A method for constructing a timing neural network dynamic enzyme constraint model based on unconventional yeast

Abstract

This invention discloses a method for constructing a dynamic enzyme constraint model based on a temporal neural network for unconventional yeast, aiming to solve the technical problems of staticity, rigid enzyme constraint updates, and poor numerical stability in traditional metabolic models. The method first constructs the framework of a static enzyme constraint model for unconventional yeast, integrates and completes enzyme kinetic parameters through deep learning, and establishes enzyme capacity constraints. It then constructs a dual-timescale coupling framework of macroscopic culture and microscopic metabolism, using enzyme constraint flux balance analysis to solve for instantaneous metabolic flux. An attention-enhanced long short-term memory network is designed to learn the temporal patterns of enzyme concentration, achieving dynamic allocation of enzyme resources. Physical laws are embedded into the loss function to train the network, and an adaptive time step ensures computational stability and efficiency. The dynamic metabolic trajectories of substrate, biomass, and products are output through iterative loops. This method is adaptable to various unconventional yeasts, can accurately simulate metabolic dynamics, and provides a general computational tool for the metabolic engineering modification of industrial strains.

Description

Technical Field

[0001] This invention relates to the field of biology, specifically to a method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast. Background Technology

[0002] Non-conventional yeasts are yeasts that are different from model yeasts such as Saccharomyces cerevisiae (Saccharomyces cerevisiae). Saccharomyces cerevisiae Kluyveromyces is a type of non-traditional yeast strain with special physiological characteristics and metabolic functions. Kluyveromyces marxianus It possesses excellent characteristics such as high temperature resistance, rapid growth rate, and the ability to utilize multiple carbon sources (including lactose, xylose, arabinose, inulin, etc.); Yersinia lipolytica ( Yarrowia lipolytica ) is a recognized safe strain (GRAS), possessing the ability to efficiently accumulate lipids and secrete proteins; Pichia pastoris ( Pichia stipitis These unconventional yeasts possess the unique advantage of fermenting xylose to produce ethanol. They have irreplaceable application value in fields such as industrial biotechnology, synthetic biology, and environmental remediation.

[0003] Genome-scale metabolic network models (GEMs) are core tools in systems biology research for top-down analysis of cellular metabolic mechanisms. Based on chemometric matrices, GEMs describe intracellular metabolic response networks using constrained modeling methods. Flux balance analysis (FBA) is the most commonly used analytical method in GEMs. It is based on the quasi-steady-state assumption, which states that intracellular metabolite concentrations remain constant over a short period, and the rate of metabolite production equals the rate of metabolite consumption. FBA predicts the optimal metabolic flux distribution of cells under specific conditions by solving a linear programming problem.

[0004] In the prior art, the patent with publication number US7711490B2 provides a method and system for simulating cell metabolism. The core is static metabolic modeling optimization based on flux balance analysis (FBA). By adding qualitative kinetic constraints, qualitative regulatory logic constraints, and differential DNA microarray data constraints, the flux boundary of FBA is tightened. Mixed integer linear programming (MILP) is used to solve the problem. It is used for gene knockout / addition, minimum metabolic response set identification, and strain phenotypic prediction. It is only applicable to model microorganisms such as Escherichia coli.

[0005] The patent with publication number CN119418751A provides a method for constructing a dynamic enzyme-constrained genome-scale metabolic network model. The core is a simple coupling of a local kinetic model and an enzyme-constrained genome-scale metabolic network (ecGEM). The kinetic flux results of local pathways such as glycolysis are used to constrain the upper and lower bounds of the flux of the corresponding reactions in the ecGEM. Then, the global dynamic flux is solved by FBA, which solves the problem that traditional ecGEM cannot output time series data, while still focusing on model bacteria such as Saccharomyces cerevisiae.

[0006] However, the existing technology has the following drawbacks:

[0007] Static assumption: FBA assumes the system is in a steady state and cannot capture the dynamic evolution of the metabolic network over time. In actual batch or fed-batch culture processes, key parameters such as substrate concentration, product concentration, and cell density continuously change, and the metabolic flux distribution also dynamically adjusts accordingly.

[0008] Ignoring enzyme resource constraints: FBA only considers stoichiometric and mass balance constraints, failing to account for the limiting effects of enzyme concentration and catalytic efficiency on metabolic flux. The total amount of protein within a cell is limited, and various metabolic pathways compete for limited enzyme resources.

[0009] Limitations of the optimality assumption: FBA assumes that cells optimize by maximizing their growth rate, but actual cell behavior may be influenced by complex regulatory networks. Summary of the Invention

[0010] The purpose of this invention is to provide a method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast, so as to solve the technical problems in the prior art, such as the inability of static models to capture metabolic dynamics, rigid enzyme constraint update mechanisms, and poor numerical stability.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast includes the following steps: S1. Obtain a genome-scale metabolic network model (GEM) of unconventional yeast, and construct the skeleton structure of a static enzyme constraint model based on chemometric matrix and gene-protein-reaction (GPR) association information.

[0012] Specifically, the GEM of the target unconventional yeast is obtained from public databases (such as KEGG, BioModels, BiGG) or literature. This model includes a metabolic response set R, a metabolite set M, a gene set G, and a chemometric matrix. .

[0013] Matrix element S ijThe stoichiometric coefficient of metabolite i in reaction j: If S ij <0 indicates that metabolite i is the substrate of reaction j; if S ij >0 indicates that metabolite i is a product of reaction j; if S ij =0 indicates that metabolite i does not participate in reaction j. Based on GPR association, a mapping relationship between genes, proteins, and reactions is established.

[0014] S2, integrase kinetic parameters, including catalytic turnover number k cat Based on the enzyme molecular weight (MW), missing enzyme parameters were supplemented using a large-scale enzyme parameter prediction database (GotEnzymes) and a deep learning model (DLKcat) to establish an enzyme capacity constraint equation.

[0015] For each enzyme-catalyzed reaction j, collect its experimentally determined k. cat value.

[0016] If experimental data is missing, GotEnzymes and the deep learning model DLK will be used. cat Replenish.

[0017] DLK cat Based on a fusion architecture of graph neural networks (GNNs) and convolutional neural networks (CNNs), this method predicts k using substrate molecular maps and protein sequences as input. cat value; GotEnzymes, on the other hand, predicts k by fusing pre-trained protein models and molecular encoders. cat .

[0018] The enzyme capacity constraint equation is: v j Let k be the flux of reaction j. cat,j E represents the enzyme turnover number catalyzing reaction j. j denoted as , where is the concentration of the enzyme catalyzing reaction j.

[0019] The total protein limit is: ; j∈R enz This represents the summation over all enzyme-catalyzed reactions / enzyme proteins, where j is the j-th enzyme (or the j-th enzyme-catalyzed reaction); Renz is the set of all enzyme proteins involved in metabolism (the subscript enz is an abbreviation for enzyme / enzyme-catalyzed reaction); MW j The molecular weight of the j-th enzyme; e j That is, the molar concentration / expression level of the j-th enzyme; MW j e j P represents the mass concentration of the j-th enzyme; total This indicates the upper limit of total enzyme protein / total protein in the cell.

[0020] S3. Construct a dual-timescale coupled framework: On the macro-timescale, establish an ordinary differential equation (ODE) system for substrate consumption and biomass accumulation; on the micro-timescale, perform enzyme-constrained flux balance analysis (ecFBA) within each discrete time step to calculate the optimal metabolic flux distribution.

[0021] Macro-timescale (cultivation timescale) The ODE system describes changes in the extracellular environment: ; Where X is the biomass concentration, S i v is the concentration of extracellular metabolite i, μ is the specific growth rate, and v is the concentration of extracellular metabolite i. ex,i Let be the exchange flux of metabolite i.

[0022] Microscopic timescale (metabolic timescale) Assuming the cell is in a pseudo-homeostatic state, perform ecFBA: ; Where max is the maximum value, c T is the coefficient row vector, and v is the metabolic flux column vector; Constraints include stoichiometry conservation Flux boundary constraints And enzyme capacity constraints: ; j∈R i Summation operator; MW i : Molecular weight of enzyme i; v j : Metabolic flux of reaction j; k cat,ij σ is the turnover number of enzyme i catalyzing reaction j. ij : Saturation coefficient / occupancy coefficient of enzyme i catalyzing reaction j; This indicates the quality requirement of reaction j for enzyme i; This represents the upper limit of available proteins for enzyme i (time-dependent).

[0023] S4. Design an attention-enhanced long short-term memory network (Attention-LSTM) that uses historical metabolic flux, enzyme concentration, and substrate concentration as inputs to learn the time-dependent patterns of dynamic changes in enzyme concentration.

[0024] The core equations of LSTM include: forget gate Input gate Candidate cell status Cell state update Output gate Hidden state update The attention mechanism dynamically adjusts the weights of information at different time steps by calculating the similarity between queries, keys, and values. ; in, Attention function, Q refers to the query matrix, K refers to the key matrix, and V refers to the value matrix; QK T Dot product similarity matrix; Scaling factor; : Scaled similarity matrix; softmax( ): softmax activation function.

[0025] S5. Establish a dynamic update equation for enzyme concentration, and couple the enzyme concentration changes predicted by LSTM with the flux constraints calculated by ecFBA to achieve dynamic optimization of enzyme resource allocation.

[0026] The dynamic changes in enzyme concentration are determined by three processes: synthesis, degradation, and dilution. ; Where, r j syn k represents the synthesis rate of enzyme j. deg,j is the degradation rate constant, and μ is the specific growth rate (dilution effect).

[0027] Enzyme synthesis rate predicted by LSTM: ; Where τ is the historical time, θ LSTM These are the network parameters. To ensure that the predicted enzyme concentration meets the total protein constraint, a projection operation is introduced: ; in, For the feasible region; S6. Construct a loss function with physical information constraints, and embed physical laws such as stoichiometry conservation, thermodynamic feasibility, and mass balance into the neural network training process.

[0028] The total loss function consists of a data fitting term and a physical constraint term: ; in, Total loss function; Data fitting loss; Regularized weights are adjustable hyperparameters of the model. Stoichiometric constraint loss; Thermodynamic constraint loss; : Loss due to mass balance constraints.

[0029] The data fitting loss expression is as follows: ; in, The summation operator integrates the biases of all experimental measurements to comprehensively evaluate the model's fit to all data. : Residual, the prediction bias of the i-th data point, quantifies the fitting error of the model to a single data point. The smaller the residual, the better the fitting effect.

[0030] Stoichiometric loss: ; : Net rate of change vector of metabolites; corresponding to the concentration change rate of the m-th metabolite.

[0031] Loss of thermodynamic feasibility: ; in, : Standard transformation Gibbs free energy change; when <0: The reaction is forward and spontaneous. >0: The reaction is forward but not spontaneous.

[0032] Mass balance loss: ; Total input of substances / isotopes; Total output of substances / isotopes.

[0033] By minimizing L total While fitting observed data, neural networks learn to obey physical laws, thereby improving the extrapolation ability and interpretability of the model.

[0034] S7. An adaptive time step strategy is adopted to dynamically adjust the integration step size according to the rate of change of metabolic state, so as to ensure numerical stability and computational efficiency.

[0035] Adaptive adjustment of the time step is based on local truncation error estimation: ; Among them, e local For local truncation error estimation, ε is the user-defined error tolerance, and α... max and α min To prevent drastic changes in the time step, δ is a small constant to prevent division by zero. During periods of rapid metabolic change (such as a sudden drop in substrate concentration or a large accumulation of products), the time step is automatically reduced to improve accuracy; during periods of relatively stable metabolic state, the time step is automatically increased to save computational resources.

[0036] S8. Iteratively execute the cycle of "ecFBA solution - flux update - ODE integration - enzyme concentration prediction" until the culture termination time is reached, and output the dynamic trajectory of substrate concentration, biomass and product concentration changes over time.

[0037] The complete iterative algorithm process includes: S8.1 At the current time step t, based on the extracellular substrate concentration S(t), calculate the upper limit of substrate uptake flux using the Michaelis-Menten kinetic equation: ; v uptake (t): Substrate uptake rate at time t; v max : Maximum reaction rate; S(t): Substrate concentration at time t; K m : Michaelis constant.

[0038] S8.2 Substitute the ingestion flux constraint into the ecFBA model to solve the linear programming problem and obtain the optimal flux distribution v(t) and specific growth rate μ(t); S8.3 Predicting enzyme synthesis rate r using LSTM syn (t); S8.4 uses a fourth-order Runge-Kutta (RK4) integral ODE system to update biomass X, substrate concentration S, and enzyme concentration e; S8.5 adaptively adjusts the time step Δt; S8.6 Update Time Repeat steps S8.1-S8.5 until the termination condition T is met.

[0039] Compared with the prior art, the beneficial effects of the present invention include: (1) Innovation of the dual timescale coupling framework: This invention is the first to deeply couple the macroscopic culture timescale with the microscopic metabolic timescale. The instantaneous metabolic state is calculated by ecFBA and the dynamic evolution is simulated by ODE integration, thus realizing the organic combination of static optimization and dynamic simulation.

[0040] (2) The introduction of attention-enhanced LSTM: Traditional dFBA methods use fixed rules or simple proportional relationships to update enzyme concentrations, which makes it difficult to capture complex temporal dependencies. This invention introduces an attention mechanism, enabling the model to adaptively focus on key historical information, which significantly improves the accuracy of enzyme concentration prediction.

[0041] (3) Training strategy with physical information constraints: This invention innovatively embeds physical laws such as stoichiometry conservation and thermodynamic feasibility into the neural network loss function, so that the model can comply with physical constraints while being data-driven, thereby improving the robustness and interpretability of the model.

[0042] (4) Adaptive time step strategy: The present invention dynamically adjusts the integral step size according to the rate of change of metabolic state, which significantly improves the computational efficiency while ensuring numerical stability and solves the problem of solving rigid ODE systems.

[0043] (5) Wide applicability: The method of the present invention is applicable to a variety of unconventional yeasts, including Kluyveromyces marxi, Yersinia lipolytica, Pichia pastoris, etc., providing a general computing platform for the metabolic engineering modification of industrial strains. Attached Figure Description

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] Figure 1 This is a metabolic pathway diagram of unconventional yeast in the examples; Figure 2 This is the amino acid prediction radar chart of iSM996 in the embodiment; Figure 3 This is a dynamic time and biomass 1-hour prediction graph based on a temporal neural network in the embodiment. Figure 4 This is a dynamic time and biomass 2h prediction graph based on a time-series neural network in the embodiment; Figure 5 This is a comparison chart of experimental results and theoretical predictions in the embodiments; Figure 6 This is a diagram illustrating the ethanol production results of the genetically engineered target in the embodiments; Figure 7 This is a diagram showing the glycerol content results of the predicted target sites in the embodiment of the genetic engineering modification; Figure 8 This is the overall flowchart of the method of the present invention. Detailed Implementation

[0046] The present invention describes a method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast, the overall flowchart of which is shown below. Figure 8 As shown, the specific implementation is as follows: Example 1 S1. Obtain a genome-scale metabolic network model (GEM) of unconventional yeast, and construct the skeleton structure of a static enzyme constraint model based on chemometric matrix and gene-protein-reaction (GPR) association information.

[0047] Specifically, the GEM of the target unconventional yeast is obtained from public databases (such as KEGG, BioModels, BiGG) or literature. This model includes a metabolic response set R, a metabolite set M, a gene set G, and a chemometric matrix. .

[0048] Matrix element S ij The stoichiometric coefficient of metabolite i in reaction j: If S ij <0 indicates that metabolite i is the substrate of reaction j; if S ij >0 indicates that metabolite i is a product of reaction j; if S ij =0 indicates that metabolite i does not participate in reaction j. Based on GPR association, a mapping relationship between genes, proteins, and reactions is established.

[0049] S2, integrase kinetic parameters, including catalytic turnover number k cat The enzyme molecular weight (MW) was used to supplement missing enzyme parameters using GotEnzymes and the deep learning model DLKcat, and an enzyme capacity constraint equation was established.

[0050] For each enzyme-catalyzed reaction j, collect its experimentally determined k. cat value.

[0051] If experimental data is missing, GotEnzymes and the deep learning model DLK will be used. cat Replenish.

[0052] DLK cat Based on a fusion architecture of graph neural networks (GNNs) and convolutional neural networks (CNNs), this method predicts k using substrate molecular maps and protein sequences as input. cat value; GotEnzymes, on the other hand, predicts k by fusing pre-trained protein models and molecular encoders. cat .

[0053] The enzyme capacity constraint equation is: v j Let k be the flux of reaction j. cat,j E represents the enzyme turnover number catalyzing reaction j. j denoted as , where is the concentration of the enzyme catalyzing reaction j.

[0054] The total protein limit is: ; j∈R enz This represents the summation over all enzyme-catalyzed reactions / enzyme proteins, where j is the j-th enzyme (or the j-th enzyme-catalyzed reaction); Renz is the set of all enzyme proteins involved in metabolism (the subscript enz is an abbreviation for enzyme / enzyme-catalyzed reaction); MW j The molecular weight of the j-th enzyme; e j That is, the molar concentration / expression level of the j-th enzyme; MW j e j P represents the mass concentration of the j-th enzyme; totalThis indicates the upper limit of total enzyme protein / total protein in the cell.

[0055] S3. Construct a dual-timescale coupled framework: On the macro-timescale, establish an ordinary differential equation (ODE) system for substrate consumption and biomass accumulation; on the micro-timescale, perform enzyme-constrained flux balance analysis (ecFBA) within each discrete time step to calculate the optimal metabolic flux distribution.

[0056] Macro-timescale (cultivation timescale) The ODE system describes changes in the extracellular environment: ; Where X is the biomass concentration, S i v is the concentration of extracellular metabolite i, μ is the specific growth rate, and v is the concentration of extracellular metabolite i. ex,i Let be the exchange flux of metabolite i.

[0057] Microscopic timescale (metabolic timescale) Assuming the cell is in a pseudo-homeostatic state, perform ecFBA: ; max v : Find the maximum value of variable v; μ: Objective function (specific cell growth rate); c T : Coefficient row vector (transposed); v: Metabolic flux column vector.

[0058] Constraints include stoichiometry conservation Flux boundary constraints And enzyme capacity constraints: ; j∈R i Summation operator; MW i : Molecular weight of enzyme i; v j : Metabolic flux of reaction j; k cat,ij σ is the turnover number of enzyme i catalyzing reaction j. ij : Saturation coefficient / occupancy coefficient of enzyme i catalyzing reaction j; This indicates the quality requirement of reaction j for enzyme i; This represents the upper limit of available proteins for enzyme i (time-dependent).

[0059] S4. Design an attention-enhanced long short-term memory network (Attention-LSTM) that uses historical metabolic flux, enzyme concentration, and substrate concentration as inputs to learn the time-dependent patterns of dynamic changes in enzyme concentration.

[0060] The core equations of LSTM include: forget gate Input gate Candidate cell status Cell state update Output gate Hidden state update The attention mechanism dynamically adjusts the weights of information at different time steps by calculating the similarity between queries, keys, and values. ; in, Attention function, Q refers to the query matrix, K refers to the key matrix, and V refers to the value matrix; QK T Dot product similarity matrix; Scaling factor; : Scaled similarity matrix; softmax( ): softmax activation function.

[0061] S5. Establish a dynamic update equation for enzyme concentration, and couple the enzyme concentration changes predicted by LSTM with the flux constraints calculated by ecFBA to achieve dynamic optimization of enzyme resource allocation.

[0062] The dynamic changes in enzyme concentration are determined by three processes: synthesis, degradation, and dilution. ; Where, r j syn k represents the synthesis rate of enzyme j. deg,j is the degradation rate constant, and μ is the specific growth rate (dilution effect).

[0063] Enzyme synthesis rate predicted by LSTM: ; Where τ is the historical time, θ LSTM These are the network parameters. To ensure that the predicted enzyme concentration meets the total protein constraint, a projection operation is introduced: ; in, For the feasible region; S6. Construct a loss function with physical information constraints, and embed physical laws such as stoichiometry conservation, thermodynamic feasibility, and mass balance into the neural network training process.

[0064] The total loss function consists of a data fitting term and a physical constraint term: ; in, Total loss function; Data fitting loss; Regularized weights are adjustable hyperparameters of the model. Stoichiometric constraint loss; Thermodynamic constraint loss; : Loss due to mass balance constraints.

[0065] The data fitting loss expression is as follows: ; in, The summation operator integrates the biases of all experimental measurements to comprehensively evaluate the model's fit to all data. : Residual, the prediction bias of the i-th data point, quantifies the fitting error of the model to a single data point. The smaller the residual, the better the fitting effect.

[0066] Stoichiometric loss: ; : Net rate of change vector of metabolites; corresponding to the concentration change rate of the m-th metabolite.

[0067] Loss of thermodynamic feasibility: ; in, : Standard transformation Gibbs free energy change; when <0: The reaction is forward and spontaneous. >0: The reaction is forward but not spontaneous.

[0068] Mass balance loss: ; Total input of substances / isotopes; Total output of substances / isotopes.

[0069] By minimizing L total While fitting observed data, neural networks learn to obey physical laws, thereby improving the extrapolation ability and interpretability of the model.

[0070] S7. An adaptive time step strategy is adopted to dynamically adjust the integration step size according to the rate of change of metabolic state, so as to ensure numerical stability and computational efficiency.

[0071] Adaptive adjustment of the time step is based on local truncation error estimation: ; Among them, e local For local truncation error estimation, ε is the user-defined error tolerance, and α... max and α min To prevent drastic changes in the time step, δ is a small constant to prevent division by zero. During periods of rapid metabolic change (such as a sudden drop in substrate concentration or a large accumulation of products), the time step is automatically reduced to improve accuracy; during periods of relatively stable metabolic state, the time step is automatically increased to save computational resources.

[0072] S8. Iteratively execute the cycle of "ecFBA solution - flux update - ODE integration - enzyme concentration prediction" until the culture termination time is reached, and output the dynamic trajectory of substrate concentration, biomass and product concentration changes over time.

[0073] The complete iterative algorithm process includes: S8.1 At the current time step t, based on the extracellular substrate concentration S(t), calculate the upper limit of substrate uptake flux using the Michaelis-Menten kinetic equation: ; v uptake (t): Substrate uptake rate at time t; v max : Maximum reaction rate; S(t): Substrate concentration at time t; K m : Michaelis constant.

[0074] S8.2 Substitute the ingestion flux constraint into the ecFBA model to solve the linear programming problem and obtain the optimal flux distribution v(t) and specific growth rate μ(t); S8.3 Predicting enzyme synthesis rate r using LSTM syn (t); S8.4 uses a fourth-order Runge-Kutta (RK4) integral ODE system to update biomass X, substrate concentration S, and enzyme concentration e; S8.5 adaptively adjusts the time step Δt; S8.6 Update Time Repeat steps S8.1-S8.5 until the termination condition T is met.

[0075] Example 2 This embodiment uses iSM966 as an example to construct a dynamic enzyme constraint model based on a time-series neural network, and uses it to predict the metabolic dynamics during its batch culture process, such as... Figure 1 The diagram shown is a metabolic pathway diagram of unconventional yeast.

[0076] A. Construction of a static enzyme constraint model Taking the publicly published iSM996 model (Marcisauskasetal., 2019) as an example, this model contains 1913 metabolic reactions, 1531 metabolites, and 996 genes. Using the GECKO toolbox, iSM996 is extended into an enzyme constraint model. The enzyme kinetic parameter k... cat The acquisition adopts a multi-source integration strategy: (1) Retrieve experimentally determined k from the BRENDA and SABIO-RK databases cat value; (2) For parameters missing in the database, use GotEnzymes to retrieve them; (3) For multi-subunit enzyme complexes, the smallest subunit k is used. cat As a limiting factor.

[0077] The total protein limit is set to P. total =0.5g·gCDW - ¹.

[0078] B. Implementation of the dual-timescale coupled framework Macro time scale: Establish a batch culture ODE system: ; Initial condition: X(0) = 0.5 g·L - ¹, S glc (0) = 20 g·L - ¹, P etoh (0) = 0 g·L - ¹, culture temperature 37°C, pH 5.5.

[0079] Microscale timescale: ecFBA is performed at each time step, with the objective function being to maximize the specific growth rate μ. Constraints include stoichiometry, substrate uptake limitation, enzyme capacity constraint, and total protein constraint.

[0080] C. Design and training of attention-enhancing LSTM Network architecture: The input layer has a feature dimension of 2074 (1913-dimensional flux vector + 156-dimensional enzyme concentration vector + 5-dimensional extracellular metabolite concentration vector). The Attention-LSTM layer uses a stacked 2-layer structure with 256 hidden units per layer, 8 attention heads, and a dropout rate of 0.2. The output layer is a fully connected layer with an output dimension of 156, predicting the synthesis rate of each enzyme. The training data includes 15 time series data collected from the literature, each containing 20-30 time points.

[0081] D. Prediction of amino acids from different substrate sources This invention analyzes the effects of different carbon sources on the intracellular amino acid metabolic profile of the strain using radar charts. Specific results are as follows: Figure 2 As shown, the results confirm that the type of carbon source can significantly regulate the amino acid metabolism of the strain, and D-glucose is the optimal carbon source for amino acid synthesis in the strain, which provides a theoretical basis for amino acid fermentation process based on carbon source optimization.

[0082] E. Comparison and verification of theoretical predictions and actual experiments regarding amino acids like Figure 5As shown, under single standard culture conditions, the flux predictions for different amino acid biosynthetic nodes show that the scatter plots of theoretical and experimental values ​​are closely distributed on both sides of the ideal concordance line (1:1 diagonal), exhibiting good symmetry and no systematic bias. Each data point corresponds to the metabolic state of different amino acid supply levels or key nodes in the biosynthetic pathway. This high clustering indicates that the enzyme-constrained metabolic network model (ecModel) can accurately capture the flux distribution patterns of the nitrogen metabolism network. Quantitative analysis shows that the coefficient of determination (R²) between the predicted and measured values ​​is high. 2 The accuracy was above 0.95, and the root mean square error (RMSE) was less than 15% of the experimental measurement uncertainty, confirming the high fidelity of the model in the quantitative prediction of amino acid metabolism flux.

[0083] The aforementioned precision stems from the catalytic ability (k) of key enzymes in the amino acid synthesis pathway. cat The enzyme abundance and concentration constraints are explicitly integrated into the ecModel, effectively solving the flux overestimation problem caused by traditional chemometric models neglecting enzyme capacity limitations. In particular, at key nodes including the synthesis of branched-chain amino acids and aromatic amino acids, the model can accurately predict the balance of carbon and nitrogen flux allocation, demonstrating that the ecModel reflects the intrinsic physiological limitations of cells under different amino acid requirements. These results not only verify the predictive accuracy of this enzyme-constrained model in amino acid metabolic network analysis but also lay a solid theoretical foundation for the rational design of high-amino acid-producing strains under model guidance.

[0084] F. Changes in glucose and biomass over time as predicted by the model. like Figure 3 and Figure 4 As shown, during the 1-2 h batch culture simulation, the substrate glucose concentration showed a linear decreasing trend with increasing culture time: when the initial glucose concentration was 10 g / L, it decreased to 9.25 g / L after 1 h of culture; when the initial glucose concentration was 20 g / L, it decreased to 18.2 g / L after 2 h of culture. Correspondingly, the biomass concentration showed a linear increasing trend with increasing culture time: when the initial biomass concentration was 0.10 g / L, it increased to 0.15 g / L after 1 min of culture and to 0.225 g / L after 2 h of culture. The specific growth rate remained at 0.04 h in the 10 g / L glucose system. - ¹, The specific growth rate was maintained at 0.08 h in a 20 g / L glucose system. - ¹, the deviations between the simulation results and the actual shaker culture experimental data are all less than 10%; compared with the traditional static ecFBA model without dynamic constraints, the simulation results of this invention have a 35% better fit with the experimental data, which fully demonstrates the accuracy and reliability of this model.

[0085] G. Prediction of Genetic Engineering Modification Targets Using the constructed dynamic model, the following gene modification strategies were simulated, such as Figure 6 and Figure 7 As shown: (1) Overexpression of key enzymes in the ethanol synthesis pathway (Pdc and Adh) predicted an increase of 35% in ethanol production, which is highly consistent with the experimental results (an increase of 38%). (2) Knocking out GPD1, a key gene in the glycerol synthesis pathway, is predicted to increase ethanol production by 22% while reducing glycerol production by 85%; (3) The dynamic regulation strategy simulated the dynamic expression of the ethanol synthesis pathway under the control of the temperature-induced promoter, and predicted that the ethanol yield could be increased by 18% compared with the constitutive expression within a 24-hour culture period.

[0086] The above embodiments verify the effectiveness and applicability of the method of the present invention. By constructing a temporal neural network dynamic enzyme constraint model, the metabolic dynamics of unconventional yeast under different culture conditions can be accurately predicted, providing a powerful computational tool for the rational design of industrial strains and the optimization of fermentation processes.

[0087] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast, characterized in that, Includes the following steps: S1. Obtain a genome-scale metabolic network model of unconventional yeast, and construct the backbone structure of a static enzyme constraint model based on chemometrics matrix and gene-protein-reaction association information; S2. Integrate enzyme kinetic parameters and establish enzyme capacity constraint equations; S3. Construct a dual-timescale coupled framework; S4. Design an attention-enhanced long short-term memory network that uses historical metabolic flux, enzyme concentration, and substrate concentration as inputs to learn the time-dependent patterns of dynamic changes in enzyme concentration. S5. Establish a dynamic update equation for enzyme concentration, and couple the enzyme concentration changes predicted by the long short-term memory network with the flux constraints calculated by the enzyme constraint flux balance analysis. S6. Construct a loss function constrained by physical information and embed physical laws into the neural network training process; S7. An adaptive time step strategy is adopted to dynamically adjust the integral step size according to the rate of change of metabolic state; S8. Iterate and cycle until the culture termination time is reached, and output the dynamic trajectory of substrate concentration, biomass, and product concentration changes over time.

2. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The enzyme kinetic parameters in step S2 include the catalytic turnover number and the enzyme molecular weight. If experimental data is missing, a large-scale enzyme parameter prediction database and a deep learning model are used to supplement it.

3. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 2, characterized in that, The large-scale enzyme parameter prediction database predicts catalytic conversion numbers by fusing pre-trained protein models and molecular encoders. The deep learning model is based on a fusion architecture of graph neural networks and convolutional neural networks, using substrate molecular graphs and protein sequences as inputs to predict catalytic conversion numbers.

4. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The enzyme capacity constraint equation mentioned in step S2 is: ； Among them, v j Let k be the flux of reaction j. cat,j E represents the enzyme turnover number catalyzing reaction j. j denoted as , where is the concentration of the enzyme catalyzing reaction j.

5. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The dual-timescale coupling framework described in step S3 includes: Ordinary differential equation systems on macroscopic time scales: ； Where X is the biomass concentration, μ is the specific growth rate, and S i v represents the concentration of extracellular metabolite i. ex,i Let f be the exchange flux of metabolite i; Optimization of enzyme-constrained flux balance analysis at the microscale time scale, expressed as: ； Where max is the maximum value, c T is the coefficient row vector, and v is the metabolic flux column vector; The constraints for enzyme-constrained flux balance analysis include stoichiometry. Flux boundary constraints And enzyme capacity constraints: ； Where S is the stoichiometric matrix, c is the objective function coefficient vector, and v lb and v ub Let R be the lower and upper bounds of flux, j∈R i It is a summation operator, MW i It is the molecular weight of enzyme i, v j It is the metabolic flux of reaction j, k cat,ij σ is the transition number of enzyme i catalyzing reaction j. ij It is the saturation coefficient / occupancy coefficient of enzyme i catalyzing reaction j. This indicates the quality requirement of reaction j for enzyme i. This indicates the upper limit of available proteins for enzyme i.

6. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The attention mechanism of the attention-enhanced long short-term memory network described in step S4 dynamically adjusts the weights of information at different time steps by calculating the similarity between queries, keys, and values. ； in, It is the attention function, where Q refers to the query matrix, K refers to the key matrix, and V refers to the value matrix; QK T It is a dot product similarity matrix; It is the scaling factor; It is the scaled similarity matrix; softmax( ) is the softmax activation function.

7. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The enzyme concentration dynamic update equation mentioned in step S5 is as follows: ； Where, r j syn k represents the synthesis rate of enzyme j. deg,j U is the degradation rate constant, and μ is the specific growth rate; The enzyme synthesis rate was predicted by a long short-term memory network: ； Where t is time t; f LSTM θ is the Long Short-Term Memory (LSTM) network function; τ is the historical time; LTSM represents the set of all trainable parameters of the LSM network; v(t-τ) is the metabolic flux vector at time t-τ; e(t-τ) is the enzyme concentration / enzyme quantity vector at time t-τ; S(t-τ) is the metabolite concentration vector at time t-τ; θ LSTM This is the set of trainable parameters for the Long Short-Term Memory (LSTM) network.

8. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The physical information constraint loss function mentioned in step S6 is: ； in, Total loss function; Data fitting loss; Regularized weights are adjustable hyperparameters of the model. Stoichiometric constraint loss; Thermodynamic constraint loss; : Loss due to mass balance constraints.

9. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The expression for the adaptive time step strategy described in step S7 is: ； Among them, e local For local truncation error estimation, ε is the user-defined error tolerance, and α... max and α min To prevent drastic changes in step size, δ is a small constant to prevent division by zero; During periods of rapid metabolic change, the time step is automatically reduced to improve accuracy; during periods of relatively stable metabolic change, the time step is automatically increased to save computational resources.

10. The method for constructing a time-series neural network dynamic enzyme constraint model based on unconventional yeast according to claim 1, characterized in that, The iterative loop described in step S8 includes: S8.1 At the current time step t, based on the extracellular substrate concentration S(t), calculate the upper limit of substrate uptake flux using the Michaelis-Menten kinetic equation: ； Among them, v uptake (t) is the substrate uptake rate at time t; v max It is the maximum reaction rate; S(t) is the substrate concentration at time t; K m It is the Michaelis constant; S8.2 Substitute the uptake flux constraint into the enzyme constraint flux balance analysis model, solve the linear programming problem, and obtain the optimal flux distribution v(t) and specific growth rate μ(t); S8.3 Predicting enzyme synthesis rate r using long short-term memory networks syn (t); S8.4 uses a fourth-order Runge-Kutta integral ordinary differential equation system to update biomass X, substrate concentration S, and enzyme concentration e; S8.5 adaptively adjusts the time step Δt; S8.6 Update Time Repeat steps S8.1-S8.5 until the termination condition T is met.

Images