Enzyme engineering multi-agent collaborative processing method based on monte carlo tree search

Abstract

This invention discloses a multi-agent collaborative processing method for enzyme engineering based on Monte Carlo tree search, relating to the fields of artificial intelligence and enzyme engineering technology. It includes: employing a coordinator at the first layer, dynamically selecting candidate tools and execution paths based on a Monte Carlo tree search mechanism by receiving natural language requests and uploaded files from the user; invoking the initialization process of the node management module based on the user's task request; scheduling a specialized agent collaborative reasoning layer at the third layer based on a structured task description and tool call sequence to achieve enzyme mutation prediction, kinetic analysis, and homologous sequence discovery; and after obtaining the task execution results, using a result analysis agent to scientifically interpret the results output by the specialized agent collaborative reasoning layer, generating result analysis content or visualizations for user display. This invention is beneficial for improving the automation level and overall processing efficiency of enzyme engineering research and development tasks.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and enzyme engineering technology, and more specifically to a multi-agent collaborative processing method for enzyme engineering based on Monte Carlo tree search. Background Technology

[0002] Enzymes, as important catalysts in biological systems, have broad application prospects in fields such as pharmaceutical manufacturing, food processing, bioremediation, and synthetic biology. Because industrial applications typically place higher demands on enzymes' catalytic activity, substrate selectivity, thermal stability, pH adaptability, and solvent tolerance, natural wild-type enzymes often cannot directly meet actual production needs. Therefore, enzyme engineering techniques are required to modify and optimize target enzymes.

[0003] In current enzyme engineering research, directed evolution remains a commonly used technical approach. This method obtains enzyme variants with improved performance by constructing mutant libraries and combining them with experimental screening. However, due to the vast space of protein mutation combinations, experimental screening often suffers from problems such as high workload, low screening efficiency, long development cycle, and high experimental costs. With the development of bioinformatics and artificial intelligence technologies, researchers are gradually adopting computational methods to assist enzyme engineering design, in order to improve the efficiency of candidate mutant screening and reduce experimental burden. Currently, common technical solutions include zero-sample mutation effect prediction methods based on protein language models and enzyme kinetic parameter prediction methods based on deep learning. The former uses probabilistic modeling of protein sequences to assess the potential impact of mutations on enzyme fitness, stability, or function by using the difference in sequence scores before and after mutation; the latter uses the fusion of protein features and substrate features to model enzyme-substrate interactions, in order to predict enzyme kinetic parameters such as catalytic efficiency and substrate affinity. These technologies have improved the prediction efficiency and preliminary analysis capabilities in enzyme engineering tasks to a certain extent.

[0004] However, existing technologies still have the following problems in practical applications: First, enzyme design and analysis involve multiple processing steps, typically including sequence processing, site screening, mutation prediction, kinetic analysis, database retrieval, data statistics, and visualization, which are lengthy and lack automation. Second, various analysis tools mostly rely on scripts and command lines, requiring researchers to write their own code and adapt and convert various data formats, making the overall operation process cumbersome and lengthy. Third, existing prediction tools mostly run as independent models, lacking a unified orchestration, collaborative scheduling, and feedback optimization mechanism oriented towards enzyme engineering task objectives, resulting in a disconnect between mutation prediction, functional retrieval, kinetic evaluation, and result analysis, making it difficult to form a closed-loop processing flow. Fourth, existing systems are not user-friendly in terms of human-computer interaction, failing to meet users' actual needs for task input, file upload, process viewing, and result understanding through a graphical interface.

[0005] Therefore, there is an urgent need to provide an intelligent collaborative processing method for enzyme engineering design, analysis, and discovery scenarios. This method should unify and collaboratively orchestrate various professional capabilities, such as enzyme mutation effect prediction, kinetic parameter evaluation, functional retrieval, code execution, database interaction, and result interpretation. It should also be able to adaptively plan execution paths based on user task objectives and intermediate processing results, thereby reducing the complexity of manual programming and switching between multiple tools, lowering the barrier to entry for the system, and ensuring task closure through an iterative code self-repair mechanism when facing underlying data anomalies such as non-standard bioinformatics file encoding and ambiguous column names. Furthermore, it should improve the human-computer interaction experience and task processing efficiency through a visual UI interface.

[0006] Therefore, proposing a multi-agent collaborative processing method for enzyme engineering based on Monte Carlo tree search to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a multi-agent collaborative processing method for enzyme engineering based on Monte Carlo tree search, to solve the problems of complex analysis processes, low automation, fragmented prediction and analysis tools, lack of a unified task orchestration mechanism, and high barriers to entry due to reliance on manual programming and command line operation in existing enzyme engineering technologies; at the same time, it solves the problems of unfriendly human-computer interaction in existing systems, and difficulty in supporting natural language input, file upload and result display through a visual UI interface, thereby improving the collaboration, ease of use and efficiency of enzyme engineering task processing.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A multi-agent cooperative processing method for enzyme engineering based on Monte Carlo tree search includes the following steps: S1. A coordinator located in the first layer is adopted, which dynamically selects candidate tools and execution paths based on the Monte Carlo tree search mechanism by receiving natural language requests and uploaded files from users. S2. Based on the user's task request, the coordinator calls the initialization process of the node management module, and activates the data agent and task agent in sequence, thereby forming the task context and tool priority information required for subsequent execution. S3. Based on structured task descriptions and tool call sequences, the professional intelligent agent collaborative reasoning layer located in the third layer is scheduled to execute professional sub-tasks including custom data analysis and visualization, enzyme mutation effect prediction, kinetic prediction parameters, and interaction with external biological databases. S4. After obtaining the task execution results, the result analysis agent scientifically interprets the results output by the professional agent collaborative reasoning layer and generates result analysis or visualization for user display.

[0009] Optionally, in S2, the data agent performs multimodal perception on uploaded non-standard bioinformatics files and user text, parses sequence metadata, extracts column name features and identifies enzyme-specific parameters, and generates a structured perception report containing enzyme-specific metadata including file summary, column names, extracted sequences, molecular weight, and EC number.

[0010] Optionally, in S2, the task agent maps user intent to six preset task classification systems, including mutation effect prediction, functional parameter prediction, sequence homology search and enzyme retrieval, database-driven enzyme discovery, general data analysis and visualization, and multi-step composite tasks.

[0011] Optionally, the professional agent collaborative reasoning layer in S3 includes: code subsystem, mutation agent, dynamic parameter agent, and functional agent; among which the code subsystem includes: code generation agent, code execution agent, and code analysis agent.

[0012] Optionally, in S3, a custom data analysis and visualization task is performed. An agent within the code execution subsystem generates a Python script by combining semantic memory and episodic memory. The code execution agent executes the script in a sandboxed Python subprocess and captures standard output and runtime error backtracking information. The code analysis agent performs error diagnosis and generates corrective code, and records all previous code attempts, error types, and repair strategies for the current task and other tasks in the same session.

[0013] Optionally, the mutant agent takes the wild-type amino acid sequence and the specified mutation as input, calculates the mutation effect score and normalizes it to obtain the final prediction. It supports full-site mode and specified-site mode, and has a built-in sub-function for predicting dynamic parameters.

[0014] Optionally, the kinetic parameter agent takes the amino acid sequence and the substrate SMILES string as input, and simultaneously predicts the maximum number of times k of a single enzyme molecule catalyzes the conversion of the substrate into the product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m And the overall catalytic capacity of an enzyme for a specific substrate (k) cat / K m It supports prediction mode and training / fine-tuning mode.

[0015] Optionally, the functional agent integrates BLASTp sequence homology search with REST API database retrieval to achieve enzyme retrieval, homology sequence discovery, structured database query, and functional information acquisition.

[0016] Optionally, the system receives natural language requests from users through a visual interface and supports uploading sequence files, table files, and related data files. The system combines user input, uploaded file parsing results, and session history information to identify the task intent and organize it into structured parameters for use by the backend module. Subsequently, the system provides feedback to the user on file loading status, task processing status, and execution results through the interface, and retains historical context through a session-level parameter saving mechanism, thereby supporting continuous, multi-round human-computer interaction analysis processes around the same enzyme engineering object.

[0017] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a multi-agent cooperative processing method for enzyme engineering based on Monte Carlo tree search, the beneficial effects of which are: 1) Enable end-to-end collaborative processing of complex enzyme engineering tasks, improving task automation and overall processing efficiency: This invention constructs a four-layer hierarchical coordination architecture for enzyme engineering scenarios. This architecture unifies multiple fragmented steps that previously required researchers to manually switch between different bioinformatics tools, prediction models, database resources, and data analysis environments into a collaborative processing flow encompassing user input, data extraction, task decomposition, specialized computation, result analysis, and feedback iteration. By introducing a Monte Carlo tree search-based task path planning mechanism, the system can adaptively orchestrate the order of multi-agent calls and execution paths around user goals, thereby reducing manual intervention and repetitive operations and improving the automation level and overall processing efficiency of enzyme engineering research and development tasks. 2) Reduce reliance on manual programming and switching between multiple tools, lowering the barrier to entry for system use: This invention addresses the challenges of complex operational procedures and frequent tool switching in existing enzyme engineering analysis processes by unifying code execution, model invocation, database interaction, and result interpretation into a single framework. Users can initiate tasks directly through a visual UI interface using natural language input and file uploads, thereby reducing the need for switching between multiple tools and manual processing, simplifying the analysis process, and improving the continuity and overall efficiency of task execution. 3) Improving the synergy between heterogeneous tools and specialized models is beneficial for forming a closed-loop processing flow oriented towards enzyme engineering goals: This invention, through a structured message passing mechanism and a multi-agent collaborative architecture, unifies and connects previously relatively independent processing capabilities such as enzyme mutation effect prediction, kinetic parameter evaluation, functional retrieval, data analysis, and result interpretation. This solves the problem in existing technologies where various tools are fragmented and difficult to continuously collaborate around specific task objectives. By providing feedback and replanning of intermediate results, this invention facilitates the formation of a closed-loop processing flow applicable to tasks such as enzyme design, enzyme analysis, and candidate enzyme discovery, thereby improving the continuous processing capability of complex enzyme engineering tasks. 4) Possesses good adaptability to anomalies and the ability to execute tasks continuously: In response to common issues in enzyme engineering data processing, such as differences in input formats, non-standard fields, missing data, or runtime errors, the code subsystem in this invention can diagnose problems by combining execution results and error information, and adjust subsequent processing steps. Compared with traditional solutions that rely on fixed script processes, this invention can maintain the continuous progress of tasks even when faced with non-standard input data and execution anomalies, thereby improving the system's adaptability to complex data environments and the stability of task execution. Attached Figure Description

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

[0019] Figure 1 A flowchart of an enzyme engineering multi-agent collaborative processing method based on Monte Carlo tree search provided by the present invention; Figure 2 This invention provides a system architecture diagram for multi-agent collaborative processing of enzyme engineering based on Monte Carlo tree search; Figure 3 This invention provides a schematic diagram of the interactive execution of multi-agent collaborative processing in enzyme engineering based on Monte Carlo tree search. Figure 4 This is a comparison diagram of the relative activity of wild-type SEAP enzyme and various point mutants provided by this invention. Detailed Implementation

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

[0021] See Figure 1 As shown, this invention discloses a multi-agent cooperative processing method for enzyme engineering based on Monte Carlo tree search, comprising the following steps: S1. A coordinator located in the first layer is adopted, which dynamically selects candidate tools and execution paths based on the Monte Carlo tree search mechanism by receiving natural language requests and uploaded files from users. S2. Based on the user's task request, the coordinator calls the initialization process of the node management module, and activates the data agent and task agent in sequence, thereby forming the task context and tool priority information required for subsequent execution. S3. Based on structured task descriptions and tool call sequences, the professional intelligent agent collaborative reasoning layer located in the third layer is scheduled to execute professional sub-tasks including custom data analysis and visualization, enzyme mutation effect prediction, kinetic prediction parameters, and interaction with external biological databases. S4. After obtaining the task execution results, the result analysis agent scientifically interprets the results output by the professional agent collaborative reasoning layer and generates result analysis content or visualizations for user display.

[0022] Specifically, S1 is executed by the coordinator located at the first layer, see [link to relevant documentation]. Figure 2 As shown, in this embodiment, the coordinator receives natural language requests and uploaded files from users through a Chainlit-based dialogue interface. It then combines the session history, existing result file paths, and previous round task parameters to perform unified modeling of the current request, forming a comprehensive decision input for the current round.

[0023] The comprehensive decision input is represented as follows:

[0024] in, Indicates the first Request the user's natural language. This indicates the uploaded file and its parsing results. Represents historical session information. This indicates that there is already a result path or intermediate result in the current session. This represents the structured parameters saved from the previous task.

[0025] Furthermore, to enhance path selection capabilities in multi-tool scenarios, a node management module is included in the coordinator. This module represents the current task status, candidate tool set, parameter status, intermediate result status, and subsequent expandable actions as nodes and branches in a search tree, and uses a Monte Carlo tree search mechanism to dynamically select candidate tools and execution paths.

[0026] Node selection is based on the following criteria:

[0027] in, Indicates the current task status node. This indicates the current task status. The following is a set of candidate tools. This refers to a candidate tool within the candidate tool set. This indicates the target tool selected through the Monte Carlo tree search mechanism. Indicates the current task status Download candidate tools The cumulative average return action obtained, Represents a node Number of visits, Candidate tool In the task status node The number of times it is selected This represents the exploration coefficient.

[0028] Cumulative average return This comprehensively reflects the degree of matching between actions and user intent, the completeness of input conditions, the probability of tool execution success, the completeness of results, and the ability to connect with subsequent tasks. Through these methods, the coordinator can dynamically select from multiple candidate paths, such as direct response, invoking a single tool, triggering multi-tool cascading, and further expanding analysis based on intermediate results. This improves the system's efficiency, robustness, and intelligence in handling complex enzyme engineering tasks.

[0029] Furthermore, in S2, the data agent performs multimodal perception on uploaded non-standard bioinformatics files and user text, parses sequence metadata, extracts column name features and identifies enzyme-specific parameters, and generates a structured perception report containing enzyme-specific metadata including file summary, column name, extracted sequence, molecular weight, and EC number.

[0030] Furthermore, in S2, the task agent maps user intent to six pre-defined task classification systems, including mutation effect prediction, functional parameter prediction, sequence homology search and enzyme retrieval, database-driven enzyme discovery, general data analysis and visualization, and multi-step composite tasks.

[0031] Specifically, based on the user's task request, the coordinator calls the initialization process of the node management module and activates the data agent and task agent located in the second layer in turn to form a unified task context, subtask dependencies and candidate tool priority information required for subsequent execution.

[0032] The coordinator first activates the data agent, which then calls upon the bioinformatics parsing library and table processing library to perform multimodal perception on the uploaded non-standard bioinformatics files and user text. This process identifies enzyme sequences, substrate samples, column name semantics, table row number mapping relationships, and enzyme-related metadata, generating a structured perception report. The structured perception report includes at least a file summary, column name information, sample row content, extracted sequences, samples information, row-level mapping relationships, and enzyme-related metadata. This structured perception report is injected as a global context into downstream agents and tool modules, allowing subsequent processing to directly reference the extracted content without repeatedly reading the source file.

[0033] After the data agent completes structured perception, the coordinator activates the task agent. Based on the structured perception report, user request content, historical session states, and existing intermediate results in the current session, the task agent categorizes the user intent into tasks, breaks down subtasks, and logically arranges them. The task agent maps user requests to a pre-defined task classification system, which includes at least mutation effect prediction tasks, enzyme kinetic parameter prediction tasks, sequence homology search and enzyme retrieval tasks, database-driven enzyme discovery tasks, general data analysis and visualization tasks, and multi-step composite tasks.

[0034] For multi-step complex tasks, the task agent breaks down the user request into multiple ordered subtasks and clarifies the input-output dependencies between each subtask to ensure that the subsequent execution layer calls the corresponding tools in the correct order.

[0035] Furthermore, the coordinator and node management module generate a set of candidate tools and their priority information based on task classification results, input condition completeness, uploaded file content, existing intermediate results, and dependencies between preceding and subsequent steps, thereby providing an initialization basis for subsequent tool scheduling and execution path search.

[0036] Furthermore, the professional agent collaborative reasoning layer in S3 includes: code subsystem, mutation agent, dynamic parameter agent, and functional agent; among which the code subsystem includes: code generation agent, code execution agent, and code analysis agent.

[0037] Furthermore, in S3, custom data analysis and visualization tasks are performed by agents within the code subsystem. The code generation agent combines semantic memory and episodic memory to generate Python scripts. The code execution agent executes the scripts in a sandboxed Python subprocess and captures standard output and runtime error backtracking information. The code analysis agent performs error diagnosis and generates corrective code, and records all previous code attempts, error types, and repair strategies for the current task and other tasks in the same session.

[0038] Furthermore, the mutant agent uses wild-type amino acid sequences and specified mutations as inputs to calculate mutation effect scores and normalize them to obtain the final prediction. It supports full-site mode and specified-site mode, and has a built-in sub-function for predicting dynamic parameters.

[0039] Furthermore, the kinetic parameter agent takes the amino acid sequence and the substrate SMILES string as input, and simultaneously predicts the maximum number of times k of a single enzyme molecule catalyzes the conversion of the substrate into the product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m And the overall catalytic capacity of an enzyme for a specific substrate (k) cat / Km It supports prediction mode and training / fine-tuning mode.

[0040] Furthermore, the functional agent integrates BLASTp sequence homology search with REST API database retrieval to achieve enzyme retrieval, homology sequence discovery, structured database query, and functional information acquisition.

[0041] Specifically, the S3 process is as follows: After S2 completes the structured task description, subtask dependencies, and candidate tool priority generation, S3 is the process by which the coordinator schedules the third-layer execution modules. The coordinator selects the appropriate dedicated execution path based on the task type, input condition completeness, and upstream / downstream dependencies to complete the computational analysis, model inference, database interaction, and result feedback for enzyme engineering tasks. The scheduling process is represented as follows:

[0042] in, This represents the set of executable modules available in the system. This indicates one of the execution modules, corresponding to the code subsystem, mutation agent, dynamic parameter agent, BLAST retrieval module, or database query module. Indicates execution module The precondition determination function is used to determine whether the current context and dependency conditions meet the execution requirements; Indicates execution module With structured task description The matching function is used to select and arrange code subsystems, mutant agents, dynamic parameter agents, homology search modules, and database query modules based on the current task objective, input data status, and previous output conditions.

[0043] When the structured task description indicates that the current task belongs to a custom data analysis, result processing, visualization, file format conversion, or other general bioinformatics computation task, the coordinator schedules the code subsystem located at the third layer. The code subsystem includes a code generation agent, a code execution agent, and a code analysis agent, which work collaboratively in a closed-loop manner. The code generation agent generates directly executable Python code based on the structured task description and jointly utilizes semantic memory and episodic memory to improve code generation quality. Semantic memory is a static, domain-specific code fragment knowledge base, covering typical calling methods, file read / write rules, and plotting and saving specifications for commonly used bioinformatics and data analysis libraries such as Biopython, pandas, matplotlib, and seaborn. The system retrieves reference fragments related to the current task from the knowledge base through keyword matching and injects prompts to enhance the adaptability of the generated code to enzymatic data formats and biological sequence processing scenarios. Episodic memory records previous code attempts, error types, failure reasons, and repair strategies for the current task and other code tasks within the current session. This information is injected as contextual prompts during subsequent code generation and repair processes to prevent the system from repeatedly using failed processing methods. The generated code prioritizes the absolute file paths provided by the perception report as input paths to improve portability and stability under different operating environments.

[0044] The code execution agent runs the generated code in an isolated Python subprocess, capturing standard output, output file paths, and runtime error backtracking information. Upon successful execution, it returns output results such as structured data tables, image files, result files, or numerical summaries. When execution fails, a non-zero exit status, an exception being thrown, or Python backtracking information in the output triggers the code analysis agent to diagnose the error and generate corrective code, which is then submitted to the code execution agent for execution again. The repair prompts explicitly include the current task's context and error pattern information from other code tasks in the same session, enabling cross-task error learning and preventing the system from repeatedly proposing the same repair strategy within the same session. If the repair loop fails within a preset maximum number of iterations, the system returns a structured error report and manual intervention suggestions to the user to prevent system silent suspension due to execution failure.

[0045] Preferably, the maximum number of repair iterations is configured via the environment variable MAX_FIX_ATTEMPTS, with a default value of 3. The design of this code subsystem is based on the fact that bioinformatics data files commonly suffer from inconsistent encoding, non-standard column names, missing fields, and heterogeneous formats. Single-round code generation is insufficient to reliably adapt to real-world application scenarios. Therefore, by separating code generation, execution, and analysis into independent but iteratively linked execution units, the completion rate and error convergence capability of complex custom analysis tasks are significantly improved.

[0046] When the structured task description indicates that the current task belongs to enzyme mutation effect prediction, the coordinator schedules the mutant agent to perform single-point mutation effect prediction on the input enzyme sequence using a protein language model. Given the wild-type amino acid sequence and candidate mutations, the mutant agent calls models such as EnzGFM and ESM2 to perform predictions, and preferably adopts the wt-marginals scoring strategy to obtain the mutation effect score. The single-point mutation effect scoring formula is expressed as:

[0047] in, Indicates the first Under a protein language model, targeting wild-type sequences and candidate mutations Mutation effect score; This indicates the prediction model number used; different models are not specified. It can be used with different models such as EnzGFM-650M, EnzGFM-150M, and ESM2-3B. Indicates the wild-type amino acid sequence; Indicates a candidate single-point mutation; Indicates the location where the mutation occurred; Indicates the wild-type sequence at position The original amino acid residues at the location; Indicates the position after mutation The target amino acid residue; Indicates except the first Sequence context outside of bits; Indicates the first Each model, given the context of the remaining sequence, determines the position of the mutated residue. The conditional probability; Indicates the first One model determines the location of wild-type residues. The conditional probability is calculated. After post-processing, the final prediction result is obtained, where a higher normalization score indicates that the mutation is more likely to produce a beneficial effect relative to wild-type residues.

[0048] The mutation agent supports at least two operating modes: a full-site scanning mode and a specified-site mode. In full-site scanning mode, it scans all possible single-point substitutions at each position of the input sequence, generating a complete mutation scoring map covering the entire sequence, and outputting it in sorted order according to the predicted scores. In specified-site mode, it only evaluates mutation sites explicitly specified by the user in the format of "original amino acid-position-target amino acid" to meet the user's site-specific validation needs. Both modes support the simultaneous running of multiple models and generate independent scoring columns for different models in the output result file, thereby supporting subsequent dual-model consensus extraction and result consistency analysis. In the implementation of this embodiment, the mutation prediction task is executed on a dedicated computing node by calling an external prediction script through a local subprocess. The model script path, model weight path, and default model type can all be configured through environment variables, keeping the user-interactive orchestration layer lightweight while delegating the high-overhead model inference task to the dedicated hardware environment. For multi-model running scenarios, the system can also automatically trigger the consensus extraction and result analysis process after prediction is completed to further improve the reliability of candidate mutation site screening.

[0049] When the structured task description indicates that the current task belongs to the kinetic parameter prediction task, or is a downstream chain analysis task following mutation prediction, the coordinator schedules the kinetic parameter agent. The kinetic parameter agent, based on the UniKP framework, uses the amino acid sequence and substrate SMILES string as joint inputs, and calculates the maximum number of times k a single enzyme molecule catalyzes the conversion of the substrate into the product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m And the overall catalytic capacity of an enzyme for a specific substrate (k) cat / K m It supports multiple protein-coding models, including the default ProtT5-XL model and models such as EnzGFM-650M and ESM2-3B. When the user has completed mutation prediction and the current session contains variant sequences, model parameters, or intermediate results, kinetic parameter prediction can be automatically triggered as a downstream chain function for mutation effect prediction. If the user has provided or the system has identified substrates (SMILES) from the file, the top-ranked mutant variants will further execute the maximum number of times (k) a single enzyme molecule catalyzes the conversion of the substrate to the product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m Predictions enable an end-to-end enzyme engineering analysis workflow, from mutation design to kinetic evaluation, within a single session, without requiring users to repeatedly upload data.

[0050] The kinetic parameter prediction supports both prediction and training / fine-tuning modes, allowing users to train, adjust, or adapt models for specific tasks based on CSV format experimental datasets. Execution is preferably performed using local asynchronous subprocess calls, offloading model inference tasks to a GPU server. Script paths, working directories, model types, and weight file paths are all configured via environment variables. The system can also backfill and display prediction results by combining the sequences, SMILES column, mutant column, and wild-type sequences from the original input table, allowing users to directly compare the differences in kinetic predictions between wild-type and mutant models.

[0051] When the structured task description indicates that the current task involves enzyme retrieval, homology sequence discovery, structured database querying, or functional information acquisition, the coordinator schedules the functional agent. The functional agent integrates two complementary types of external information acquisition capabilities: BLASTp-based sequence homology search capabilities and REST API-based structured database retrieval capabilities.

[0052] For sequence homology search tasks, the functional agent invokes BLASTp to perform searches on the protein database, supporting both single-query and batch-query modes. Input can be amino acid sequences, UniProt accession numbers, or batch sequences from uploaded files. When a user provides a UniProt accession number, the system automatically retrieves the corresponding complete amino acid sequence via the database interface before submitting the BLAST search, and then inputs it as the search sequence into the homology search module. Search parameters can be automatically parsed and configured according to user requests, including the minimum sequence consistency threshold, the maximum number of hits per query, and the E-value cutoff. In batch search scenarios, the system can limit the number of simultaneous searches through a concurrency control mechanism to reduce the instantaneous access pressure on external services.

[0053] For structured database retrieval tasks, the functional agent obtains structured annotation information of the target enzyme through REST interfaces provided by databases such as UniProt and PDB. This includes functional classification, catalytic residues, active sites, binding sites, domain information, protein sequence, kinetic parameters, known structures, and pre-calculated structural models. When the user only provides the enzyme name without providing the standard database number, the system can first perform candidate entry search and matching, and then return the corresponding structured results. Both the homology search results and database retrieval results mentioned above can be further passed to the downstream results analysis module, where the results analysis agent sorts, summarizes, scientifically interprets, and visualizes the results.

[0054] Furthermore, the system receives natural language requests from users through a visual interface and supports uploading sequence files, table files, and related data files. The system combines user input, uploaded file parsing results, and session history information to identify the task intent and organize it into structured parameters for use by backend modules. Subsequently, the system provides feedback to the user on file loading status, task processing status, and execution results through the interface, and retains historical context through a session-level parameter saving mechanism, thereby supporting continuous, multi-round human-computer interaction analysis processes around the same enzyme engineering object.

[0055] For details, see Figure 3 As shown, the specific process of S4 is as follows: S4 is executed by the result analysis agent located in the fourth layer. It is used to uniformly receive, structure, visualize and scientifically interpret the raw results output by the specialized execution module in the third layer, and then send the enriched results back to the coordinator to support subsequent iterative tasks.

[0056] The results analysis agent takes the output of the execution module, task description, current user input text, historical session status, and existing result paths in the current session as joint inputs to automatically profile the result file, identify the row and column structure, field type, missing proportion, numerical statistical features, and field semantics of the data table, and further determine whether it belongs to the mutation effect prediction result, the dual-model consensus result, the dynamic parameter prediction result, or the general analysis result.

[0057] For mutation effect prediction results, the results analysis agent ranks candidate mutations according to prediction scores, identifies high-confidence beneficial mutations, hotspot sites, and substitution amino acid preferences, and automatically generates ranking curves, site frequency maps, substitution preference maps, or other adaptation charts. For dual-model consensus results, the results analysis agent further generates a consistency comparison map for the two models to identify high-confidence candidate mutations that are supported by both models. For kinetic parameter prediction results, the results analysis agent determines the maximum number of times k of a single enzyme molecule catalyzes the conversion of substrate to product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m And the overall catalytic capacity of an enzyme for a specific substrate (k) cat / K mStructured summarization, sorting, and comparison display are performed to support comparative analysis between wild-type and mutant strains. Generated charts, result tables, and scientific explanations are saved to the session output path and returned to the user via the interactive interface. After completing a result analysis, the result analysis agent also writes back the sorting results, recommended candidates, hotspot sites, result file paths, and structured summaries to the current session context. The coordinator then continues subsequent routing and scheduling based on the enrichment context, enabling users to continuously complete an iterative enzyme engineering workflow of "mutation prediction—result screening—kinetic parameter evaluation—further analysis" within a single conversation without repeatedly uploading data or performing pre-calculations.

[0058] See Figure 4 As shown, Figure 4 This is a comparison of the relative activities of wild-type SEAP enzymes and different point mutants. The relative activities of the wild-type enzyme were normalized to 1. Figure 4 It was found that the relative activities of the selected point mutants were all higher than those of the wild type, with mutant 1 exhibiting the highest relative activity. These results demonstrate that the method of this invention can provide effective support for the screening of candidate mutation sites and the optimization of enzyme activity in enzyme engineering.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-agent cooperative processing method for enzyme engineering based on Monte Carlo tree search, characterized in that, Includes the following steps: S1. A coordinator located in the first layer is adopted, which dynamically selects candidate tools and execution paths based on the Monte Carlo tree search mechanism by receiving natural language requests and uploaded files from users. S2. Based on the user's task request, the coordinator calls the initialization process of the node management module, and activates the data agent and task agent in sequence, thereby forming the task context and tool priority information required for subsequent execution. S3. Based on structured task descriptions and tool call sequences, the professional intelligent agent collaborative reasoning layer located in the third layer is scheduled to execute professional sub-tasks including custom data analysis and visualization, enzyme mutation effect prediction, kinetic prediction parameters, and interaction with external biological databases. S4. After obtaining the task execution results, the result analysis agent scientifically interprets the results output by the professional agent collaborative reasoning layer and generates result analysis or visualization for user display.

2. The enzyme engineering multi-agent collaborative processing method based on Monte Carlo tree search according to claim 1, characterized in that, In S2, the data agent performs multimodal perception on uploaded non-standard bioinformatics files and user text, parses sequence metadata, extracts column name features and identifies enzyme-specific parameters, and generates a structured perception report containing enzyme-specific metadata such as file summary, column name, extracted sequence, molecular weight, and EC number.

3. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 1, characterized in that, In S2, the task agent maps user intent to six pre-defined task classification systems, including mutation effect prediction, functional parameter prediction, sequence homology search and enzyme retrieval, database-driven enzyme discovery, general data analysis and visualization, and multi-step composite tasks.

4. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 1, characterized in that, The collaborative reasoning layer of specialized intelligent agents in S3 includes: code subsystem, mutation agent, dynamic parameter agent, and functional agent; among which the code subsystem includes: code generation agent, code execution agent, and code analysis agent.

5. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 4, characterized in that, In S3, custom data analysis and visualization tasks are performed by agents within the code execution subsystem. The code generation agent combines semantic memory and episodic memory to generate Python scripts. The code execution agent executes the scripts in a sandboxed Python subprocess and captures standard output and runtime error backtracking information. The code analysis agent performs error diagnosis and generates corrective code, and records all previous code attempts, error types, and repair strategies for the current task and other tasks in the same session.

6. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 4, characterized in that, The mutant agent takes wild-type amino acid sequences and specified mutations as inputs, calculates mutation effect scores, and normalizes them to obtain the final prediction. It supports full-site mode and specified-site mode, and has a built-in sub-function for predicting dynamic parameters.

7. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 4, characterized in that, The kinetic parameter agent takes the amino acid sequence and the substrate SMILES string as input and predicts the maximum number of times k of a single enzyme molecule catalyzes the conversion of the substrate into the product per unit time. cat The substrate concentration K at which the enzyme reaches half of its maximum catalytic rate m And the overall catalytic capacity of an enzyme for a specific substrate (k) cat / K m It supports prediction mode and training / fine-tuning mode.

8. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 4, characterized in that, The functional agent integrates BLASTp sequence homology search with REST API database retrieval to achieve enzyme retrieval, homology sequence discovery, structured database query, and functional information acquisition.

9. The enzyme engineering multi-agent cooperative processing method based on Monte Carlo tree search according to claim 1, characterized in that, The system receives natural language requests from users through a visual interface and supports uploading sequence files, table files, and related data files. It combines user input, uploaded file parsing results, and session history information to identify the task intent and organize it into structured parameters for use by backend modules. The system then provides feedback to the user on file loading status, task processing status, and execution results through the interface. It also retains historical context through a session-level parameter saving mechanism, thereby supporting continuous, multi-round human-computer interaction analysis processes around the same enzyme engineering object.

Images