Genome scale metabolic network model deletion reaction prediction and filling method and system

Abstract

The invention discloses a genome scale metabolic network model deletion reaction prediction and filling method and system, and the method comprises the steps: obtaining a plurality of genome scale metabolic network models from a metabolic network database, and constructing a data set with a supervision label; extracting and fusing the molecular sequence and molecular map characteristics of the metabolite to obtain an initial characteristic vector; constructing a metabolic directed graph and a metabolic reaction hypergraph based on a positive reaction sample, performing feature directivity enhancement by using the directed graph, and extracting high-order topological information through a hypergraph convolutional neural network to obtain final feature representation of metabolites; on the basis of the feature representation, adopting an attention mechanism to predict candidate reaction confidence and training a model; and finally, screening a high-confidence reaction from the candidate reaction pool and filling the target model with the high-confidence reaction. According to the method, the metabolite multi-dimensional molecular characteristics and the network high-order topological information are deeply fused, the prediction accuracy is remarkably improved, the method does not depend on experimental data, and the method is suitable for efficient metabolic network model correction and optimization.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and bioinformatics technology, and in particular to a method and system for predicting and filling missing responses in a genome-scale metabolic network model. Background Technology

[0002] Genome-scale metabolic models (GEMs) are important bioinformatics tools for analyzing cellular metabolism and simulating physiological behavior. A GEM is a collection of such models, while a GEM refers to a single genome-scale metabolic network model. GEMs systematically describe gene-protein-response (GPR) relationships in cells, linking metabolites and reactions through a stoichiometric matrix and connecting reactions to corresponding enzymes through a reaction-gene matrix. GEMs reveal the underlying mechanisms from genotype to phenotype and can be combined with methods such as flux balance analysis (FBA) to predict cellular metabolic flux. Therefore, they are widely used in biomanufacturing and biomedical fields such as the rational design of chassis strains, metabolic engineering, therapeutic target identification, and drug development. Constructing high-quality genome-scale metabolic network models is of great value. In recent years, automated GEM construction tools such as ModelSEED, CarveMe, gapseq, and Merlin have emerged, significantly improving the efficiency of GEM construction. However, due to errors in genome functional annotation, the existence of unknown responses and pathways, GEMs generated by these automated construction tools often have many gaps (i.e. missing responses), which seriously affect the prediction accuracy of GEMs. Therefore, it is necessary to develop effective GEM gap-filling tools to correct the models.

[0003] Early gap-filling methods were primarily based on optimization strategies. Methods such as GrowMatch and MIRAGE addressed the inconsistency between GEM predictions and experimental data by introducing additional responses. However, these methods heavily rely on experimental phenotypic data, which is often difficult to obtain, especially in non-model organisms, thus significantly limiting their practicality. To overcome this dependence on experimental data, gap-filling methods based on metabolic network topology have emerged, such as FASTGAPFILL and Meneco. These methods do not require phenotypic data, but their accuracy is limited, particularly when dealing with large-scale, complex metabolic networks, where precise filling results are often difficult to achieve.

[0004] In recent years, machine learning algorithms have been introduced into the gap-filling task of GEMs, mainly transforming the problem of predicting missing responses in metabolic networks into a hyperedge prediction problem in hypergraph learning. For example, C3MM is a hyperedge prediction algorithm based on the cluster closure assumption; NHP uses graph convolutional networks to predict candidate responses that may be missing in metabolic networks; CHESHIRE uses Chebyshev spectral graph convolutional networks for feature refinement and performs gap filling based on the predicted scores of candidate responses; HGNNP is a hypergraph neural network for hyperedge prediction that can effectively mine higher-order correlations in the network; DSHCNet uses a dual-scale fusion hypergraph convolution method to predict missing responses, effectively distinguishing between substrates and products. However, the node feature embeddings of these methods are mostly learned purely from the topology of the GEM, failing to fully utilize the molecular biochemical information of the metabolites themselves, thus limiting further performance improvements. Multi-HGNN is currently the latest method for predicting missing responses in GEMs using hypergraph learning. It attempts to extract the biochemical features of metabolites from the molecular graph and use them as node feature embeddings for hypergraph learning. Nevertheless, existing methods still have shortcomings in deeply integrating multidimensional molecular features of metabolites with high-order topological information of metabolic networks, and their predictive performance needs further improvement. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the problems of existing fill-in methods being highly dependent on experimental data, having high computational complexity, long cycle, and low accuracy, as well as existing machine learning methods failing to fully integrate multidimensional molecular features of metabolites with high-order topological information of metabolic networks, resulting in limited performance in predicting missing reactions.

[0006] To address the aforementioned technical problems, this invention provides a method for predicting and imputing missing responses in a genome-scale metabolic network model, comprising the following steps: S1: Obtain multiple genome-scale metabolic network models from a metabolic network database, extract the real responses from the genome-scale metabolic network models as positive response samples, and generate negative response samples corresponding to the real response samples; assign labels to the positive response samples and the negative response samples respectively to form a supervised labeled dataset; S2: For all metabolites in the dataset, extract the molecular sequence representation and molecular graph structure representation of each metabolite; based on the molecular sequence representation and molecular graph structure representation, extract and fuse the initial feature vector of each metabolite to obtain the initial feature vector of the metabolite. S3: Based on all positive reaction samples in the dataset, construct a directed metabolic graph and a metabolic response hypergraph to characterize the metabolic network structure; use the directed metabolic graph to perform directional enhancement on the initial feature vector of the metabolites to obtain enhanced metabolite features; input the enhanced metabolite features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, and output the final feature representation of metabolites that fuses the network topology; S4: Based on the final feature representation of the metabolites, use an attention mechanism to predict the confidence of any candidate reaction consisting of a set of metabolites to obtain a prediction model; S5: Train the prediction model using the dataset to obtain the final response prediction model; S6: Obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; use the final reaction prediction model to predict the confidence scores of the reactions in the candidate reaction pool, and select a preset number of candidate reactions based on the predicted confidence scores to fill the target metabolic network model.

[0007] In one embodiment of the present invention, in step S3, the method of enhancing the directionality of the initial feature vector of the metabolite using the directed metabolic graph to obtain the enhanced metabolite feature is as follows: construct a directed metabolic graph based on all positive reaction samples, calculate the directional smoothing matrix and the directional guide number matrix using the adjacency matrix, degree matrix and Laplacian matrix of the directed metabolic graph, and transform the initial feature vector of the metabolite using the directional smoothing matrix and the directional guide number matrix to obtain the enhanced metabolite feature with directional information.

[0008] In one embodiment of the present invention, in step S3, the method of inputting the metabolite enhancement features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction and outputting the final metabolite feature representation of the fused network topology is as follows: construct a metabolic response hypergraph based on all positive response samples, input the metabolite enhancement features into the hypergraph convolutional neural network, perform aggregation from node features to hyperedge features and aggregation from hyperedge features to node features in sequence, extract high-order interaction information of the metabolic network, and output the final metabolite feature representation.

[0009] In one embodiment of the present invention, the method for predicting the confidence level of any candidate reaction consisting of a set of metabolites using an attention mechanism to obtain a prediction model in step S4 is as follows: For a candidate reaction consisting of a set of metabolite nodes, the feature vectors of each metabolite node are input into the position feedforward network and the multi-head attention layer, respectively, to obtain the static embedding vector and dynamic embedding vector of each node. Calculate the Hadamard square of the difference between the static embedding vector and the dynamic embedding vector of each node; The squaring result is input into the neural network to obtain the score of each node; The confidence score of the candidate response is obtained by averaging the scores of all nodes. Based on the confidence scores, a predictive model is constructed to predict whether a candidate response belongs to the target metabolic network model.

[0010] In one embodiment of the present invention, the method for training the prediction model using the dataset to obtain the final response prediction model in step S5 is as follows: The true labels of positive and negative reaction samples in the dataset are used as supervision signals; The confidence score output by the prediction model for the training samples is input into the binary cross-entropy loss function along with the corresponding true label to calculate the loss value; Based on the loss value, all learnable parameters of the prediction model are iteratively updated through an optimization algorithm until the model converges, thus obtaining the trained final response prediction model.

[0011] In one embodiment of the present invention, in step S1, the method for obtaining the negative reaction sample corresponding to the real reaction sample is as follows: randomly select a metabolite node from the real reaction sample, and randomly select a metabolite node that does not belong to the real reaction sample from the preset global metabolite set to replace it, thereby obtaining the corresponding negative reaction sample.

[0012] In one embodiment of the present invention, in step S2, based on the molecular sequence representation and the molecular graph structure representation, the initial feature vector of each metabolite is extracted and fused to obtain the initial feature vector of the metabolite. The SMILES sequences of metabolites are obtained as molecular sequence representations; the SMILES sequences are transformed into two-dimensional molecular graphs containing atoms, chemical bonds, and corresponding properties, as molecular graph structure representations; one-dimensional molecular features are extracted from the SMILES sequences using a molecular pre-training model based on the Transformer architecture; two-dimensional molecular features are extracted from the two-dimensional molecular graphs using a molecular pre-training model based on a graph neural network; the one-dimensional molecular features of each metabolite are concatenated with the two-dimensional molecular features, and fused into an initial feature vector of the metabolite through a feedforward neural network.

[0013] In one embodiment of the present invention, in step S6, the confidence scores of the reactions in the candidate reaction pool are predicted using the final reaction prediction model, and a preset number of candidate reactions are selected based on the predicted confidence scores and added to the target metabolic network model. The method is as follows: Set a threshold for the number of reactions to be filled; select reactions with confidence scores higher than the preset threshold from the candidate reaction pool; sort the selected reactions according to their similarity to existing reactions in the target metabolic network model; select the preset number of top-ranked reactions to fill the gaps.

[0014] This invention also provides a system for predicting and imputing missing responses in a genome-scale metabolic network model, comprising: The data construction module is used to obtain multiple genome-scale metabolic network models from a metabolic network database, extract real reactions from the genome-scale metabolic network models as positive reaction samples, generate negative reaction samples corresponding to the real reaction samples, and assign labels to the positive reaction samples and the negative reaction samples respectively to form a supervised labeled dataset. The feature extraction and fusion module is used to extract the molecular sequence representation and molecular graph structure representation of each metabolite for all metabolites in the dataset; and to extract and fuse the molecular sequence representation and molecular graph structure representation to obtain the initial feature vector of each metabolite. The graph neural network feature enhancement module is used to construct a directed metabolic graph and a metabolic response hypergraph representing the metabolic network structure based on all positive response samples in the dataset; to perform directional enhancement on the initial feature vector of the metabolites using the directed metabolic graph to obtain enhanced metabolite features; to input the enhanced metabolite features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, and to output the final feature representation of metabolites that fuses the network topology. The reaction prediction module is used to predict the confidence of any candidate reaction consisting of a set of metabolites based on the final feature representation of the metabolites using an attention mechanism, thereby obtaining a prediction model. The model training module is used to train the prediction model using the dataset to obtain the final response prediction model; The filling application module is used to obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; the final reaction prediction model is used to predict the confidence scores of the reactions in the candidate reaction pool, and a preset number of candidate reactions are selected based on the predicted confidence scores and filled into the target metabolic network model.

[0015] The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the genome-scale metabolic network model missing response prediction and imputation method.

[0016] The technical solution of the present invention has the following advantages over the prior art: This invention significantly improves the accuracy and reliability of missing reaction prediction by deeply fusing multi-dimensional molecular features of metabolites with high-order topological information of metabolic networks. This method effectively overcomes the dependence of existing technologies on experimental data, achieving efficient prediction based solely on information from the metabolic network itself, and possesses broad applicability and strong practicality. By combining directional graph learning, hypergraph convolution, and attention mechanisms, the system can fully capture the dynamic roles and interactions of metabolites in the reaction network, thereby maintaining high discriminative ability while improving the model's stability and generalization performance in real-world scenarios. The metabolic network model filled using this method shows substantial improvements in both structural integrity and phenotypic prediction reliability, providing strong support for model construction and optimization in metabolic engineering, synthetic biology, and related fields. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the method for predicting and imputing missing responses in a genome-scale metabolic network model provided in this embodiment of the invention. Figure 2 This is a schematic diagram of a partial implementation structure for enhancing local features in an embodiment of the present invention; Figure 3 Box plot showing the results of reaction prediction verification in the experiment; Figure 4 Box plot of the recovery rate prediction results for negative samples of the test set obtained by random sampling in the experiment; Figure 5 This is a box plot showing the predicted recovery rate of the entire database as negative samples in the test set during the experiment. Figure 6 Box plots showing the phenotypic prediction performance after GEM gap filling with other methods in the experiment. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0020] Example 1: like Figure 1 As shown, this invention provides a method for predicting and imputing missing responses in a genome-scale metabolic network model, comprising the following steps: S1: Obtain multiple genome-scale metabolic network models from a metabolic network database, extract the real responses from the genome-scale metabolic network models as positive response samples, and generate negative response samples corresponding to the real response samples; assign labels to the positive response samples and the negative response samples respectively to form a supervised labeled dataset; S2: For all metabolites in the dataset, extract the molecular sequence representation and molecular graph structure representation of each metabolite; based on the molecular sequence representation and molecular graph structure representation, extract and fuse the initial feature vector of each metabolite to obtain the initial feature vector of the metabolite. S3: Based on all positive reaction samples in the dataset, construct a directed metabolic graph and a metabolic response hypergraph to characterize the metabolic network structure; use the directed metabolic graph to perform directional enhancement on the initial feature vector of the metabolites to obtain enhanced metabolite features; input the enhanced metabolite features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, and output the final feature representation of metabolites that fuses the network topology; S4: Based on the final feature representation of the metabolites, use an attention mechanism to predict the confidence of any candidate reaction consisting of a set of metabolites to obtain a prediction model; S5: Train the prediction model using the dataset to obtain the final response prediction model; S6: Obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; use the final reaction prediction model to predict the confidence scores of the reactions in the candidate reaction pool, and select a preset number of candidate reactions based on the predicted confidence scores to fill the target metabolic network model.

[0021] The method of this invention first constructs a labeled dataset to extract and fuse multi-dimensional molecular features of metabolites. Then, it utilizes directed metabolic graphs and hypergraphs to learn the directional structure and high-order interaction information of the network, respectively, to obtain a metabolite characterization that integrates topological and biochemical properties. Based on this, an attention mechanism is used to construct a prediction model to evaluate the confidence level of candidate reactions belonging to the target metabolic network. Finally, high-confidence reactions are selected and filled based on the model output. This method effectively fuses metabolite molecular information and network topological features, significantly improving prediction accuracy, and is independent of experimental data, making it suitable for efficient construction and correction of genome-scale metabolic network models.

[0022] In this embodiment of the invention, the method is named HGLMA. In step S1, multiple genome-scale metabolic network models are obtained from the metabolic network database. The real reactions in the genome-scale metabolic network models are extracted as positive reaction samples, and negative reaction samples corresponding to the real reaction samples are generated. Labels are assigned to the positive reaction samples and the negative reaction samples respectively to form a supervised labeled dataset.

[0023] Specifically, when constructing the training and testing datasets for predicting missing responses in genome-scale metabolic network models, 108 genome-scale metabolic network models were selected from the BiGG database, and all metabolite and response information contained in these models was obtained, totaling 10,393 metabolites and 16,337 metabolic responses, to construct the response set used for training and testing. Simultaneously, a general response database was downloaded from the BiGG database as a candidate response pool for prediction and imputation in subsequent steps.

[0024] Simultaneously, the SMILES sequence information of the aforementioned metabolites was obtained from the publicly available MetaNetX website and used as the molecular sequence representation of the metabolites. The RDKit tool was used to convert the SMILES sequence of each metabolite into a corresponding two-dimensional molecular graph representation. Specifically, atomic properties such as atomic number, hybridization, formal charge, and aromaticity were extracted from the molecular graph constructed by RDKit and converted into tensor form as graph node features. Bond properties were extracted as edge features, and the connections between atoms were extracted as edge indices. A graph data structure was constructed using the Data method of the publicly available PyTorch Geometric library.

[0025] Each metabolite A two-dimensional molecular diagram can be represented as ,in It is an atomic property matrix. This is a key attribute matrix containing edge connection information. Based on this, for each genome-scale metabolic network model, each real reaction is directly used as a positive sample and assigned label 1. To generate corresponding negative samples, a random replacement strategy is used: a metabolite node is randomly selected from the positive sample reaction, and then a metabolite node not belonging to that reaction is randomly selected from the global metabolite set (10393 metabolites) to replace it, thus constructing a negative reaction sample with the same number and label 0. This strategy effectively simulates abnormal or impossible metabolite combinations in the reaction, ensuring the diversity and opposition of the dataset, and providing high-quality negative example support for subsequent supervised learning.

[0026] Furthermore, in step S2, multi-dimensional features of all metabolites in the dataset are extracted and fused to obtain an initial feature vector for each metabolite.

[0027] Specifically, such as Figure 2 As shown, for the dataset containing To fully extract molecular information from the SMILES sequences, a pre-trained model based on the Transformer architecture, ChemBERTa, was used to encode the sequences and extract their one-dimensional sequence features, resulting in corresponding feature representation vectors. ,in .

[0028] The ChemBERTa model is a bidirectional encoder based on the BERT architecture, consisting of 6 encoder layers (12 attention heads per layer). It employs byte-level BPE word segmentation and is trained on a large corpus of the benchmark dataset ZINC using the RoBERTa pre-training method based on a random masking strategy. The output dimension is... It is an inherent property of the model architecture, representing the richness of molecular sequence features learned by the model.

[0029] Simultaneously, the pre-trained graph MVP model based on graph neural networks was used to analyze metabolites. Two-dimensional molecular diagram Encode the molecules, extract their two-dimensional molecular features, and implicitly extract their three-dimensional molecular features to obtain feature vectors. ,in .

[0030] GraphMVP employs a graph-based multi-view pre-trained molecular encoding framework. It performs self-supervised learning through the correspondence and consistency between two-dimensional topological structures and three-dimensional geometric views. It randomly selects 50,000 qualified molecules with two-dimensional and three-dimensional structures from the GEOM molecular conformation dataset for pre-training, forming a two-dimensional graph representation of metabolite molecules with latent three-dimensional features. and After splicing, the data is input into a single-layer feedforward neural network to obtain metabolites. Initial feature vector after multi-dimensional feature fusion : , in, , and These are the learnable weight matrix and bias vector, respectively.

[0031] This step effectively fuses and refines the sequence and structural information of each metabolite, outputting an initial feature vector of the metabolite containing rich biochemical semantics. This provides a high-quality, multi-dimensional fusion of node features for subsequent graph-based feature enhancement and response prediction.

[0032] Furthermore, in step S3, after obtaining the initial feature vector of the metabolites, graph representation learning is performed based on the metabolic network structure composed of all positive reaction samples in the dataset to enhance the feature representation.

[0033] Construct a directed metabolic graph based on the substrate-product relationships between metabolites in the positive reaction sample. ,in For a set of metabolite nodes, This is a set of directed edges representing the direction of reactions. Based on the metabolic directed graph, its adjacency matrix is ​​used... Obtain the degree diagonal matrix and Laplace matrix The directional smoothing matrix and the directional guide matrix are calculated, and then the initial feature vector of the metabolite is enhanced directionally.

[0034] Specifically, calculate the Laplacian matrix. of The eigenvectors corresponding to the smallest eigenvalues and vector field matrix , in, , , Representation matrix The row and number Column elements, Represents column vectors The Each element.

[0035] Furthermore, find the directed smoothing matrix. : And find the guided matrix. : , in, ,matrix From the vector field matrix get: , in, , For sufficiently small positive numbers, Representation matrix The OK, Represents a matrix or vector Take the absolute value of each element. Representing a real vector Norm.

[0036] The initial feature vector of the metabolite is transformed using the directional smoothing matrix and the directed guide matrix to obtain enhanced metabolite features with directional information. Directed graph learning explicitly introduces the flow information from substrate to product in the metabolic reaction, enabling feature encoding to distinguish the role of metabolites in the reaction (such as reactants or products), thus enhancing the biochemical rationality of the features.

[0037] Furthermore, to further extract higher-order interaction information at the reaction level in the metabolic network, a metabolic reaction hypergraph is constructed based on the same positive reaction samples.

[0038] Based on the initial feature vectors of metabolites in the metabolic network model, directed graph learning is used to obtain node metabolite feature vectors that incorporate network structure information. : , , , in, , , As initial features, For all nodes at the 1st The feature matrix of the layer.

[0039] Furthermore, the metabolic network model structure composed of positive reaction samples is represented as a hypergraph. ,in For a set of metabolite nodes, This represents the set of hyperedges corresponding to positive reactions, thus describing the higher-order interaction information between metabolites and positive reaction samples in the metabolic network model. A metabolic hypergraph is constructed based on positive samples to learn the higher-order relationships between samples and to learn the metabolite encoding embeddings.

[0040] According to the Hypergraph Correlation Matrix Find the diagonal matrix describing the hypermarginality. diagonal matrix of node degree .

[0041] The feature representation obtained from the directed graph network Enter to A layered hypergraph convolutional neural network obtains the hyperedge feature representation after aggregating the features of neighboring nodes in each layer. The node feature representation updated after aggregating the features of neighboring superedges : , , in, , ,and This is a learnable weight matrix. Therefore, the output node feature representation is obtained: , in For node metabolites The corresponding feature vectors yield the final feature representation of the metabolites, which integrates the network's higher-order topology and directional information. The final metabolite feature representation simultaneously encodes the metabolite's molecular properties, its directional role in the directed network, and its higher-order interaction patterns in multi-reaction synergy, providing a rich and structured representation basis.

[0042] Furthermore, in step S4, based on the final feature representation of metabolites, an attention mechanism is used to predict the confidence of any candidate reaction consisting of a set of metabolites, thus obtaining a prediction model.

[0043] Specifically, a multi-head self-attention mechanism is used to focus on any set of metabolite nodes. The candidate reactions are used to predict their confidence scores, obtaining the probability score that the reaction belongs to the target metabolic network model. First, for each metabolite node in the candidate reaction... , and its corresponding final feature vector The input is fed into a location feedforward network to obtain the static embedding vector of that node. : , in, , It is a learnable static embedding weight matrix. The static embedding weight matrix is ​​used to capture the inherent, context-independent feature representations of metabolite nodes.

[0044] At the same time, the same feature vector Input to a head number The multi-head attention layer aggregates information from other metabolites within the reaction through a self-attention mechanism to obtain the dynamic embedding vector of the node. : , in, For each point of attention, For the value matrix, the attention score satisfy: , in, , , These are the query matrix and the key matrix, respectively. Dynamic embedding can capture the associations and dependencies of metabolites in a specific reaction context.

[0045] Furthermore, to integrate static and dynamic information and highlight the importance of their differences for response discrimination, the Hadamard square of the difference between the static and dynamic embedding vectors of each node is calculated and input into a single-layer neural network to obtain the probability score of each node. : , in, , It is a learnable weight matrix. It is a learnable bias vector.

[0046] right Output of all nodes Calculate the average value to obtain The probability fraction of the corresponding metabolite constituting a candidate reaction : .

[0047] probability fraction By integrating the static properties and dynamic interactions of each metabolite within the reaction, this value directly reflects the probability that the reaction belongs to the target metabolic network model. The closer the value is to 1, the higher the confidence level of predicting it as a missing reaction (i.e., one that should be filled). Based on this confidence score, a binary classification prediction model can be constructed to predict whether a candidate reaction belongs to the target metabolic network model.

[0048] Further, in step S5, the prediction model is trained using the dataset to obtain the final response prediction model.

[0049] Specifically, based on the set batch size The training samples are divided into multiple training batches. For each training batch, steps S2 to S4 combine the metabolic network model structure consisting of all positive samples in the training set, as well as all positive and negative reaction samples and candidate reactions. Molecular information of each metabolite is used to predict the probability score of all reaction samples in each training batch. The binary cross-entropy loss function for each training batch is selected as follows: , in, , Indicates batch size, For the first The true label value of each reaction.

[0050] The loss function forces the model's output score to converge towards the true label, effectively measuring the prediction error. Based on the loss value, the Adam optimization algorithm is used to backpropagate and iteratively update all learnable parameters involved in steps S2 to S4. This training process is repeated until the required number of training epochs is reached. The hyperparameter settings for model construction and training are shown in Table 1.

[0051] Table 1:

[0052] Further, in step S6, all known reactions outside the target genome-scale metabolic network model are obtained from the metabolic reaction database to form a candidate reaction pool; the confidence scores of the reactions in the candidate reaction pool are predicted using the final reaction prediction model, and a preset number of candidate reactions are selected based on the predicted confidence scores and added to the target metabolic network model.

[0053] Specifically, all known real reactions, excluding those from the current target genome-scale metabolic network model, are retrieved from the BiGG database to form a candidate reaction pool. Each reaction in the candidate reaction pool is input into the pre-trained final reaction prediction model, and its corresponding confidence score is calculated. Based on a pre-set number of reactions to fill in and a pre-set confidence score threshold, the final reaction is then determined. (For example ), and select all candidates from the candidate reaction pool that meet the confidence score greater than The responses are used as a high-confidence candidate set.

[0054] Calculate the similarity score between each high-confidence candidate reaction and all existing reactions in the target genome-scale metabolic network model. This similarity score reflects the overall consistency between the candidate reaction and the existing metabolic network. The higher the score, the more similar the candidate reaction is to the existing reaction in terms of metabolite composition or function.

[0055] To introduce diversity and avoid redundancy during the imputation process, high-confidence candidate responses are sorted in ascending order of similarity scores, prioritizing responses with lower similarity to existing networks. Finally, a predetermined number of candidate responses are selected from the sorted list and formally imputed into the target genome-scale metabolic network model, thereby completing the prediction and imputation of missing responses and constructing a more structurally complete and functionally accurate metabolic network model.

[0056] By combining confidence scores and similarity scores as a dual screening mechanism, the high reliability of the filling response is ensured, while the structural diversity and functional coverage of the metabolic network are enhanced, significantly improving the construction quality and application value of genome-scale metabolic network models.

[0057] experiment: By observing the model's performance on the validation set, the model's hyperparameters were adjusted to optimize performance. Finally, all models were trained and tested on the same dataset. The models' prediction scores and predicted labels were compared across evaluation metrics including the area under the precision-recall curve (AUPRC), recall, F1 score, and accuracy.

[0058] Five existing methods were selected for comparison: C3MM, a hyperedge prediction algorithm based on the proposed cluster closure hypothesis; NHP, a graph neural network based on time-point processes, which uses graph convolutional networks to predict candidate reactions that may be missing in the metabolic network, refines features using Chebyshev spectral graph convolutional networks, and can fill gaps based on the predicted scores of candidate reactions; HGNNP, a hypergraph neural network that can be used for hyperedge prediction and mines higher-order correlations; and Multi-HGNN, which uses biochemical features of metabolites extracted from molecular graphs and feature embeddings extracted from directed graphs for further hypergraph learning.

[0059] Numerical results comparing the performance of the model in this embodiment with other models for four metrics: AUPRC, recall, F1 score, and accuracy. Table 2 shows the performance metrics of multiple prediction models in predicting missing responses in genome-scale metabolic network models, based on 108 high-quality BiGG genome-scale metabolic network models. Figure 3 As shown, the model in this embodiment outperforms other models on most datasets. In the corresponding 108 high-quality BiGG genome-scale metabolic network datasets, the response prediction performance of the method (HGLMA) in this embodiment outperforms the second-ranked method in four metrics: AUPRC, recall, F1 score, and accuracy, with performance improvements of approximately 5%, 22.4%, 15.9%, and 12.6%, respectively.

[0060] Table 2:

[0061] As shown in Table 3 and Figure 4The negative reactions shown in the test set are real reactions randomly selected from the BiGG database that do not belong to the target metabolic network. The top 25, top 50, top 100, and top N reactions with the highest prediction scores (where N is the number of manually removed reactions) were selected for testing, and the recovery rate (i.e., the proportion of reactions from the target metabolic network among the selected reactions) was used to evaluate the model's performance. The method (HGLMA) in this embodiment of the invention also outperforms the second-ranked model by 38.9%, 38.9%, 49.9%, and 47.9% in the recovery rates for the top 25, top 50, top 100, and top N, respectively. This demonstrates that the method of the present invention can effectively distinguish between real reactions and reactions belonging to the current metabolic network when trained using randomly generated negative samples.

[0062] Table 3:

[0063] As shown in Table 4 and Figure 5 The figure shows the recovery rate performance when all real BiGG reactions not belonging to the target metabolic network were used as negative reactions in the test set. Compared with other models, the model in this embodiment of the invention achieved the best predictive performance. In the recovery rate experiment using real reactions as the negative sample test set, after introducing all BiGG reactions, the recovery rate index decreased overall due to the sharp increase in the number of candidate reaction samples (from 465±257 to approximately 17,000), but it was still 58.8%, 62.7%, 78.6%, and 70% higher than the second-ranked model, reflecting the model's stability when a large number of candidate reactions were involved.

[0064] Table 4:

[0065] The method in this embodiment of the invention is used to fill GEM vacancies, and the COBRApy tool is used to predict the phenotype of the filled GEMs using different methods, inferring the fermentation products in the culture medium, and verifying the effectiveness of the method in this embodiment of the invention for GEM filling.

[0066] Specifically, initial GEMs of 24 bacteria constructed using the CarveMe method were selected. Based on fermentation experimental data provided in papers related to the Gapseq method, the predictive ability of the initial GEMs constructed using CarveMe, and GEMs filled with gaps using this invention and other methods (filling 200 reactions) were examined to assess the secretion of nine fermentation products. The performance of GEMs filled with gaps using different methods was evaluated using AUPRC, recall, F1 score, and accuracy metrics.

[0067] As shown in Table 5 and Figure 6As shown, for the initial GEM, random imputation offered almost no performance improvement, while other imputation methods demonstrated some performance improvement. The GEM imputed using this invention achieved optimal performance across all metrics. Compared to the GEM imputed by the second method (CHESHIRE), this invention improved AUPRC by 0.4%, recall by 24%, F1 score by 41.8%, and precision by 12.5%.

[0068] Table 5:

[0069] Example 2: Based on the same inventive concept as in Embodiment 1, this invention also provides a genome-scale metabolic network model missing response prediction and imputation system, used to implement the steps of the genome-scale metabolic network model missing response prediction and imputation method described in Embodiment 1, including the following modules: The data construction module is used to obtain multiple genome-scale metabolic network models from a metabolic network database, extract real reactions from the genome-scale metabolic network models as positive reaction samples, generate negative reaction samples corresponding to the real reaction samples, and assign labels to the positive reaction samples and the negative reaction samples respectively to form a supervised labeled dataset. The feature extraction and fusion module is used to extract the molecular sequence representation and molecular graph structure representation of each metabolite for all metabolites in the dataset; and to extract and fuse the molecular sequence representation and molecular graph structure representation to obtain the initial feature vector of each metabolite. The graph neural network feature enhancement module is used to construct a directed metabolic graph and a metabolic response hypergraph representing the metabolic network structure based on all positive response samples in the dataset; to perform directional enhancement on the initial feature vectors of metabolites using the directed metabolic graph to obtain enhanced metabolite features; and to input the enhanced metabolite features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, outputting a final metabolite feature representation that fuses the network topology. The reaction prediction module is used to predict the confidence of any candidate reaction consisting of a set of metabolites based on the final feature representation of the metabolites using an attention mechanism, thereby obtaining a prediction model. The model training module is used to train the prediction model using the dataset to obtain the final response prediction model; The filling application module is used to obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; the final reaction prediction model is used to predict the confidence scores of the reactions in the candidate reaction pool, and a preset number of candidate reactions are selected based on the predicted confidence scores and filled into the target metabolic network model.

[0070] Each module of the genome-scale metabolic network model missing response prediction and imputation system proposed in this embodiment is used to implement each step in the genome-scale metabolic network model missing response prediction and imputation method in Embodiment 1. To avoid redundancy, these steps will not be repeated here.

[0071] Example 3: The present invention also provides a computer storage medium storing a computer software product, the computer software product including several instructions for causing a computer device to execute the genome-scale metabolic network model missing response prediction and imputation method described in Embodiment 1.

[0072] 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.

[0073] 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.

[0074] 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 1The function specified in one or more boxes.

[0075] 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.

[0076] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for predicting and imputing missing responses in a genome-scale metabolic network model, characterized in that, Includes the following steps: S1: Obtain multiple genome-scale metabolic network models from a metabolic network database, extract the real responses from the genome-scale metabolic network models as positive response samples, and generate negative response samples corresponding to the real response samples; assign labels to the positive response samples and the negative response samples respectively to form a supervised labeled dataset; S2: For all metabolites in the dataset, extract the molecular sequence representation and molecular graph structure representation of each metabolite; based on the molecular sequence representation and molecular graph structure representation, extract and fuse the initial feature vector of each metabolite to obtain the initial feature vector of the metabolite. S3: Based on all positive reaction samples in the dataset, construct a directed metabolic graph and a metabolic response hypergraph to characterize the metabolic network structure; use the directed metabolic graph to perform directional enhancement on the initial feature vector of the metabolites to obtain enhanced metabolite features; The enhanced metabolite features are input into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, and the final metabolite feature representation with fused network topology is output. S4: Based on the final feature representation of the metabolites, use an attention mechanism to predict the confidence of any candidate reaction consisting of a set of metabolites to obtain a prediction model; S5: Train the prediction model using the dataset to obtain the final response prediction model; S6: Obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; use the final reaction prediction model to predict the confidence scores of the reactions in the candidate reaction pool, and select a preset number of candidate reactions based on the predicted confidence scores to fill the target metabolic network model.

2. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S3, the method for enhancing the directionality of the initial feature vector of the metabolites using the directed metabolic graph to obtain enhanced metabolite features is as follows: a directed metabolic graph is constructed based on all positive reaction samples; the directional smoothing matrix and the directional guide number matrix are calculated using the adjacency matrix, degree matrix, and Laplacian matrix of the directed metabolic graph; and the initial feature vector of the metabolites is transformed using the directional smoothing matrix and the directional guide number matrix to obtain enhanced metabolite features with directional information.

3. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S3, the method for inputting the metabolite enhancement features into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction and outputting the final metabolite feature representation of the fused network topology is as follows: construct a metabolic response hypergraph based on all positive response samples, input the metabolite enhancement features into the hypergraph convolutional neural network, perform aggregation from node features to hyperedge features and aggregation from hyperedge features to node features in sequence, extract high-order interaction information of the metabolic network, and output the final metabolite feature representation.

4. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S4, the method for using an attention mechanism to predict the confidence level of any candidate reaction consisting of a set of metabolites to obtain a prediction model is as follows: For a candidate reaction consisting of a set of metabolite nodes, the feature vectors of each metabolite node are input into the position feedforward network and the multi-head attention layer, respectively, to obtain the static embedding vector and dynamic embedding vector of each node. Calculate the Hadamard square of the difference between the static embedding vector and the dynamic embedding vector of each node; The squaring result is input into the neural network to obtain the score of each node; The confidence score of the candidate response is obtained by averaging the scores of all nodes. Based on the confidence scores, a predictive model is constructed to predict whether a candidate response belongs to the target metabolic network model.

5. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S5, the method for training the prediction model using the dataset to obtain the final response prediction model is as follows: The true labels of positive and negative reaction samples in the dataset are used as supervision signals; The confidence score output by the prediction model for the training samples is input into the binary cross-entropy loss function along with the corresponding true label to calculate the loss value; Based on the loss value, all learnable parameters of the prediction model are iteratively updated through an optimization algorithm until the model converges, thus obtaining the trained final response prediction model.

6. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S1, the method for obtaining the negative reaction sample corresponding to the real reaction sample is as follows: randomly select a metabolite node from the real reaction sample, and randomly select a metabolite node that does not belong to the real reaction sample from the preset global metabolite set to replace it, thereby obtaining the corresponding negative reaction sample.

7. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S2, based on the molecular sequence representation and molecular graph structure representation, the initial feature vector of each metabolite is extracted and fused. The method for obtaining the initial feature vector of the metabolite is as follows: The SMILES sequences of metabolites are obtained as molecular sequence representations; the SMILES sequences are converted into two-dimensional molecular graphs containing atoms, chemical bonds and corresponding properties as molecular graph structure representations; one-dimensional molecular features are extracted from the SMILES sequences using a molecular pre-trained model based on the Transformer architecture. Two-dimensional molecular features are extracted from the two-dimensional molecular graph using a molecular pre-trained model based on a graph neural network; the one-dimensional molecular features of each metabolite are concatenated with the two-dimensional molecular features and fused into an initial feature vector of the metabolite through a feedforward neural network.

8. The method for predicting and imputing missing responses in a genome-scale metabolic network model according to claim 1, characterized in that: In step S6, the confidence scores of the reactions in the candidate reaction pool are predicted using the final reaction prediction model. Based on the predicted confidence scores, a predetermined number of candidate reactions are selected and added to the target metabolic network model. The method is as follows: Set a threshold for the number of reactions to be filled; select reactions with confidence scores higher than the preset threshold from the candidate reaction pool; sort the selected reactions according to their similarity to existing reactions in the target metabolic network model; select the preset number of top-ranked reactions to fill the gaps.

9. A system for predicting and imputing missing responses in a genome-scale metabolic network model, characterized in that, include: The data construction module is used to obtain multiple genome-scale metabolic network models from a metabolic network database, extract real reactions from the genome-scale metabolic network models as positive reaction samples, generate negative reaction samples corresponding to the real reaction samples, and assign labels to the positive reaction samples and the negative reaction samples respectively to form a supervised labeled dataset. The feature extraction and fusion module is used to extract the molecular sequence representation and molecular graph structure representation of each metabolite for all metabolites in the dataset; and to extract and fuse the molecular sequence representation and molecular graph structure representation to obtain the initial feature vector of each metabolite. The graph neural network feature enhancement module is used to construct a metabolic directed graph and a metabolic response hypergraph representing the structure of the metabolic network based on all positive response samples in the dataset; and to use the metabolic directed graph to perform directional enhancement on the initial feature vector of the metabolites to obtain enhanced metabolite features. The enhanced metabolite features are input into a hypergraph convolutional neural network based on the metabolic response hypergraph for information extraction, and the final metabolite feature representation with fused network topology is output. The reaction prediction module is used to predict the confidence of any candidate reaction consisting of a set of metabolites based on the final feature representation of the metabolites using an attention mechanism, thereby obtaining a prediction model. The model training module is used to train the prediction model using the dataset to obtain the final response prediction model; The filling application module is used to obtain all known reactions outside the target genome-scale metabolic network model from the metabolic reaction database to form a candidate reaction pool; the final reaction prediction model is used to predict the confidence scores of the reactions in the candidate reaction pool, and a preset number of candidate reactions are selected based on the predicted confidence scores and filled into the target metabolic network model.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer software product, the computer software product including several instructions for causing a computer device to execute the genome-scale metabolic network model missing response prediction and imputation method according to any one of claims 1 to 8.

Images