Metabolic regulation method based on genome-scale metabolic network model after gap filling

Abstract

The application provides a metabolic regulation method based on a genome-scale metabolic network model after gap filling, comprising obtaining metabolite and known reaction data of a GEM to be filled, constructing a stoichiometric matrix and a binary correlation matrix; based on the stoichiometric matrix and the binary correlation matrix, obtaining a structural prior representation of the metabolite; inputting the structural prior representation into a Transformer encoder to output a metabolite refined representation containing context correlation; performing directional pooling aggregation on the metabolite refined representation matrix based on the composition of substrates and products in the candidate reaction to obtain a gap filling prediction score of the candidate reaction; selecting a plurality of candidate reactions according to the gap filling prediction score to perform gap filling processing on the GEM to be filled to obtain a GEM after gap filling; and generating a candidate regulation target list by using the GEM after gap filling. The application can improve the accuracy of genome-scale metabolic network model analysis and prediction, and further realize accurate metabolic regulation of target microorganisms.

Description

Technical Field

[0001] This application relates to the fields of bioinformatics, computational biology, and metabolic engineering, and specifically to a metabolic regulation method based on a gap-filled genome-scale metabolic network model. Background Technology

[0002] Genome-scale metabolic network models (GEMs) describe intracellular metabolic processes through genome annotation, biochemical reaction knowledge, and stoichiometric constraints, and are widely used in flux balance analysis, metabolic engineering design, microbial fermentation capacity prediction, and systems biology research. The predictive power of GEMs is highly dependent on the completeness of the reaction network. For newly sequenced, non-pattern, or insufficiently annotated microorganisms, automatically reconstructed draft GEMs often lack key reactions, leading to flux interruptions, failure to generate target metabolites, or distorted phenotypic predictions.

[0003] Existing gap-filling methods typically rely on pre-defined biomass targets, culture medium hypotheses, phenotypic measurements, or expert manual review. In recent years, metabolic network gap-filling methods based on graph neural networks or hypergraph learning have begun to transform this problem into candidate response ranking or hyperedge prediction, reducing reliance on experimental data. However, many graph or hypergraph neural network methods still suffer from two shortcomings: first, local message passing can easily lead to representation smoothing near highly connected metabolites, weakening the functional differences between money metabolites and pathway-specific metabolites; second, many methods treat reactions as disordered sets of metabolites, ignoring the opposing roles of substrates and products in the reaction, resulting in the loss of directional information about the reaction.

[0004] A literature search of existing technologies revealed Chinese patent CN116665764A, which proposes a method for predicting missing reactions in metabolic networks. Based on the topology of the metabolic network, this method utilizes metabolic reaction information and biological network data from publicly available databases to predict missing reactions, providing a more general, accurate, and scalable solution for predicting and completing missing reactions in metabolic networks. However, this method does not consider the opposing roles of substrates and products in the reaction, which can easily lead to prediction distortion.

[0005] Therefore, there is a need for an automated metabolic network gap-filling method and metabolic regulation method that can utilize the global context of metabolic networks while preserving the directionality of symbolic stoichiometry, in order to improve the accuracy of candidate reaction ranking, model completion, and downstream fermentation phenotype prediction. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a metabolic regulation method based on a gap-filled genome-scale metabolic network model.

[0007] According to the first aspect of this application, a metabolic regulation method based on a gap-filled genome-scale metabolic network model is provided, comprising: Obtain metabolite and known reaction data of the GEM to be filled, construct stoichiometric matrix and binary correlation matrix; obtain a general reaction library containing several candidate reactions; Based on the stoichiometry matrix and the binary correlation matrix, the prior structural representation of the metabolite is obtained. The prior representation of the structure is input into a Transformer encoder based on a global self-attention mechanism, and the output is a refined metabolite representation containing contextual associations. Based on the refined metabolite representation, a refined metabolite representation matrix is ​​constructed. For each candidate reaction in the general reaction library, the refined metabolite representation matrix is ​​subjected to directed pooling aggregation based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. Based on the gap-filling prediction score, several candidate responses are selected, and the selected candidate responses are used to perform gap-filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. Flux balance analysis was performed using the gap-filled GEM, and a list of candidate regulatory targets was generated based on the flux balance analysis results.

[0008] Optionally, the step of obtaining metabolite and known reaction data of the GEM to be filled, and constructing a stoichiometric matrix and a binary correlation matrix, includes: Obtain the set of metabolites and known reaction data from the genome-scale metabolic network model to be filled; Based on the set of metabolites and the known reaction data, a binary correlation matrix and a stoichiometric matrix are constructed; the binary correlation matrix is ​​used to characterize the topological relationship between the metabolites in the set of metabolites and the reactions in the known reaction data; the stoichiometric matrix contains the stoichiometric coefficients of each metabolite in the known reaction data.

[0009] Optionally, obtaining the prior structural representation of the metabolite based on the stoichiometry matrix and the binary correlation matrix includes: Calculate the metabolite co-occurrence matrix based on the binary correlation matrix; A metabolite co-occurrence map is constructed using the metabolite co-occurrence matrix, the normalized graph Laplacian matrix of the metabolite co-occurrence map is calculated, and the Laplacian position code is extracted. Random walk codes are extracted from the transition matrix of the metabolite co-occurrence map; Extract the identifier of the metabolite and generate a compartment code based on the compartment suffix in the metabolite identifier; Based on the stoichiometry matrix, the number of reactions in which each metabolite participates is statistically encoded, and the consumption and generation state of each metabolite in the reaction are analyzed and encoded as dead ends. The Laplace position encoding, the random walk encoding, the compartment encoding, the degree encoding, and the dead end encoding are fused to obtain the prior structural representation of the metabolite.

[0010] Optionally, the Transformer encoder includes a multi-head self-attention layer, a residual connection layer, a layer normalization module, and a feedforward neural network, wherein: The multi-head self-attention layer is used to perform multi-head self-attention computation on the prior representation of the structure, extract the global dependencies between the metabolites, and output the attention feature representation. The residual connection layer is used to superimpose and fuse the prior structural representations of each metabolite with the attention feature representations; and the layer normalization module is used to normalize the superimposed and fused features to obtain intermediate features. The intermediate features are nonlinearly mapped and transformed by the feedforward neural network to output a refined metabolite representation of each metabolite that includes contextual association.

[0011] Optionally, the step of constructing a metabolite refinement representation matrix based on the refined metabolite representation, and for each candidate reaction in the general reaction library, performing directed pooling aggregation on the metabolite refinement representation matrix based on the composition of the substrate and product in the candidate reaction to obtain a gap-filling prediction score for each candidate reaction, includes: The refined representations of each metabolite are arranged sequentially according to the order of the metabolites in the stoichiometric matrix to obtain the refined representation matrix of metabolites. For each candidate reaction in the general reaction library, determine the stoichiometric vector corresponding to each candidate reaction; The negative and positive terms in the stoichiometric vectors corresponding to the candidate reactions are separated, and the separated negative terms are used as the substrate allocation vectors and the positive terms are used as the product allocation vectors; the substrate allocation vectors and the product allocation vectors are then normalized. The metabolite refined representation matrix is ​​weighted and pooled using the normalized substrate allocation vector to obtain the substrate feature representation; the metabolite refined representation matrix is ​​weighted and pooled using the normalized product allocation vector to obtain the product feature representation. Two different weight matrices are used to perform weighted transformations and fusions on the substrate feature representation and the product feature representation, respectively, to obtain the reaction level feature vector; The feature vectors of the reaction levels are scored to obtain the gap-filling prediction scores for each candidate reaction.

[0012] Optionally, the step of selecting a plurality of candidate responses based on the gap-filling prediction score, and using the selected candidate responses to perform gap-filling processing on the GEM to be filled, to obtain a gap-filled GEM, includes: All candidate responses are sorted in descending order according to the gap-filling prediction score, and the top preset number of candidate responses after sorting are selected as the target response. The target reaction is added to the GEM to be filled to obtain the GEM after gap filling.

[0013] Optionally, the flux balance analysis includes at least one of microbial fermentation phenotype prediction, metabolic pathway completion analysis, and strain performance evaluation.

[0014] According to a second aspect of this application, a metabolic regulation system based on a gap-filled genome-scale metabolic network model is provided, comprising: The data initialization module is used to acquire metabolite and known reaction data of the GEM to be filled, construct stoichiometric matrix and binary correlation matrix; and acquire a general reaction library containing several candidate reactions. The structure prior extraction module is used to obtain the structure prior representation of metabolites based on the stoichiometry matrix and the binary correlation matrix. The context encoding module is used to input the prior structural representation into a Transformer encoder based on a global self-attention mechanism, and output a refined metabolite representation containing contextual associations. The directed scoring module is used to construct a metabolite refinement representation matrix based on the refined metabolite representation. For each candidate reaction in the general reaction library, the module performs directed pooling aggregation on the metabolite refinement representation matrix based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. The network completion module is used to select several candidate responses based on the gap-filling prediction score, and use the selected candidate responses to perform gap-filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. The analysis and control module is used to perform flux balance analysis using the gap-filled GEM and generate a list of candidate control targets based on the flux balance analysis results.

[0015] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the metabolic regulation method based on a gap-filled genome-scale metabolic network model as described in the first aspect of this application.

[0016] According to a fourth aspect of this application, an electronic device is provided, comprising: At least one memory for storing program instructions; At least one processor is configured to invoke program instructions stored in the memory and execute the steps of the metabolic regulation method based on a gap-filled genome-scale metabolic network model as described in the first aspect of this application, according to the obtained program instructions.

[0017] This application provides a metabolic regulation method based on a gap-filled genome-scale metabolic network model. The structural prior representation is input into a Transformer encoder. By introducing a global self-attention mechanism, the method effectively alleviates the information oversmoothing problem caused by traditional local message passing mechanisms in genome-scale metabolic network models, enhancing the ability of gap-filling prediction scores to represent global relationships within the model. Directed pooling of substrates and products preserves reaction directionality and avoids the loss of directional information. Selecting suitable candidate reactions based on gap-filling prediction scores improves the accuracy of analysis and prediction in genome-scale metabolic network models. This method utilizes the internal structure of the GEM (Genomic Image Processing Model) and a universal reaction library for GEM completion, improving the accuracy of GEM analysis and prediction, and ultimately generating precise metabolic regulation strategies for target microorganisms.

[0018] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0019] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of a metabolic regulation method using a genome-scale metabolic network model in one embodiment of this application; Figure 2 This is a flowchart of a metabolic network gap-filling method in one embodiment of this application; Figure 3 This is a comparison chart of gap-filling performance under candidate reaction budgets in one embodiment of this application, wherein (a) is a data chart of aggregate F1 values ​​under different candidate budgets; and (b) is a chart of extended verification results when candidate budget K=250. Figure 4 This is an HPLC detection report of the fermentation product of sample 13-2 in one embodiment of this application; wherein, (a) is a chromatogram of the RID signal corresponding to the fermentation product 13-2; (b) is a chromatogram of the VWD signal corresponding to the fermentation product 13-2 at 210 nm; and (c) is a table of chromatographic peak integration results of the RID signal corresponding to the fermentation product 13-2, the integration results of which include at least retention time, peak width, peak area, peak height and peak area percentage. Figure 5This is an HPLC detection report of the fermentation product of sample 15-2 in one embodiment of this application; wherein, (a) is the chromatogram of the RID signal corresponding to the fermentation product of 15-2; (b) is the chromatogram of the VWD signal corresponding to the fermentation product of 15-2 at 210 nm; and (c) is a table of chromatographic peak integration results of the RID signal corresponding to the fermentation product of 15-2. Figure 6 This is a schematic diagram of the metabolic regulation system of a genome-scale metabolic network model in one embodiment of this application; Figure 7 This is a structural diagram of a metabolic network gap-filling system in one embodiment of this application. Detailed Implementation

[0020] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all fall within the protection scope of the present application. Parts not described in detail in the following embodiments can be implemented using existing technology.

[0021] It should be noted that all information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0022] Genome-scale metabolic network models, which describe intracellular metabolic processes through genome annotation, biochemical reaction knowledge, and stoichiometric constraints, are widely used in flux balance analysis, metabolic engineering design, microbial fermentation capacity prediction, and systems biology research. Existing gap-filling methods for genome-scale metabolic network models have the following shortcomings: First, local message passing can easily cause representation smoothing near highly connected metabolites, weakening the functional differences between money metabolites and pathway-specific metabolites; second, treating reactions as disordered collections of metabolites leads to the loss of reaction directionality information. Based on these problems, this application provides a metabolic regulation method based on a gap-filled genome-scale metabolic network model to address these issues.

[0023] Reference Figure 1 As shown in the embodiments of this application, a metabolic regulation method based on a gap-filled genome-scale metabolic network model is provided, including: S1. Obtain the metabolites and known reaction data of the GEM to be filled, and construct the stoichiometric matrix and binary correlation matrix; obtain a general reaction library, which contains several candidate reactions; S2. Obtain the prior structural representation of metabolites based on the stoichiometry matrix and the binary correlation matrix; S3. Input the structural prior representation into the Transformer encoder based on the global self-attention mechanism, and output the metabolite refined representation containing contextual association; S4. Construct a metabolite refining representation matrix based on the metabolite refining representation. For each candidate reaction in the general reaction library, perform directed pooling aggregation on the metabolite refining representation matrix based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. S5. Select several candidate responses based on the gap filling prediction score, and use the selected candidate responses to perform gap filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. S6. Perform flux balance analysis using the gap-filled GEM, and generate a list of candidate regulatory targets based on the flux balance analysis results.

[0024] For example, in step S6, based on the GEM after gap filling, a flux balance analysis is performed under preset culture conditions and reaction flux boundary conditions. Based on the flux balance analysis results, a list of candidate regulatory targets is generated for subsequent metabolic engineering analysis.

[0025] In the embodiments described above, the structural prior representation is input into the Transformer encoder. By introducing a global self-attention mechanism, the excessive smoothing of information caused by the traditional local message passing mechanism in genome-scale metabolic network models can be effectively alleviated, enhancing the ability of gap-filling prediction scores to represent global correlations in genome-scale metabolic network models. Directed pooling aggregation of substrates and products preserves reaction directionality and avoids the loss of reaction directionality information. Selecting suitable candidate reactions based on gap-filling prediction scores improves the accuracy of analysis and prediction in genome-scale metabolic network models. The method of this application can utilize the internal structure of GEMs and a universal reaction library to complete GEMs, improving the accuracy of GEM analysis and prediction, and thereby generating precise metabolic regulation strategies for target microorganisms.

[0026] In some specific embodiments of this application, obtaining metabolite and known reaction data of the GEM to be filled, and constructing a stoichiometric matrix and a binary correlation matrix, may further include: S11. Obtain the set of metabolites and known reaction data in the genome-scale metabolic network model to be filled; S12. Based on the metabolite set and known reaction data, construct a binary correlation matrix and a stoichiometric matrix; the binary correlation matrix is ​​used to characterize the topological relationship between metabolites in the metabolite set and reactions in the known reaction data; the stoichiometric matrix contains the stoichiometric coefficients of each metabolite in the known reaction data. If the stoichiometric coefficient is negative, it means that the metabolite is consumed as a substrate in the corresponding reaction; If the stoichiometric coefficient is positive, it means that the metabolite is generated as a product in the corresponding reaction; If the stoichiometric coefficient is zero, it means that the metabolite did not participate in the reaction.

[0027] For example, a stoichiometric matrix is ​​used to represent the reaction participation of metabolites and is labeled. The stoichiometric matrix is ​​an m×n dimensional matrix, where m is the number of metabolites in the metabolite set, and n is the total number of reactions in the known reaction data; matrix elements S ij The stoichiometric coefficient of the i-th metabolite in the j-th reaction is represented by the stoichiometric coefficient, which includes negative, positive and zero values, corresponding to the metabolite as a substrate, product or not participating in the reaction, respectively.

[0028] In some specific embodiments of this application, obtaining the prior structural representation of a metabolite based on a stoichiometric matrix and a binary correlation matrix may further include: S21. Calculate the metabolite co-occurrence matrix based on the binary correlation matrix; S22. Construct a metabolite co-occurrence map using the metabolite co-occurrence matrix, calculate the normalized graph Laplacian matrix of the metabolite co-occurrence map, and extract the Laplacian position code. S23. Extract random walk codes from the transition matrix of the metabolite co-occurrence map; S24. Extract the identifiers of metabolites and generate compartment codes based on the compartment suffixes in the metabolite identifiers; S25. Based on the stoichiometric matrix, statistically analyze the reaction quantity and generation degree coding of each metabolite, and analyze the consumption and generation state coding of each metabolite in the reaction. S26. Feature fusion of Laplace position coding, random walk coding, compartment coding, degree coding and dead end coding is performed to obtain the structural prior representation of metabolites.

[0029] For example, in step S2, based on the stoichiometric matrix, the binary correlation matrix, and the signed stoichiometric matrix, multidimensional structural and biological characteristics of metabolites are extracted to construct an initial structural prior representation of the metabolites. The specific steps are as follows: The transpose multiplication operation is performed on the binary correlation matrix to obtain the metabolite co-occurrence matrix. A metabolite co-occurrence graph is constructed using this matrix, and the normalized graph Laplacian matrix of the graph is calculated. Eigenvalue decomposition is performed on the normalized graph Laplacian matrix, and feature vectors are extracted as Laplacian position codes. The state transition probabilities between nodes are calculated based on the metabolite co-occurrence graph, generating a transition matrix. A random walk operation is performed on the transition matrix to generate a walk sequence, and embedding learning is performed on the walk sequence to obtain a random walk code. The compartment suffixes in the metabolite identifiers are parsed and mapped to generate compartment codes. The number of reactions in which metabolites participate in the genome-scale metabolic network model is counted to generate degree codes. Based on the stoichiometric coefficients of metabolites in each reaction in the stoichiometric matrix, the metabolite is determined to be in a strictly consumed, strictly generated, equilibrium, or isolated state to generate a dead-end code. The Laplacian position code, random walk code, compartment code, degree code, and dead-end code are concatenated to obtain the prior structural representation of the metabolite.

[0030] In some specific embodiments of this application, the Transformer encoder in step S3 includes a multi-head self-attention layer, a residual connection layer, a layer normalization module, and a feedforward neural network, wherein: Multi-head self-attention computation is performed on the structural prior representation through a multi-head self-attention layer to extract the global dependencies between various metabolites and output attention feature representations. The structural prior representations of each metabolite are superimposed and fused with the attention feature representations through a residual connection layer; and the superimposed and fused features are normalized through a layer normalization module to obtain intermediate features. By performing a nonlinear mapping transformation on intermediate features through a feedforward neural network, a refined metabolite representation containing contextual associations is output for each metabolite.

[0031] For example, the structural prior representation is input into a Transformer encoder based on a global self-attention mechanism. The self-attention mechanism captures the contextual association of metabolites in the global genome-scale metabolic network model and outputs a refined metabolite representation of the metabolite, so that each metabolite representation can directly integrate information of any other metabolite in the genome-scale metabolic network model.

[0032] In some specific embodiments of this application, for each candidate reaction in the general reaction library, directed pooling polymerization is performed on the metabolite refining representation based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction, which may further include: S41. Arrange the refined metabolite representations corresponding to each metabolite in the order of their arrangement in the stoichiometric matrix to obtain a refined metabolite representation matrix; wherein each row in the refined metabolite representation matrix corresponds to the refined metabolite representation of one metabolite. S42. For each candidate reaction in the general reaction library, determine the stoichiometric vector corresponding to each candidate reaction; S43. Separate the negative and positive terms from the stoichiometric vectors corresponding to the candidate reactions, and use the separated negative terms as the substrate partition vector and the positive terms as the product partition vector; normalize the substrate partition vector and the product partition vector respectively. S44. Use the normalized substrate allocation vector to perform weighted pooling on the metabolite refinement representation matrix to obtain the substrate feature representation; use the normalized product allocation vector to perform weighted pooling on the metabolite refinement representation matrix to obtain the product feature representation. S45. Two different weight matrices are used to perform weighted transformations and fusions on the substrate feature representation and the product feature representation respectively to obtain the reaction level feature vector; S46. Score the feature vectors of the reaction level to obtain the gap-filling prediction score for each candidate reaction.

[0033] For example, for candidate reactions in a general reaction library, firstly, the substrate, product, and their corresponding stoichiometric coefficients of the candidate reaction are obtained from the general reaction library, and a stoichiometric vector for the candidate reaction is constructed accordingly. Next, based on the signed stoichiometric vector of the candidate reaction, negative and positive terms are taken from the stoichiometric coefficients of the candidate reaction, respectively, to extract the substrate allocation vector and product allocation vector. Then, the substrate allocation vector and product allocation vector are normalized respectively, and the metabolite refinement representation matrix is ​​weighted and pooled using the normalized substrate allocation vector to obtain the substrate feature representation. The metabolite refinement representation matrix is ​​weighted and pooled using the normalized product allocation vector to obtain the product feature representation. The substrate feature representation and product feature representation are aggregated to generate a reaction-level feature vector. The reaction-level feature vector is obtained by adding the substrate feature representation and product feature representation after transformation with different weight matrices, so that reactions consuming the same metabolite and reactions generating the same metabolite are represented as different candidate reactions. Finally, the gap-filling prediction score of the candidate reaction is calculated based on the reaction-level feature vector.

[0034] Specifically, in the weighted pooling process, non-zero elements (i.e., stoichiometric coefficients of negative terms) in the substrate allocation vector are used as weights to select the refined metabolite representation corresponding to the candidate reaction substrate from the metabolite refined representation matrix, and weighted aggregation is performed according to the substrate allocation vector to obtain the substrate feature representation characterizing the overall characteristics of the candidate reaction substrate; non-zero elements (i.e., stoichiometric coefficients of positive terms) in the product allocation vector are used as weights to select the refined metabolite representation corresponding to the candidate reaction product from the metabolite refined representation matrix, and weighted aggregation is performed according to the product allocation vector to obtain the product feature representation characterizing the overall characteristics of the candidate reaction product; in the process of generating reaction-level feature vectors, the two weight matrices are weight matrices with different parameters to characterize the consumption direction attribute of the substrate and the generation direction attribute of the product in the candidate reaction, respectively, in the feature space.

[0035] In some specific embodiments of this application, a number of candidate responses are selected based on the gap-filling prediction score, and the selected candidate responses are used to perform gap-filling processing on the GEM to be filled, thereby obtaining a gap-filled GEM. This may further include: S51. Sort all candidate responses in descending order according to the gap-filling prediction score, and select the first preset number of candidate responses after sorting as the target response. S52. Add the target reaction to the GEM to be filled to obtain the GEM after gap filling.

[0036] For example, in step S5, candidate responses are ranked according to the gap-filling prediction score, and the top-ranked candidate responses are added to the genome-scale metabolic network model to be filled, resulting in the gap-filled genome-scale metabolic network model. The number of top-ranked candidate responses is a preset candidate budget K, where K ≤ 300.

[0037] Furthermore, before performing gap-filling prediction scoring for each candidate response on the gap-filling GEM to be filled, a candidate response scoring model is constructed and trained. The candidate response scoring model includes a structure prior extraction module, a Transformer encoder for the context encoding module, and a directed scoring module. Observed responses from multiple known GEMs are obtained as positive samples, and synthetic negative samples are generated. Based on the positive samples and synthetic negative samples, the structure prior extraction module, Transformer encoder, and directed scoring module are jointly trained using a binary cross-entropy loss function to obtain the trained candidate response scoring model. The gap-filling GEM to be filled is then input into the trained candidate response scoring model to obtain the gap-filling prediction scores for the candidate responses.

[0038] For example, for any positive sample, a predetermined proportion of participating metabolites are randomly retained from all participating metabolites contained in the positive sample, and the remaining participating metabolites are replaced with random metabolites that did not participate in the original reaction to obtain a synthetic negative sample, so that the generated synthetic negative sample has the same or similar number of participating metabolites as the original reaction; the generated synthetic negative sample is deduplicated.

[0039] In some specific embodiments of this application, the flux balance analysis in step S6 includes at least one of microbial fermentation phenotype prediction, metabolic pathway completion analysis, and strain performance evaluation.

[0040] For example, the gap-filled genome-scale metabolic network model can be used for microbial fermentation phenotype prediction, metabolic pathway completion, strain performance evaluation, strain design, or recommendation of candidate regulatory targets for metabolic engineering.

[0041] The gap-filled GEM can be used to perform metabolic simulation analysis on target microorganisms. Based on the results of the metabolic simulation analysis, metabolic regulation strategies targeting the target microorganisms can be generated. Among them, the metabolic simulation analysis includes at least one of microbial fermentation phenotype prediction, metabolic pathway completion analysis, and strain performance evaluation; the metabolic regulation strategy includes at least one of the following: a metabolic engineering candidate regulatory target recommendation scheme and a strain design scheme.

[0042] Metabolic simulation analysis includes at least one of the following: Microbial fermentation phenotype prediction: Under preset culture conditions, the growth phenotype of the target microorganism and / or the synthetic ability of the target product are predicted using gap-filled GEM. Metabolic pathway completion analysis: Based on the topology and flux distribution of the gap-filled GEM, the missing metabolic pathways in the specific product synthesis pathways of the target microorganism are identified, and a recommended list of candidate completion reactions is output. Strain performance evaluation: Under set environmental constraints, the intracellular metabolic flux distribution of the target microorganism and / or the theoretical maximum yield of the target product are evaluated by solving the gap-filled GEM. Metabolic regulation strategies include at least one of the following: Recommendation of candidate regulatory targets for metabolic engineering: Based on the results of metabolic simulation analysis, we recommend gene knockout, gene overexpression and / or introduction of exogenous genes that can improve the synthesis ability of target products. Strain design scheme: Based on the metabolic simulation analysis results, output a strain design scheme that includes specific modification sites, modification methods and expected phenotypic effects.

[0043] This application uses metabolites as the basic representation object. It first integrates structural priors such as topological location, random walks, compartments, degrees, and dead-end states, then obtains a global context representation through a Transformer encoder. Finally, it aggregates substrate and product representations based on the symbolic stoichiometry coefficients of candidate reactions to obtain a gap-filling prediction score. The gap-filled GEM obtained in this application improves the accuracy of flux balance analysis and fermentation phenotypic prediction. This method can automatically recommend reactions using the metabolic network's own structure even in the absence of large-scale phenotypic data, and reduces information loss caused by disordered hypergraph representation by preserving reaction directionality. This application reduces excessive smoothing in metabolic networks through a global self-attention mechanism; preserves reaction directionality through independent pooling on the substrate and product sides; connects the automatic recommendation results with flux balance analysis (FBA) phenotypic prediction through a candidate reaction ranking mechanism; and can still complete the model using the internal structure of the GEM and a general reaction library even in the absence of large-scale wet experimental labeled data.

[0044] The following specific embodiments / application examples will be used to further illustrate this application in order to better understand the above technical solutions of this application. It should be understood that the following are only some examples and are not intended to limit this application.

[0045] Example 1: Refer to Figure 2 The image shows a method for filling gaps in a metabolic network, which specifically includes the following steps: S1. Obtain the set of metabolites, the set of known reactions, and the candidate set of general reaction libraries from the draft GEM (i.e., the genome-scale metabolic network model to be filled). Construct a signed stoichiometric matrix S and a binary correlation matrix H. In the stoichiometric matrix S, negative values ​​indicate that the metabolite is consumed in the corresponding reaction, positive values ​​indicate that the metabolite is generated in the corresponding reaction, and zero values ​​indicate that the metabolite does not participate in the reaction. The binary correlation matrix H is used to indicate whether the metabolite participates in the corresponding reaction.

[0046] S2. Initialize the prior representation of metabolite structure: Specifically, calculate the metabolite co-occurrence matrix based on the binary correlation matrix H, and use the co-occurrence matrix to calculate the normalized graph Laplace matrix to extract the Laplace position code; extract the local diffusion structure code / random walk code based on the random walk transition matrix; extract the compartment code based on the metabolite identifier suffix; extract the degree code based on the number of reactions in which the metabolite participates; and determine whether the metabolite is strictly consumed, strictly generated, in equilibrium, or isolated based on the stoichiometric matrix S to generate a dead-end code. The above Laplace position code, random walk code, compartment code, degree code, and dead-end code are fused with the metabolite participation mode projection vector to form the initial prior representation of the structure.

[0047] S3. Input the initial structural prior representation into the Transformer encoder. The encoder contains a multi-head self-attention layer, a feedforward neural network, residual connections, and layer normalization. Unlike graph neural networks that rely solely on neighboring reaction propagation, the Transformer allows each metabolite to directly access other metabolites in the global network, thereby obtaining a refined metabolite representation that includes the global network context.

[0048] S4. Directed scoring of candidate reactions. For any candidate reaction, the negative terms in its stoichiometric vector are normalized to a substrate allocation vector, and the positive terms are normalized to a product allocation vector. The metabolite refining representation matrix is ​​weighted and pooled using the substrate allocation vector and product allocation vector respectively to obtain the substrate feature representation and product feature representation. Then, different weight matrices are used to transform and merge the results, which are then input into the classifier to obtain a normalized gap-filling prediction score, such as a score between 0 and 1.

[0049] S5. Rank the candidate reactions according to the gap-filling prediction score, and add the Top-K candidate reactions to the draft GEM to obtain the gap-filled GEM. The gap-filled GEM can be further used for FBA analysis, fermentation product prediction, pathway completion, metabolic engineering design, or strain design.

[0050] In one specific implementation, the GEM model is trained using known reactions as positive samples and generates synthetic negative samples through topological substitution. Specifically, for a positive sample reaction, some participating metabolites are retained, and the remaining participating metabolites are replaced with random metabolites from the network that did not participate in that reaction, thereby constructing negative samples with similar reaction sizes but unreasonable topological combinations. The GEM model is trained using a binary cross-entropy loss function to learn metabolite context representation and reaction plausibility discrimination capabilities. For ranking real candidate reactions, the model uses symbolic stoichiometry coefficients provided by a general reaction library to perform directed substrate / product scoring.

[0051] The above metabolic network gap-filling method uses global context to represent metabolites and preserves substrate / product directionality when scoring candidate reactions.

[0052] Application Example 1: Computational Validation Based on the Zimmermann / CarveMe (Automated Genome-Scale Metabolic Model Construction Tool) Benchmark This application example uses 24 bacterial draft GEMs reconstructed by CarveMe and a Zimmermann / CarveMe external validation task consisting of 10 fermentation products. After incorporating the Top-K candidate reactions recommended by the method in this application into the draft model, FBA was used to analyze the fermentation product production capacity and compared with experimental phenotypic labels.

[0053] Reference Figure 3As shown in Table 1, at K=100, the aggregation F1 value of the method in this application is 0.3150, which is close to that of the strong baseline method; at K=200, the aggregation F1 value increases to 0.3692; and at K=300, the aggregation F1 value reaches 0.4173, which is higher than the 0.3538 of the existing method CHESHIRE under the same budget. This result indicates that as the candidate reaction budget increases, the candidate ranking advantage of the method in this application can be translated into more real positive fermentation phenotype recovery.

[0054] Table 1: Aggregate F1 Values ​​under Different Candidate Budgets Application Example 2: Extended Validation To demonstrate the effectiveness of this method under larger candidate budgets and wider reaction tank conditions, this application example uses 30 sketch models and 12 fermentation products as validation objects, recording the phenotypic recovery after the addition of candidate reactions at K=250. (Refer to...) Figure 3 The results show that the F1 score of the method in this application is 0.4582, and it recovers more target product generation phenotypes compared to the unfilled sketch model. This application example further illustrates the deployment, statistical methods, and graphical representation of the method in this application under different candidate budgets and reaction pool sizes.

[0055] Application Example 3: Detection of Independent Fermentation Products and HPLC Phenotypic Labeling This application example uses a fermentation detection panel consisting of three GEMs and seven fermentation products, including succinic acid, DL-lactic acid, formic acid, acetic acid, propionic acid, butyric acid, and acetone. Butyric acid and acetone were not detected in their respective strains and can be used as negative controls. Binary phenotypic labels were assigned based on the HPLC quantitative signals; a detected quantitative concentration was recorded as positive, and no detection was recorded as negative.

[0056] Table 2: Detection results of fermentation assay panel for different products In this fermentation assay panel, the unfilled draft model recovered 10 positive phenotypes, corresponding to a recall of 0.667 and an F1 score of 0.800. After adding the Top-300 candidate reactions recommended in this application's method, 14 positive phenotypes were recovered, corresponding to a recall of 0.933 and an F1 score of 0.966, and no false positives were introduced in the measured negative control products. (Refer to...) Figure 4 and Figure 5 It can be seen that the provided HPLC-RID single injection report can be used to illustrate the chromatographic detection, peak area statistics, and retention time recording methods of fermentation samples.

[0057] Application Example 4: Verification of Reactive Directional Ablation To verify the role of symbolic stoichiometry in candidate reaction ranking, the method of this application was compared with several ablation scoring methods, including substrate and product role reversal, merging substrate and product into an undirected representation, using only the substrate-side representation, and using only the product-side representation. The results show that in 24 Zimmermann / CarveMe sketch GEMs, the Top-K candidate sets obtained for each ablation variant deviated significantly from the complete directed scoring results. At K=300, the average overlap rate between the role reversal scoring and the complete model's Top-300 set was 0.254, the undirected merged scoring was 0.451, the substrate-only scoring was 0.488, and the product-only scoring was 0.443. These results indicate that preserving the symbolic directionality of substrates and products is a crucial technical feature of the candidate ranking method of this application.

[0058] Reference Figure 6 As shown, based on the same inventive concept, another embodiment of this application provides a metabolic regulation system based on a gap-filled genome-scale metabolic network model. The metabolic regulation system 100 includes: The data initialization module 110 is used to acquire the metabolites and known reaction data of the GEM to be filled, construct the stoichiometric matrix and the binary correlation matrix, and acquire a general reaction library containing several candidate reactions. The structure prior extraction module 120 is used to obtain the structure prior representation of metabolites based on the stoichiometry matrix and the binary correlation matrix; The context encoding module 130 is used to input the structural prior representation into a Transformer encoder based on a global self-attention mechanism, and output a refined metabolite representation containing context association. The directed scoring module 140 is used to construct a metabolite refinement representation matrix based on the metabolite refinement representation. For each candidate reaction in the general reaction library, the metabolite refinement representation matrix is ​​subjected to directed pooling aggregation based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. The network completion module 150 is used to select several candidate responses based on the gap-filling prediction score, and to use the selected candidate responses to perform gap-filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. The analysis and regulation module 160 is used to perform flux balance analysis using the gap-filled GEM and generate a list of candidate regulation targets based on the flux balance analysis results.

[0059] It should be noted that the modules in the metabolic regulation system based on the gap-filled genome-scale metabolic network model provided in the above embodiments of this application correspond to the steps of the metabolic regulation method based on the gap-filled genome-scale metabolic network model in any of the above embodiments. Those skilled in the art can refer to the step features of the metabolic regulation method based on the gap-filled genome-scale metabolic network model to implement the corresponding modules in the metabolic regulation system based on the gap-filled genome-scale metabolic network model, which will not be repeated here.

[0060] Reference Figure 7 The method described in this application can be deployed as a local computing program, a server-side inference service, or an auto-completion module integrated into a metabolic model reconstruction process. After the user inputs the GEM to be filled and a general reaction library, the system automatically completes matrix construction, structure prior extraction, Transformer encoding, directed reaction scoring, and Top-K candidate reaction output. The output results can be a gap-filled SBML model, a candidate reaction ranking table, an FBA phenotypic prediction table, and a reaction evidence report for manual review.

[0061] The above embodiments are only some implementations of this application. Those skilled in the art can adjust the prior structure type, number of Transformer layers, candidate reaction budget, negative sample generation ratio, universal reaction library source, and FBA constraints without departing from the concept of this application. These adjustments are all equivalent variations of the technical solution of this application.

[0062] In another embodiment of this application, an electronic device is also provided, including a memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions stored in the memory and execute the steps of the metabolic regulation method based on the gap-filled genome-scale metabolic network model as described above according to the obtained program instructions.

[0063] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs and functional modules that implement the above methods), computer instructions, etc., and the aforementioned computer programs and computer instructions can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0064] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0065] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0066] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0067] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0069] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0071] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0072] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A metabolic regulation method based on a gap-filled genome-scale metabolic network model, characterized in that, include: Obtain metabolite and known reaction data of the GEM to be filled, construct stoichiometric matrix and binary correlation matrix; obtain a general reaction library containing several candidate reactions; Based on the stoichiometry matrix and the binary correlation matrix, the prior structural representation of the metabolite is obtained. The prior representation of the structure is input into a Transformer encoder based on a global self-attention mechanism, and the output is a refined metabolite representation containing contextual associations. Based on the refined metabolite representation, a refined metabolite representation matrix is ​​constructed. For each candidate reaction in the general reaction library, the refined metabolite representation matrix is ​​subjected to directed pooling aggregation based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. Based on the gap-filling prediction score, several candidate responses are selected, and the selected candidate responses are used to perform gap-filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. Flux balance analysis was performed using the gap-filled GEM, and a list of candidate regulatory targets was generated based on the flux balance analysis results.

2. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The process of obtaining metabolite and known reaction data of the GEM to be filled, and constructing a stoichiometric matrix and a binary correlation matrix includes: Obtain the set of metabolites and known reaction data from the genome-scale metabolic network model to be filled; Based on the set of metabolites and the known reaction data, a binary correlation matrix and a stoichiometric matrix are constructed; the binary correlation matrix is ​​used to characterize the topological relationship between the metabolites in the set of metabolites and the reactions in the known reaction data; the stoichiometric matrix contains the stoichiometric coefficients of each metabolite in the known reaction data.

3. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The step of obtaining the prior structural representation of the metabolite based on the stoichiometry matrix and the binary correlation matrix includes: Calculate the metabolite co-occurrence matrix based on the binary correlation matrix; A metabolite co-occurrence map is constructed using the metabolite co-occurrence matrix, the normalized graph Laplacian matrix of the metabolite co-occurrence map is calculated, and the Laplacian position code is extracted. Random walk codes are extracted from the transition matrix of the metabolite co-occurrence map; Extract the identifier of the metabolite and generate a compartment code based on the compartment suffix in the metabolite identifier; Based on the stoichiometry matrix, the number of reactions in which each metabolite participates is statistically encoded, and the consumption and generation state of each metabolite in the reaction are analyzed and encoded as dead ends. The Laplace position encoding, the random walk encoding, the compartment encoding, the degree encoding, and the dead end encoding are fused to obtain the prior structural representation of the metabolite.

4. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The Transformer encoder includes a multi-head self-attention layer, a residual connection layer, a layer normalization module, and a feedforward neural network, wherein: The multi-head self-attention layer is used to perform multi-head self-attention computation on the prior representation of the structure, extract the global dependencies between the metabolites, and output the attention feature representation. The residual connection layer is used to superimpose and fuse the prior structural representations of each metabolite with the attention feature representations; and the layer normalization module is used to normalize the superimposed and fused features to obtain intermediate features. The intermediate features are nonlinearly mapped and transformed by the feedforward neural network to output a refined metabolite representation of each metabolite that includes contextual association.

5. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The process involves constructing a metabolite refinement representation matrix based on the refined metabolite representation. For each candidate reaction in the general reaction library, the metabolite refinement representation matrix is ​​subjected to directed pooling aggregation based on the composition of the substrate and product in the candidate reaction to obtain a gap-filling prediction score for each candidate reaction, including: The refined representations of each metabolite are arranged sequentially according to the order of the metabolites in the stoichiometric matrix to obtain the refined representation matrix of metabolites. For each candidate reaction in the general reaction library, determine the stoichiometric vector corresponding to each candidate reaction; The negative and positive terms in the stoichiometric vectors corresponding to the candidate reactions are separated, and the separated negative terms are used as the substrate allocation vectors and the positive terms are used as the product allocation vectors; the substrate allocation vectors and the product allocation vectors are then normalized. The metabolite refined representation matrix is ​​weighted and pooled using the normalized substrate allocation vector to obtain the substrate feature representation; the metabolite refined representation matrix is ​​weighted and pooled using the normalized product allocation vector to obtain the product feature representation. Two different weight matrices are used to perform weighted transformations and fusions on the substrate feature representation and the product feature representation, respectively, to obtain the reaction level feature vector; The feature vectors of the reaction levels are scored to obtain the gap-filling prediction scores for each candidate reaction.

6. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The step of selecting several candidate responses based on the gap-filling prediction score, and using the selected candidate responses to perform gap-filling processing on the GEM to be filled, to obtain a gap-filled GEM, includes: All candidate responses are sorted in descending order according to the gap-filling prediction score, and the top preset number of candidate responses after sorting are selected as the target response. The target reaction is added to the GEM to be filled to obtain the GEM after gap filling.

7. The metabolic regulation method based on a gap-filled genome-scale metabolic network model according to claim 1, characterized in that, The flux balance analysis includes at least one of the following: microbial fermentation phenotype prediction, metabolic pathway completion analysis, and strain performance evaluation.

8. A metabolic regulation system based on a gap-filled genome-scale metabolic network model, characterized in that, include: The data initialization module is used to acquire metabolite and known reaction data of the GEM to be filled, construct stoichiometric matrix and binary correlation matrix; and acquire a general reaction library containing several candidate reactions. The structure prior extraction module is used to obtain the structure prior representation of metabolites based on the stoichiometry matrix and the binary correlation matrix. The context encoding module is used to input the prior structural representation into a Transformer encoder based on a global self-attention mechanism, and output a refined metabolite representation containing contextual associations. The directed scoring module is used to construct a metabolite refinement representation matrix based on the refined metabolite representation. For each candidate reaction in the general reaction library, the module performs directed pooling aggregation on the metabolite refinement representation matrix based on the composition of the substrate and product in the candidate reaction to obtain the gap-filling prediction score of each candidate reaction. The network completion module is used to select several candidate responses based on the gap-filling prediction score, and use the selected candidate responses to perform gap-filling processing on the GEM to be filled, so as to obtain the gap-filled GEM. The analysis and control module is used to perform flux balance analysis using the gap-filled GEM and generate a list of candidate control targets based on the flux balance analysis results.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: At least one memory for storing program instructions; At least one processor is configured to invoke program instructions stored in the memory and execute the steps of the method as described in any one of claims 1-7 according to the obtained program instructions.

Images