightweight AI agent framework to prevent duplicate resource allocation.

[0038] The service deployment module is used to leverage data application rapid construction tools to promote user interaction, acquire interaction sessions and chat history data, store them in a local database, and integrate with the single sign-on service of the business development center through open authorization to enable third-party account access and authentication.

[0039] The reasoning output module is used to extract user input queries from historical chat records, break them down into multiple sub-questions, convert them into keyword search instructions, and use a language interaction model to search for answers to each sub-question. When the total number of search rounds reaches the upper limit or all sub-questions are answered, the search ends, and all search results are summarized to generate the final answer.

[0040] According to a third aspect of the present invention, a computer device is provided.

[0041] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0042] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0043] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.

[0044] The technical solutions provided by the embodiments of this invention can include the following beneficial effects: Deeply integrating general technologies (search enhancement, keyword decomposition, and dynamic result adjustment) with the entire process of synthetic biology strain development; constructing a search index based on synthetic biology literature to accurately match key information in strain development such as gene modification, metabolic regulation, culture medium settings, and fermentation conditions; significantly optimizing the DBTL (design-build-test-learn) process efficiency of biomanufacturing enterprises and invention institutions in strain development, effectively reducing trial-and-error costs; significantly reducing the consumption of manpower and resources in the experimental process, while providing standardized data interfaces and verification support for the development of AI-driven synthetic biology tools, powerfully promoting the transformation and upgrading of the biomanufacturing field towards an algorithm-driven paradigm.

[0045] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0046] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0047] Figure 1 This is a flowchart illustrating an intelligent retrieval and reasoning method for strain development according to an exemplary embodiment;

[0048] Figure 2 This is a block diagram illustrating the principle of an intelligent retrieval and reasoning system for strain development, according to an exemplary embodiment.

[0049] Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment;

[0050] Figure 4 This is a performance comparison of different inference backends (DeepSeekV3, Gemini-2.0-flash, Claude-3.7-sonnet) in synthetic biology tasks, as illustrated by an exemplary embodiment. Detailed Implementation

[0051] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some embodiments may include or substitute parts and features of other embodiments. The scope of the embodiments herein encompasses the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0052] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientations or positional relationships are based on the orientations or positional relationships shown in the accompanying drawings. They are used solely for the convenience of describing the document and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements; they can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0053] In this document, unless otherwise stated, the term "multiple" means two or more.

[0054] In this article, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0055] In this article, the term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0056] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0057] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0058] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0059] Figure 1 An embodiment of an intelligent retrieval reasoning method for strain development according to the present invention is shown.

[0060] In this optional embodiment, the intelligent retrieval and reasoning method developed for microbial species includes:

[0061] Step S101: Collect portable document format (PDF) documents from open access databases, process the documents to generate lightweight markup language files, and semantically split them into text paragraphs to form a dataset;

[0062] Step S102: Using the split text segments as input, use the retrieval plugin to generate sparse vectors and dense vectors for each text segment, establish a vector hybrid index framework, and store them in the database respectively.

[0063] Step S103: Build the application interface layer, providing data interfaces for full-text search (FULLTEXT_SEARCH interface) and paragraph close reading (SECTION_DETAIL_READER interface). Based on the user's question, dynamically call external tools to write prompt words and input them into the language interaction model (SynBioGPTv2.0, the first large language model expert system in the field of biomanufacturing, focusing on data integration and intelligent question answering in the fields of synthetic biology and metabolic engineering. It significantly improves the accuracy of answers through search enhancement generation technology and supports core functions such as strain modification, gene mutation analysis, and metabolic pathway design). Utilize the language interaction model to autonomously perform query decomposition and content reasoning to generate the final answer to the user's input query.

[0064] In this optional embodiment, when collecting portable document format documents from an open-access database, processing them through document parsing to generate lightweight markup language files, and semantically splitting them into text paragraphs to form a dataset, the following steps can be taken: collecting open-access portable document format documents and the unique digital object identifiers of each portable document format document, converting them into a standardized dataset format; using open-source document parsing tools to convert portable document format documents into lightweight markup language files; using a document layout processing library to extract the paragraph structure of the lightweight markup language files, and removing citations, sensitive information, headers, and footers to generate refined lightweight markup language columns; extracting the table of contents from the lightweight markup language columns, and storing the table of contents information in a separate table of contents column.

[0065] In this alternative embodiment, the open-source document parsing tool disables its built-in image processing, optical character recognition, and table enhancement features during operation.

[0066] In this optional embodiment, the retrieval plugin includes a sparse vector plugin based on the best matching term frequency algorithm (BM25 term frequency algorithm) and a dense vector plugin based on a fine-tuned encoder model (fine-tuned ModernBERT-based large model); the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

[0067] ModernBERT is an advanced encoder-only model that, as the successor to BERT, achieves significant improvements in several aspects. ModernBERT comprehensively surpasses the older BERT model, supporting token lengths of up to 8192, and boasting better downstream performance and faster processing speed. The fine-tuning process primarily involves collecting question-answer pairs from specified documents, and then performing general fine-tuning based on these pairs to ensure that the vector distribution generated by the fine-tuned model more closely resembles the distribution of synthetic biological information.

[0068] In this optional embodiment, when the prompt words are dynamically written and input into the language interaction model based on the user's question using external tools, and the language interaction model autonomously performs query decomposition and content reasoning to generate the final answer to the user's input query, a lightweight AI agent framework code generation intelligent agent can be initialized to prevent redundant resource allocation; a data application rapid construction tool is used to promote user interaction, obtain interaction sessions and chat history data, store them in a local database, and integrate with the single sign-on service of the business development center through open authorization to achieve third-party account access and authentication; user input queries in historical chat records are extracted, decomposed into multiple sub-questions, and converted into keyword retrieval instructions (the language interaction model is directly instructed to decompose the keywords according to the given query and context, and return them in English), and the language interaction model is used to retrieve the answers to each sub-question. When the total number of retrieval rounds reaches the upper limit or all sub-questions are answered, the retrieval ends, and all retrieval results are summarized to generate the final answer.

[0069] Figure 2 An embodiment of an intelligent retrieval and reasoning system for strain development according to the present invention is shown. In this optional embodiment, the intelligent retrieval and reasoning system for strain development includes:

[0070] The data acquisition and cleaning module 201 is used to collect portable document format (PDF) documents from open access databases, process the documents to generate lightweight markup language files, and semantically split them into text paragraphs to form a dataset.

[0071] The index creation and storage module 202 is used to take the split text segments as input, use the retrieval plugin to generate sparse vectors and dense vectors for each text segment, establish a vector hybrid index framework, and store them separately in the database.

[0072] The service query reasoning module 203 is used to build the application interface layer, providing data interfaces for full-text search and paragraph in-depth reading. It dynamically calls external tools to write prompt words based on user questions and inputs them into the language interaction model. The language interaction model autonomously performs query decomposition and content reasoning to generate the final answer to the user's input query.

[0073] In this optional embodiment, the data acquisition and cleaning module includes: a document acquisition module, a format conversion module, a document simplification module, and a directory storage module. The document acquisition module is used to acquire open-access portable document format documents and the unique digital object identifiers of each portable document format document, and convert them into a standardized dataset format. The format conversion module is used to convert portable document format documents into lightweight markup language files using open-source document parsing tools. The document simplification module is used to extract the paragraph structure of the lightweight markup language files using a document layout processing library, and remove citations, sensitive information, headers, and footers to generate a refined lightweight markup language column. The directory storage module is used to extract the directory from the lightweight markup language column and store the directory information in a separate directory column.

[0074] In this alternative embodiment, the open-source document parsing tool disables its built-in image processing functions and enables optical character recognition and table enhancement functions during operation.

[0075] In this optional embodiment, the retrieval plugin includes a sparse vector plugin based on the best matching word frequency algorithm and a dense vector plugin based on the fine-tuning encoder model; the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

[0076] In this optional embodiment, the service query inference module includes: an initialization module, a service deployment module, and an inference output module. The initialization module initializes the code-generating intelligent agent of the lightweight AI agent framework to prevent redundant resource allocation. The service deployment module utilizes data application rapid construction tools to facilitate user interaction, acquires interaction sessions and chat history data, stores it in a local database, and integrates with the single sign-on service of the business development center through open authorization to enable third-party account access and authentication. The inference output module extracts user input queries from historical chat records, breaks them down into multiple sub-questions, converts them into keyword search instructions, uses a language interaction model to retrieve answers to each sub-question, and ends the search when the total number of search rounds reaches the upper limit or all sub-questions are answered, summarizing all search results to generate the final answer.

[0077] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0078] The design of the intelligent retrieval reasoning system (integrated language interaction model: SynBioGPTv2.0) for strain development in this invention is divided into three stages: data collection and cleaning, index creation and storage, and index loading and external service deployment.

[0079] I. Data Acquisition and Cleaning:

[0080] In the initial phase, the system systematically collects open-access PDF (Portable Document Format) documents, each with a unique DOI (Digital Object Identifier), and converts them to the HuggingFace (AI Model Community and Platform) dataset format to preserve the original data. Then, Docling (an open-source document parsing tool) is used to convert the PDFs to Markdown (Lightweight Markup Language) files. During the process, the open-source document parsing tool disables its built-in image processing functions but enables optical character recognition and table enhancement. The resulting Markdown files are then integrated into the dataset. spaCyLayout (a document layout processing library) is used to extract paragraph structure and remove citations, sensitive information, headers, and footers, generating refined Markdown columns. Furthermore, spaCyLayout is used to extract the Table of Contents (TOC) from the refined Markdown and stores this information in a separate TOC column. Metadata obtained during the open-access collection process (such as author details, country, abstract, title, and category) is also expanded and stored as separate columns in the dataset.

[0081] II. Index Creation and Storage:

[0082] In a subsequent phase, SynBioGPT v2.0 established its indexing framework. The system uses the PostgreSQL BM25 retrieval plugin to create and store sparse indexes for each paragraph of each article.

[0083] III. External Service Deployment:

[0084] In the final stage, the platform shifts to providing external services. An API server based on FastAPI (a high-performance web framework) is built to handle user requests. Additionally, CodeAgent (a code generation intelligent agent) from smolagents (a lightweight AI agent framework) is pre-initialized to prevent redundant resource allocation during concurrent requests. On the front end, Streamlit (a rapid data application build tool) facilitates user interaction, and session and chat history data is stored in a local SQLite database (an embedded relational database). Finally, integration with BDC's (Bio Design Center) Single Sign-On (SSO) service is achieved through OAuth (Open Authorization), ensuring secure third-party account access and strong authentication.

[0085] Building upon the indexed database, an API layer is constructed, providing two key interfaces for full-text search: FULLTEXT_SEARCH and SECTION_DETAIL_READER. These interfaces are written as prompts in a FUNCTION-CALLING format and provided to the language interaction model. The language interaction model dynamically calls these interfaces during various stages, including question decomposition, determining whether the search results are sufficient to answer the question, and finally generating the answer. The FUNCTION-CALLING uses a completely generic approach, clearly describing the API's functionality and parameter assignments. The language interaction model returns the name of the function to be called and its parameters.

[0086] Specifically, the query processing flow executed by the application language interaction model includes the following steps:

[0087] Step 1: User inputs query (e.g., "How to optimize L-lysine production by Corynebacterium glutamicum?");

[0088] Step 2: The language interaction model generates a sub-question based on the user input and the information given in the prompts, such as "What are the metabolic engineering strategies for the production of L-lysine by Corynebacterium glutamicum?"

[0089] For example, a user might query: "How can I optimize the production of L-lysine by Corynebacterium glutamicum?"

[0090] The language interaction model breaks down the problem into a series of sub-problems to address various dimensions of the query, including:

[0091] 2.1 In the dimension of metabolic engineering optimization, the sub-question "What are the metabolic engineering strategies for the production of L-lysine by Corynebacterium glutamicum?" can be generated.

[0092] 2.2 In terms of genetic engineering optimization, the sub-question "What are the optimized genetic engineering methods for the production of L-lysine by Corynebacterium glutamicum?" can be generated.

[0093] 2.3 In the dimension of metabolic engineering optimization - cofactor engineering, the sub-question "What are the optimization strategies for the supply of NADPH (reduced nicotinamide adenine dinucleotide phosphate) in the L-lysine biosynthesis pathway of Corynebacterium glutamicum?" can be generated.

[0094] 2.4 In the dimension of fermentation engineering optimization / bioprocess optimization, the sub-question "What are the fermentation optimization strategies for L-lysine production by Corynebacterium glutamicum?" can be generated.

[0095] 2.5 In the dimension of metabolic engineering optimization - flux redirection, the sub-question "What are the strategies for optimizing carbon flux and reducing byproducts in the production of L-lysine by Corynebacterium glutamicum?" can be generated.

[0096] Step 3: The language interaction model calls FULLTEXT_SEARCH to perform full-text search, or calls SECTION_DETAIL_READER to retrieve the answers to the sub-questions in the previous step for a specific paragraph of a specific document.

[0097] Step 4: Based on the answers to the sub-questions, determine if there is sufficient information to answer the user's question. If so, construct the final answer directly; otherwise, generate further sub-questions (e.g., what are the fermentation optimization strategies for L-lysine production by Corynebacterium glutamicum? What are the strategies for optimizing carbon flux and reducing byproducts in L-lysine production by Corynebacterium glutamicum?), and repeat the previous step to obtain more information. (Each question and answer includes all relevant context, and the prompts ask whether the language interaction model considers the existing content sufficient to answer the user's question and provide a quantified percentage. If the percentage is greater than 0.8, it is considered satisfactory.)

[0098] Step 5: Repeat steps 1 to 4 of the query processing flow until the language interaction model obtains enough information to answer the user's question, or the total number of retrieval rounds reaches the set limit. At this point, summarize all search results and generate the final answer (the answer provides the reference information corresponding to all information).

[0099] like Figure 4 As shown, this invention provides a performance comparison of different inference backends (DeepSeekV3, Gemini-2.0-flash, and Claude-3.7-sonnet) in synthetic biology tasks. Specifically, SynBioGPT uses different models and knowledge bases to answer 100 manually compiled questions, providing statistical results. (A) Comparison of the performance of DeepSeekV3, Gemini-2.0-flash, and Claude-3.7-sonnet in answering synthetic biology questions with and without retrieval enhancement. (B) Performance comparison of the previous version (SynBioGPT v1.0) and the updated SynBioGPT v2.0.

[0100] To evaluate the effectiveness of large language models (LLMs) in synthetic biology applications, this invention uses a test set containing 100 questions, including 71 specific factual questions and 29 open-ended reasoning questions. These questions cover key areas of synthetic biology such as gene mutation, overexpression, co-expression, exogenous gene integration, precursor utilization, promoter dynamics, and competitive pathway exploration.

[0101] In this invention, the performance of the domain-specific language model SynBioGPT v2.0 was evaluated using three different inference backends—DeepSeekV3, Gemini-2.0-flash, and Claude-3.7-sonnet—to investigate the impact of backend selection on model results. Furthermore, this invention benchmarked SynBioGPT v2.0 against its predecessor, SynBioGPT v1.0, which was based on the Llama3-8B-Instruct model and had been identified as the best-performing model in previous tests.

[0102] Experimental results show that the introduction of search-enhanced generation (RAG) technology significantly improved accuracy, increasing it by 26%-37%. Figure 4 A). These findings highlight the crucial role of domain-specific context in optimizing model performance. Notably, although the core architecture of SynBioGPTv2.0 remained unchanged in the evaluation, its performance on synthetic biology tasks was significantly influenced by the choice of inference backend. Among the tested backends, Claude-3.7-sonnet achieved the highest accuracy, surpassing DeepSeekV3 and Gemini-2.0-flash. Figure 4 A). This superior performance demonstrates that Claude-3.7-sonnet possesses stronger reasoning capabilities and the ability to integrate domain-specific knowledge, making it particularly adept at handling the complex interdisciplinary nature of synthetic biology queries. In contrast, DeepSeekV3 and Gemini-2.0-flash exhibit higher error rates on specific questions requiring precise factual recall, suggesting potential limitations in their depth of knowledge or contextual understanding within this domain.

[0103] A comparative analysis of SynBioGPT v2.0 and v1.0 further highlights the improvements of the newer versions. On the same 100-question test set, SynBioGPT v1.0, based on Llama3-8B-Instruct, achieved an accuracy of 88% with 12% incorrect answers. SynBioGPT v2.0, paired with the Claude-3.7-sonnet backend, achieved an accuracy of 98% and reduced the error rate to only 2%. Figure 4 (B) This 10% improvement highlights the effectiveness of the improvements introduced in v2.0, which may include expanded training on synthetic biology-specific datasets, improved handling of complex queries, or optimization of model interactions with advanced backends. Notably, the 98% accuracy is only available with the Claude-3.7-sonnet configuration, demonstrating the crucial role the backend plays in unlocking the full potential of v2.0.

[0104] The data above highlights the interplay between model design and inference infrastructure when achieving high performance in specialized domains. The superior performance of Claude-3.7-sonnet suggests that its underlying architecture or training scheme may be better suited to the needs of synthetic biology, such as resolving technical terms or inferring multi-step biological processes. Conversely, the relatively weaker performance of DeepSeekV3 and Gemini-2.0-flash on specific problems indicates a need for further investigation into their limitations, which may be related to the scope of training data coverage or inference efficiency in terms of inference time.

[0105] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0106] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0107] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0108] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0109] Those skilled in the art will understand that all or part of the processes in the methods of 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, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0110] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. An intelligent retrieval and reasoning method for strain development, characterized in that, The method includes: Portable Document Format documents are collected from open access databases, parsed to generate lightweight markup language files, and semantically segmented into text paragraphs to form a dataset. Using the split text segments as input, a retrieval plugin is used to generate sparse and dense vectors for each text segment, a vector hybrid index framework is established, and the vectors are stored separately in the database. An application interface layer is built to provide data interfaces for full-text search and paragraph in-depth reading. Based on user questions, external tools are dynamically invoked to write prompt words for input into the language interaction model. The language interaction model is then used to autonomously perform query decomposition and content reasoning to generate the final answer to the user's input query.

2. The intelligent retrieval and reasoning method for strain-based development according to claim 1, characterized in that, The process involves collecting portable document format documents from open-access databases, parsing them to generate lightweight markup language files, and then semantically splitting them into text paragraphs to form a dataset, which includes: Collect open-access portable document format documents, along with the unique digital object identifiers for each portable document format, and convert them into a standardized dataset format; Use open-source document parsing tools to convert portable document format documents into lightweight markup language files; The paragraph structure of lightweight markup language files is extracted using a document layout processing library, and references, sensitive information, headers and footers are removed to generate a refined lightweight markup language column. Extract the directory from the lightweight markup language column and store the directory information in a separate directory column.

3. The intelligent retrieval and reasoning method for strain-based development according to claim 2, characterized in that, The open-source document parsing tool disables its built-in image processing functions and enables optical character recognition and table enhancement functions during operation.

4. The intelligent retrieval and reasoning method for strain development according to claim 1, characterized in that, The retrieval plugin includes a sparse vector plugin based on the best matching word frequency algorithm and a dense vector plugin based on the fine-tuning encoder model; the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

5. The intelligent retrieval and reasoning method for strain-based development according to claim 1, characterized in that, The process of dynamically calling external tools to write prompts based on user questions and inputting them into the language interaction model, and then using the language interaction model to autonomously perform query decomposition and content reasoning to generate the final answer to the user's input query includes: The code for initializing the lightweight AI agent framework generates intelligent agents, preventing redundant resource allocation. Leverage data applications to quickly build tools to facilitate user interaction, acquire interaction sessions and chat history data, store them in a local database, and integrate with the business development center's single sign-on service through open authorization to enable third-party account access and authentication; Extract user input queries from historical chat logs, break them down into multiple sub-questions, and convert them into keyword search instructions. Use a language interaction model to retrieve the answers to each sub-question. When the total number of search rounds reaches the limit or all sub-questions are answered, the search ends, and all search results are summarized to generate the final answer.

6. An intelligent retrieval and reasoning system for microbial strain development, characterized in that, The system includes: The data acquisition and cleaning module is used to collect portable document format documents from open access databases, process the documents through parsing, generate lightweight markup language files, and semantically split them into text paragraphs to form a dataset. The index creation and storage module is used to take the split text segments as input, use the retrieval plugin to generate sparse and dense vectors for each text segment, establish a vector hybrid index framework, and store them separately in the database. The service query reasoning module is used to build the application interface layer, providing data interfaces for full-text search and paragraph in-depth reading. It also dynamically calls external tools to write prompt words based on user questions and inputs them into the language interaction model. The language interaction model then autonomously performs query decomposition and content reasoning to generate the final answer to the user's input query.

7. The intelligent retrieval and reasoning system for strain development according to claim 6, characterized in that, The data acquisition and cleaning module includes: a document acquisition module, a format conversion module, a document simplification module, and a directory storage module, wherein... The document acquisition module is used to acquire open-access portable document format documents and the unique digital object identifier of each portable document format document, and convert them into a standardized dataset format; The format conversion module is used to convert portable document format documents into lightweight markup language files using open-source document parsing tools; The document simplification module is used to extract the paragraph structure of lightweight markup language files using a document layout processing library, and remove citations, sensitive information, headers and footers to generate a refined lightweight markup language column. The directory storage module is used to extract the directory from the lightweight markup language column and store the directory information in a separate directory column.

8. The intelligent retrieval and reasoning system for strain development according to claim 7, characterized in that, The open-source document parsing tool disables its built-in image processing functions and enables optical character recognition and table enhancement functions during operation.

9. The intelligent retrieval and reasoning system for strain development according to claim 6, characterized in that, The retrieval plugin includes a sparse vector plugin based on the best matching word frequency algorithm and a dense vector plugin based on the fine-tuning encoder model; the sparse vector plugin is used to generate a sparse vector for each text segment; the dense vector plugin is used to generate a dense vector for each text segment.

10. The intelligent retrieval and reasoning system for strain development according to claim 6, characterized in that, The service query inference module includes: an initialization module, a service deployment module, and an inference output module, wherein... The initialization module is used to initialize the code-generating intelligent agent of the lightweight AI agent framework to prevent duplicate resource allocation. The service deployment module is used to leverage data application rapid construction tools to promote user interaction, acquire interaction sessions and chat history data, store them in a local database, and integrate with the single sign-on service of the business development center through open authorization to enable third-party account access and authentication. The reasoning output module is used to extract user input queries from historical chat records, break them down into multiple sub-questions, convert them into keyword search instructions, and use a language interaction model to search for answers to each sub-question. When the total number of search rounds reaches the upper limit or all sub-questions are answered, the search ends, and all search results are summarized to generate the final answer.

Images