Metabolic pathway node priority evaluation and cross-species pathway intelligent exploration method

Abstract

This invention discloses a method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways. The method comprises three stages: Stage α utilizes a large language model to rapidly assess the eight-dimensional modification potential of each molecule at each node in the metabolic pathway, and combines this with a small database query to quickly generate a node priority ranking; Stage β only calls a cross-species full-pathway retrieval tool for high-priority nodes confirmed by the user, returning cross-species alternative pathway data in complete pathway form, and incorporates a four-factor weighted ranking and truncation algorithm based on literature evidence, number of steps, enzyme characterization completeness, and thermodynamic feasibility; Stage γ (integration reasoning) uses the large language model to synthesize pathway data and output a structured strategy suggestion table. Testing showed that this method reduces database calls by 78.4% and contextual term consumption by 82.1% compared to a full-node, full-species indiscriminate retrieval scheme, while achieving an expert adoption rate of 76.5% for strategy suggestions, higher than the 54.2% of the unstratified control group.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of bioinformatics and artificial intelligence, and in particular to methods for prioritizing metabolic pathway nodes and intelligently exploring pathways across species. Background Technology

[0002] In metabolic engineering, when selecting heterologous pathways or modifying endogenous pathways, researchers need to answer a core question: "Which node in the pathway is most worth modifying, and what are the options for modification methods?" This question involves multiple aspects such as cross-species pathway comparison, enzyme source screening, and thermodynamic feasibility assessment.

[0003] The main limitations of existing technology are as follows: (1) Retrosynthesis tools perform indiscriminate searches on all nodes, resulting in wasted computational resources and excessive noise in the results. Existing retrosynthesis planning tools (such as RetroPath2.0, MINE Database, and novoStoic) search backward from the target product to the host endogenous metabolites, performing a full-species enzyme search at each step of the pathway. For a pathway containing 8 intermediate nodes, if each node is associated with an average of 15 enzymes from different species in the KEGG database, a complete search will generate more than 120 enzyme information records. In the application scenario of large language model agents, injecting all this data into the context window will quickly reach the word limit (current mainstream models have 128K to 200K words), leading to a decrease in the quality of subsequent inference. Experimental measurements show that when more than 60,000 words of pathway data are injected into the context, the accuracy of the large language model in referencing key parameters drops from 94.3% to 61.7%.

[0004] (2) Existing tools return isolated response fragments, lacking pathway-level context. The interfaces of the KEGG and MetaCyc databases both return single responses as the unit of return, requiring researchers to manually string together multiple responses into a complete pathway. This is particularly disadvantageous for large language model agents: without complete pathway information, large language models cannot determine key issues such as "how many genes are needed to introduce this alternative pathway, how much net cofactor is consumed, and whether there are toxic intermediates".

[0005] (3) Lack of a hierarchical decision-making model of "strategy first, tactics later". In the current bioinformatics toolchain, data retrieval and strategy reasoning are separated - the tool returns all data, and researchers have to judge which data is valuable. Under the framework of large language model agents, this model results in a large amount of invalid data occupying the context space, squeezing the space for large language models to perform high-quality reasoning.

[0006] (4) Cross-species pathway data lacks a structured output format for large language models. The return formats of existing database interfaces (KGML format for KEGG, BioPAX format for MetaCyc, etc.) are machine-readable but not friendly to large language models, requiring extensive preprocessing to effectively inject context. Summary of the Invention

[0007] Based on the technical problems existing in the background art, this invention proposes a method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways.

[0008] This invention proposes a method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways based on hierarchical reasoning using a large language model, comprising the following three stages: Phase α: Node Strategy Analysis (Combination of Lightweight Scanning and Large Language Model Reasoning) Given a target metabolic pathway P, which consists of n metabolite nodes M1 to M2... The components are connected sequentially (where M1 is the host's endogenous metabolite, M...). (for the target product), for each intermediate node M (i ranges from 2 to n-1), the system performs the following eight-dimensional transformation potential assessment: Dimension 1 – Branching Degree Assessment: Utilize the competitive response retrieval tool to obtain information from node M. Departure or by M The number of all known metabolic reactions of the substrate is denoted as B(M). The branching score S1 equals B(M). Divide the result by 5 and take the smaller value between 1.0 and 5, where 5 is a normalization constant. A higher branching degree indicates a greater risk of flux leakage at that node and thus a higher value for modification. This step consumes one database query per node, returning approximately 200 to 500 tokens.

[0009] Dimension 2 – Thermodynamic Bottleneck Assessment: Utilizing thermodynamic assessment tools, the standard Gibbs free energy change ΔrG'° for each step of the reaction pathway is obtained in one go. For the reaction from M... To M +1 The reaction step is marked as thermodynamically unfavorable if ΔrG'° is greater than 0 kJ / mol. The score S2 maps the value of ΔrG'° divided by 10 to between 0 and 1 using the sigmoid function, mapping positive values ​​(thermodynamically unfavorable) to higher scores. This call is shared once for the entire pathway and returns approximately 300 tokens.

[0010] Dimension 3 – Enzyme Activity Bottleneck Assessment: Utilizing enzyme kinetic parameter query tools to obtain catalytic M... The catalytic rate constant kcat value of the transformed enzyme in the host. The score S3 is equal to the smaller of 1 minus kcat divided by 100 and 1.0. The lower the kcat, the higher the score, indicating greater value in modifying the enzyme by replacing it with a more efficient enzyme source. Each node consumes one query, approximately 150 terms.

[0011] Dimension 4 – Cofactor Consumption Assessment: This dimension uses a large language model to infer the types of cofactors consumed in this reaction step based on the reaction equation. If scarce cofactors are consumed (e.g., NADPH instead of NADH, or high-energy compounds like ATP), the S4 score will increase accordingly. This dimension requires no tool usage and is purely inference-based.

[0012] Dimension 5 – Precursor Availability Assessment: Utilize flux balance analysis tools to predict M The maximum endogenous supply flux F(M) in the host The score S5 equals 1 minus F(M). Divide by 10 and take the smaller value of 1.0. The lower the throughput, the more insufficient the endogenous supply of the precursor and the higher the demand for modification. Each node consumes 1 query, approximately 200 tokens.

[0013] Dimension 6 – Toxicity and Inhibition Assessment: Using a toxicity threshold query tool to obtain M The half-maximal inhibitory concentration (IC50) or growth-inhibiting concentration for the host. Simultaneously, the allosteric regulation query tool is invoked to examine catalytic M. Does the transformed enzyme exhibit product feedback inhibition? If known toxicity or feedback inhibition is present, the score S6 will increase accordingly. Each node consumes a maximum of 2 queries, approximately 300 terms.

[0014] Dimension 7 – Potential for Natural Alternative Pathways: Using large language models to infer whether other species can bypass M based on their knowledge of metabolic networks. Alternative metabolic pathways. For example, inferring whether there are shorter bypass pathways or alternative routes that do not pass through this intermediate. This dimension does not consume tool calls and is a pure inference dimension.

[0015] Dimension 8 – Existing Engineering Cases: Extract the number and effectiveness of successful modification cases at this node from the literature scan results of Phase 1. The presence of successful cases indicates high feasibility for modification. This dimension does not require additional tool calls.

[0016] For each node, a comprehensive transformation potential score is calculated. The comprehensive score is a weighted sum of the scores from eight dimensions, with a default weight vector of (0.18, 0.14, 0.16, 0.08, 0.14, 0.10, 0.08, 0.12). This weight was determined through a Delphi method survey of 12 metabolic engineering experts and further optimized by backtracking and validation using 30 historical cases.

[0017] The large language model sorts all nodes in descending order of their overall scores and outputs a node analysis report. The report is presented in tabular form and includes the following: node name, branching score and number of competing reactions, thermodynamic data (ΔrG'° value and whether it is unfavorable), current enzyme kcat value and score, cofactor consumption type, endogenous flux value and score, toxicity data, alternative pathway potential (judged by the large language model), number of literature cases, overall score, and suggested directions for further exploration. It then proceeds to a lightweight checkpoint (mini-CP) to request user confirmation of which nodes should be explored in depth.

[0018] Quantitative analysis of resource consumption in the α-stage: For a pathway with n intermediate nodes, the total number of tool calls in the α-stage does not exceed 4n+1 (one of which is a shared thermodynamic assessment across the entire pathway, and four calls for each node: competitive reaction, enzyme parameters, throughput analysis, and toxicity query). Taking a typical 6-node pathway as an example, there are 25 calls and a total data volume of approximately 8,000 terms. In comparison, the number of calls for a full-node, full-species deep search is approximately 270 (6 nodes multiplied by 15 species multiplied by 3 databases), and the total data volume is approximately 180,000 terms. The α-stage achieves a 10.8-fold reduction in call volume and a 22.5-fold reduction in data volume.

[0019] Phase β: Cross-species deep pathway exploration (data-intensive, high-priority nodes only) After confirming lightweight checkpoints (typically selecting the top 2 to 3 nodes based on overall score), users invoke the core advanced tool "Cross-Species Pathway Explorer" for each selected node. This tool accepts the following input parameters: target node metabolite (name or KEGG identifier), search direction (starting from the metabolite, reaching the metabolite, or bidirectional), maximum exploration depth (default 5 steps), species range (specifying a list of species or all species), and whether to include non-natural engineered pathways reported in the literature (default includes).

[0020] The tool internally executes the following four sub-steps in sequence: Sub-step one, KEGG cross-species pathway retrieval: Find all metabolic pathway maps containing this metabolite (covering all annotated species in the KEGG database), and for each pathway, trace the complete reaction path from (or to) this metabolite up to the specified depth step. Each reaction step is accompanied by the enzyme classification number EC and the complete reaction equation.

[0021] Sub-step two, MetaCyc complementary search: using the metabolite as a substrate or product to search for reactions, tracing downstream or upstream to the complete pathway. The MetaCyc database contains non-model organism pathways missing from KEGG, and the two data sources are complementary.

[0022] Sub-step three: Automatic enzyme parameter association: For each reaction's EC number, the system automatically queries the Brenda database for the corresponding enzyme's catalytic rate constant kcat, Michaelis constant Km, optimum pH, and optimum temperature. For enzymes lacking parameters in Brenda, the DLKcat deep learning model is used to predict and label the parameters with confidence levels.

[0023] Sub-step four, supplementing with engineered pathways from literature: If the user has enabled the option to include non-natural pathways, additionally search for artificial bypass pathways constructed using synthetic biology techniques reported in the literature.

[0024] After all sub-steps are completed, the system performs intelligent sorting and truncation operations on all collected pathways. The sorting is calculated based on the following four weighted factors to determine the comprehensive pathway score: literature evidence weight of 5 points (whether there are successful implementation cases of the pathway in engineered bacteria, if so, add 5 points), enzyme characterization integrity weight of 4 points (the proportion of all enzymes in the pathway with kinetic characterization data multiplied by 4), step economy weight of 3 points (the reciprocal of the number of steps normalized and multiplied by 3, with fewer steps given priority), and thermodynamic feasibility weight of 3 points (the proportion of all steps with negative ΔrG'° multiplied by 3).

[0025] The algorithm sorts and returns the top K pathways with the highest scores (K equals 10 by default). Each pathway's information includes: pathway identifier and name, source species, complete reaction pathway (response equations, enzyme information, kinetic parameters, and cofactor requirements listed step-by-step), net reaction equation, net cofactor balance (net consumption or generation of each cofactor), total number of steps, a Boolean marker indicating whether all enzymes have been characterized, and a Digital Object Identifier (DOI) for literature evidence. Summary information includes: total number of pathways retrieved, number of pathways returned, list of species covered, number of steps in the shortest path, and identifier of the highest-scoring pathway.

[0026] Phase γ: Integration analysis of large language models (pure inference with limited validation) After receiving the cross-species pathway data returned from the β stage, the large language model performs the following inference task: First, pathway feasibility assessment. For each candidate pathway, the feasibility of introducing it into the host is evaluated, including: the number of exogenous genes to be introduced, the presence of enzymes that have not yet been kinetically characterized (marked as risk factors), whether the cofactor requirements are compatible with the host's existing metabolic network, and whether the intermediate metabolites are toxic to the host.

[0027] Second, strategy comparison and combination. By comparing pathway schemes from different species, we can identify combinable "chimeric pathways"—that is, combining the optimal enzyme of one step of the reaction in species A with the optimal enzyme of another step of the reaction in species B.

[0028] Third, portability assessment. For top-ranked candidate pathways, pathway portability assessment tools can be used for detailed evaluation. Assessment indicators include codon preference fit, presence of signal peptides that need to be removed, and protein folding compatibility in the host cytoplasmic environment.

[0029] Fourth, output a strategy recommendation table. Detailed information for each introduction strategy is output in a structured table format. The table columns include: strategy number and description, name of the introduced gene and its source species, expected effects (quantified description), a list of risk factors, compatibility assessment with other modification strategies, and a comprehensive feasibility score (indicated by stars).

[0030] The proposed method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways has the following beneficial effects: (1) The number of database calls is reduced by 78.4%. In the pathway exploration tasks of 8 test cases (covering different target compounds and host combinations such as naringenin, mucoconic acid, lactic acid, and β-carotene), the average total number of calls in the three stages of this method was 47.3 times, while the average number of calls in the control group (non-discriminatory retrieval of all nodes and all species) was 218.6 times.

[0031] (2) Contextual lexical consumption was reduced by 82.1%. The total amount of contextual data injected into the large language model by this method averaged 14,200 lexical units (approximately 8,000 lexical units in the α stage plus approximately 6,200 lexical units in the top ten pathways of the β stage), while the control group averaged 79,400 lexical units. The contextual compression ratio was 5.6 to 1.

[0032] (3) The reasoning quality of the large language model is significantly improved. Under low context load conditions (this method), the large language model's accuracy in referencing key parameters (kcat, ΔrG'°, cofactor requirements) is 93.7%, while the control group's accuracy drops to 61.7% under high context load conditions. The internal logical consistency of the strategy recommendations (assessed by blind expert review on a scale of 1 to 5) is 4.2 ± 0.5 points for this method and 3.1 ± 0.8 points for the control group.

[0033] (4) The expert adoption rate increased by 22.3 percentage points. Twelve experts in the field of metabolic engineering conducted a blind review and evaluation of the strategy recommendations of eight cases each from the method and the control group. Among the strategy recommendations of the method, 76.5% were rated by the experts as "directly adoptable" or "adoptable with slight modifications", while the figure for the control group was 54.2%.

[0034] (5) Discovery of high-value alternative pathways. In the case of naringenin / Saccharomyces cerevisiae, the exploration of the malonyl-CoA node at the β stage revealed the MatB bypass from Rhodopseudomonas palustris (a single-step reaction that replaces the two-step reaction of acetyl-CoA carboxylase ACC, increasing the catalytic rate by 11.2 times). This approach had previously only been reported in two papers for fatty acid synthesis and had not yet been applied to the malonyl-CoA supply enhancement of the flavonoid pathway. In the case of 3-hydroxypropionic acid / Escherichia coli, a 3-HP / 4-HB cycle fragment from Vibrio chloroprene was discovered, which can serve as an alternative carbon dioxide fixation module. Attached Figure Description

[0035] Figure 1 The flowchart shows the metabolic pathway node priority assessment and cross-species pathway intelligent exploration method proposed in this invention. Figure 2 This is a flowchart of the three-stage hierarchical pathway exploration method proposed in this invention, which is a method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways. Figure 3 This diagram illustrates the internal execution logic and intelligent truncation algorithm of the cross-species pathway explorer in the proposed metabolic pathway node priority assessment and cross-species pathway intelligent exploration method. Detailed Implementation

[0036] Reference Figure 1-3 The production of naringenin using Saccharomyces cerevisiae is used as an example. The target pathway is: L-phenylalanine → cinnamic acid → p-coumaric acid → p-coumaroyl-CoA → naringenin chalcone → naringenin. The synthesis of naringenin chalcone is catalyzed by chalcone synthase (CHS) and requires 3 molecules of malonyl-CoA as a substrate.

[0037] Implementation Step 1 (A1 Phase): Node Strategy Analysis.

[0038] The pathway has five intermediate nodes (cinnamic acid, p-coumaric acid, p-coumaryl-CoA, naringenin chalcone, and malonyl-CoA as a key substrate supply node), and the system performs an eight-dimensional evaluation of each node.

[0039] The system sequentially performed the following database queries: Querying the competing reactions at the pyruvate node returned 5 results, with the reaction catalyzed by pyruvate decarboxylase being the main flux leakage pathway, achieving a branching score of 1.0 (full marks); querying the competing reactions at the malonyl-CoA node returned 2 results, with fatty acid synthesis catalyzed by fatty acid synthase FAS1 being the main consumption pathway; a thermodynamic evaluation of the entire pathway revealed that the standard Gibbs free energy change for step 4 (the CHS-catalyzed reaction, where 3 molecules of malonyl-CoA condense with p-coumaroyl-CoA to form naringenin chalcone) was positive 3.1 kJ / mol, indicating a slightly unfavorable thermodynamic outcome; a query for the catalytic rate constant of parsley-derived chalcone synthase showed a kcat of only 0.02 s⁻¹. -1 The flux of malonyl-CoA in Saccharomyces cerevisiae is extremely low; the maximum endogenous supply flux of malonyl-CoA in Saccharomyces cerevisiae is predicted to be 0.21 mmol / gDW / h, which is extremely low; according to the toxicity of naringenin chalcone to Saccharomyces cerevisiae, the IC50 is greater than 50 mM, indicating low toxicity.

[0040] The α phase involved 18 database queries, adding approximately 6,800 words to the context.

[0041] The priority sorting table of the output nodes of the large language model is as follows: The node ranked first is malonyl-CoA, with a comprehensive score of 0.87. The key bottleneck is the extremely low supply (only 0.21 mmol / g DW / h) combined with low acetyl-CoA carboxylase (ACC) activity and competitive consumption by fatty acid synthase. The recommended directions for exploration are alternative supply pathways, ACC enzyme source replacement, and FAS downregulation.

[0042] The second-ranked node is naringenin chalcone (i.e., the CHS reaction site), with a composite score of 0.79. The key bottleneck is the extremely low CHS catalytic rate (kcat only 0.02 s). -1 Given the slightly unfavorable thermodynamics of this step, it is recommended to explore the screening of highly active CHS isoenzymes and thermodynamically driven strategies.

[0043] The third-ranked node is p-coumaroyl-CoA, with a comprehensive score of 0.41. It has moderate branching and a general 4CL enzyme activity level, and its modification priority is moderate.

[0044] Paracoumaric acid and cinnamic acid ranked 4th and 5th respectively, with comprehensive scores of 0.28 and 0.22, indicating low modification value.

[0045] The system has entered a lightweight checkpoint. It is recommended to prioritize in-depth cross-species exploration of the top two ranked nodes. The user confirms and agrees.

[0046] Step 2 (β phase): Exploration of deep pathways across species.

[0047] For the malonyl-CoA node, the system invoked the cross-species pathway explorer, setting the search direction to "reach this metabolite," the maximum exploration depth to 3 steps, the entire species scope, and including engineered pathways. The explorer returned 23 pathways, and after a four-factor weighted ranking, the top three results are as follows: The number one pathway is the malonic acid bypass pathway from *Rhodopseudomonas palustris*, which requires only one step: malonic acid reacts with coenzyme A under ATP-driven catalysis by malonyl-CoA synthase MatB to produce malonyl-CoA, with MatB having a kcat of 45.2 s. -1 The overall score for the pathway was 12.8.

[0048] The second-ranked pathway involves replacing the native ACC in Saccharomyces cerevisiae with a fortified acetyl-CoA carboxylase McACC from Mucor circinelloides. This requires a two-step reaction, and McACC has a kcat of 8.7 s. -1 (4.3 times that of wild-type ACC), with a pathway comprehensive score of 11.2.

[0049] The third-ranked pathway is the 3-HP cycle fragment from Chloroflexus aurantiacus, which requires three steps and involves multiple enzymes such as malonyl-CoA reductase (MCR). The pathway has a comprehensive score of 9.4.

[0050] For the CHS reaction node, the system invoked a cross-species enzyme comparison tool to compare the kinetic parameters of chalcone synthase (EC: 2.3.1.74) from five plant sources. The results showed that the kcat of HaCHS from *Hypericum androsaemum* was 0.15 s. -1 The Km for malonyl-CoA is 18 μM, and the catalytic efficiency (kcat / Km) is 8333, which is higher than that of currently used parsley-derived PcCHS (kcat is 0.02 s). -1 The catalytic efficiency of MsCHS from alfalfa (Medicagosativa) was 7.5 times that of parsley (Medicagosativa) at 35 μM, and 5.5 times that of PhCHS from petunia (Petunia hybrida).

[0051] The β phase involved 6 high-level tool calls and 4 intermediate-level supplementary queries, adding approximately 5,800 terms to the context.

[0052] Step 3 (γ phase): Large language model integration analysis.

[0053] The large language model integrates the beta-stage data and outputs a strategy suggestion table: Strategy A involves introducing Rhodopseudomonas palustris MatB and supplementing with exogenous malonic acid substrate, which is expected to increase the supply of malonyl-CoA by 5 to 8 times. The risks are that malonic acid needs to be added as a component of the culture medium and needs to be coordinated with the carbon source strategy of dimension D6. The overall feasibility is rated as five stars (the highest).

[0054] Strategy B involves replacing the native ACC of Saccharomyces cerevisiae with Mucor truncatula McACC, which is expected to increase the catalytic activity of ACC by 4.3 times. The risks are that the McACC protein has a large molecular weight (2350 amino acids), a high expression burden, and needs to be coordinated with the ATP cofactor supply of dimension D3. The overall feasibility is rated as four stars.

[0055] Strategy C involves replacing parsley PcCHS with St. John's wort HaCHS, which is expected to increase the catalytic rate by 7.5 times. The risk is that it is necessary to verify whether the substrate specificity is compatible with naringenin synthesis. This strategy is independent of other dimensions and has extremely high compatibility. Its overall feasibility is rated as five stars.

[0056] Strategy D is a combination of strategy A and strategy C, and is expected to produce a synergistic effect (simultaneously solving the two bottlenecks of insufficient substrate supply and low catalytic efficiency). The risk is that the protein expression burden will increase by two heterologous genes, which needs to be verified by enzyme constraint model. The overall feasibility is rated as four stars.

[0057] Strategy E is a combination of strategy B and strategy C, which is expected to produce synergy but has a higher ATP burden and a higher protein burden than strategy D. Its overall feasibility is rated as three stars.

[0058] The γ phase involves no additional tool calls and is purely for inference. The total context consumption for the entire three-phase process is approximately 14,600 words: approximately 6,800 words for the α phase, approximately 5,800 words for the β phase, and approximately 2,000 words for the γ phase inference.

[0059] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways, characterized in that, Includes the following steps: (1) α phase: Node strategy analysis phase Targeting each intermediate node M in the target metabolic pathway P (i=2 to n-1), perform an eight-dimensional transformation potential assessment for each node, including: Branching degree assessment: Obtain the number of competitive responses B (M) related to the node. ), calculate the branching score S1; Thermodynamic bottleneck assessment: Obtain the standard Gibbs free energy change ΔrG'° of the reaction and calculate the score S2; Enzyme activity bottleneck assessment: Obtain the catalytic rate constant kcat of the catalytic enzyme and calculate the score S3; Cofactor consumption assessment: Based on the reaction equation, infer the cofactor consumption situation and calculate the score S4; Precursor availability assessment: Obtain the intra-node source flux F(M) ), calculate the score S5; Toxicity and inhibition assessment: Obtain toxicity parameters and feedback inhibition information, and calculate the score S6; Assessment of the potential of natural alternative pathways: Inferring the potential of alternative pathways based on metabolic networks and calculating the score S7; Evaluation of existing engineering cases: Score S8 was calculated based on literature data; The scores of the above eight dimensions are weighted and summed according to preset weights to obtain the comprehensive transformation potential score of each node, and the nodes are sorted to output the node priority analysis results. Lightweight checkpoints are set based on the sorting results, allowing users to select high-priority nodes to proceed to the next stage. (2) β stage: Cross-species deep pathway exploration stage For the high-priority node selected by the user, invoke the cross-species pathway explorer and perform the following steps: Cross-species pathway retrieval based on metabolic databases to obtain complete response pathways containing target nodes; Supplement the pathway search to obtain information on non-model organisms or supplementary pathways; The enzyme kinetic parameters associated with each reaction in the pathway; Search the literature for engineered non-natural pathways; The obtained pathway set was weighted, scored, and ranked based on the following four factors: Documentary evidence; Completeness of enzyme characterization; Step-by-step economy; Thermodynamic feasibility; Output the sorted top K candidate pathways and their complete pathway information; (3) γ stage: Integrative reasoning stage Based on the pathway data output from the β stage, the large language model performs the following analysis: feasibility assessment of the pathway; Comparative and combined analysis of multi-pathway strategies; Pathway portability assessment; Output a structured strategy suggestion table.

2. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: In the eight-dimensional assessment, the branching score S1 is determined by the number of competing responses B (M). The result is obtained by dividing by the normalization constant and taking the smaller value of 1.

3. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: In the thermodynamic bottleneck assessment, the score S2 is obtained by mapping ΔrG'° to the 0 to 1 interval using the sigmoid function.

4. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: In the enzyme activity bottleneck assessment, the score S3 is obtained by subtracting the normalized value of kcat from 1.

5. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: In the precursor availability assessment, the score S5 is calculated by subtracting the endogenous flux F(M) from 1. The normalized value is obtained.

6. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: The comprehensive renovation potential score is calculated using the following weight vector: (0.18, 0.14, 0.16, 0.08, 0.14, 0.10, 0.08, 0.12）。 7. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: The cross-species pathway explorer includes the following sub-modules: pathway retrieval module; pathway supplement retrieval module; enzyme parameter association module; and literature pathway supplementation module.

8. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: The pathway ranking was performed using a four-factor weighted scoring method, with the following weights for each factor: literature evidence weighted at 5; enzyme characterization integrity weighted at 4; step economy weighted at 3; and thermodynamic feasibility weighted at 3.

9. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: The candidate pathway information returned includes: pathway identifier and name; source species; complete reaction pathway; enzyme information and kinetic parameters; net reaction equation; cofactor balance; number of steps; enzyme characterization integrity markers; and literature evidence markers.

10. The method for prioritizing metabolic pathway nodes and intelligently exploring cross-species pathways according to claim 1, characterized in that: The strategy recommendation table includes the following fields: strategy description; introduced genes and source species; expected effects; risk factors; strategy compatibility; comprehensive feasibility score; the total number of database calls in the α phase shall not exceed 4n+1 times, where n is the number of intermediate nodes.

Images