Metabolic engineering DBTL circulation intelligent optimization system based on machine learning

Abstract

The invention relates to the technical field of bioengineering, in particular to a metabolic engineering DBTL circulation intelligent optimization system based on machine learning. The system integrates a path design module based on a graph neural network, a hybrid modeling engine fusing a physical mechanism and deep learning, an analysis module based on a Shapley additive explanatory value and an adjustment module based on deep reinforcement learning and online Raman spectrum monitoring. The optimal synthesis path is predicted through heterogeneous graph convolution, high-precision metabolic flow prediction is achieved through a physical information neural network under the condition of data scarcity, a gene modification target is accurately positioned through Shapley additive interpretation analysis, and in the fermentation stage, a deep reinforcement learning agent is combined with real-time spectrum feedback, so that the optimal metabolic flow prediction is achieved. And a dynamic feeding and environment control strategy is autonomously learned and executed, so that the yield and the conversion rate of high-added-value chemicals are remarkably improved. According to the method, closed-loop optimization from strain design to intelligent control of the production process is realized, and the research and development cycle of a microbial cell factory is greatly shortened.

Description

Technical Field

[0001] This invention relates to the field of bioengineering technology, specifically to a machine learning-based intelligent optimization system for metabolic engineering DBTL cycles. Background Technology

[0002] Metabolic engineering aims to construct microbial cell factories capable of efficiently producing fuels, chemicals, and drugs by targeting and modifying cellular metabolic networks. The development of metabolic engineering generally follows an iterative cycle of "Design-Build-Test-Learn" (DBTL). In the "design" phase, researchers select targets based on biochemical knowledge; in the "build" phase, gene editing technology is used to modify strains; in the "test" phase, fermentation culture is conducted and the products are detected; and in the "learn" phase, patterns are extracted from the data to guide the next round of design.

[0003] However, cellular metabolic networks are extremely complex, involving thousands of metabolites, enzymes, and regulatory factors, and exhibiting non-linear interactions. Existing rational designs often rely on researchers' intuition or simplified biochemical models, resulting in high trial-and-error costs, and many designs fail to achieve the expected yields or productivity in actual testing.

[0004] To aid in design, researchers have developed various mathematical models. Genome-scale metabolic network models (GEMs) and flux balance analysis (FBA) are currently the most widely used tools. FBA, based on stoichiometry and homeostasis assumptions, predicts metabolic flux distribution by optimizing an objective function (such as maximizing biomass). However, FBA typically fails to capture dynamic changes in metabolic processes, enzyme kinetic limitations, and the effects of gene expression regulation.

[0005] In recent years, artificial intelligence and machine learning technologies have made groundbreaking progress in the application of biology. Deep learning models have been used to predict protein structure, enzyme function, and metabolic pathways. However, purely data-driven "black box" models face two major bottlenecks in metabolic engineering: Data scarcity: Compared to the hundreds of millions of datasets in image recognition or natural language processing, the cost of acquiring high-quality metabolomics and throughput omics data is extremely high, and the amount of data is often insufficient to train large-scale deep neural networks, which can easily lead to overfitting.

[0006] Poor interpretability: While deep learning models offer high prediction accuracy, they often fail to explain "why" a gene knockout increases yield. For biologists, predictions lacking mechanistic explanations are difficult to translate into concrete gene-editing strategies, limiting the application of AI in rational design.

[0007] In the "construction" and "testing" phases, the level of control over the fermentation process directly determines the performance of the bacterial strain. Traditional fermentation control mainly relies on PID controllers to maintain setpoints for temperature, pH, and dissolved oxygen. However, microbial fermentation is a nonlinear and time-varying dynamic process. Fixed setpoints often cannot adapt to the changing physiological needs of cells at different growth stages (such as lag phase, exponential growth phase, and stationary phase). Although model predictive control has been proposed for optimizing fermentation, its performance is highly dependent on the accuracy of the process model, and accurate biological process models are extremely difficult to obtain.

[0008] In summary, there is currently a lack of a system that can run through the entire DBTL process, deeply integrate biological mechanism knowledge with data-driven AI models, and achieve closed-loop optimization from upstream strain design to downstream fermentation control. Summary of the Invention

[0009] This invention provides a machine learning-based intelligent optimization system for metabolic engineering DBTL cycles, aiming to solve problems such as low prediction accuracy, strong data dependence, poor interpretability, and rigid fermentation control strategies in existing technologies. By integrating graph neural networks, physical information neural networks, interpretable AI analysis, and deep reinforcement learning, this invention achieves efficient design and real-time autonomous optimization of the production process of microbial cell factories.

[0010] The machine learning-based metabolic engineering DBTL cycle intelligent optimization system includes a data acquisition module, a data processing module, a pathway design module, a hybrid modeling module, and a regulation module. The data acquisition module is used to acquire spectral data of the culture medium in the bioreactor using a spectral sensor. The data processing module is used to receive the culture medium spectral data collected by the data acquisition module, and to estimate the state variables of the fermentation process in real time based on the data. The pathway design module is used to generate metabolic pathway design schemes using graph neural networks, and includes a graph construction unit and a feature embedding unit. The simulation module is used to simulate the metabolic pathway design scheme in order to simulate and predict the metabolic phenotype of the strain. The adjustment module is used to utilize a deep reinforcement learning agent to output control actions based on a preset reward function according to the state variables in order to adjust the operating parameters of the bioreactor.

[0011] Preferably, the graph construction unit is used to construct a heterogeneous graph containing metabolite nodes and enzyme nodes. The heterogeneous graph includes: metabolite nodes representing metabolic substrates, intermediates and products, and enzyme nodes representing catalytic enzymes or encoding genes. The edges in the heterogeneous graph include substrate-enzyme connection edges, enzyme-product connection edges representing the direction of biochemical reactions, and regulatory edges representing gene regulatory relationships.

[0012] Preferably, the feature embedding unit is used to: extract the molecular fingerprint or structural feature vector of the metabolite node using a pre-trained molecular characterization model; extract the amino acid sequence feature or functional feature vector of the enzyme node using a pre-trained protein language model; and use the structural feature vector and the sequence feature vector as node input features of the graph neural network.

[0013] Preferably, the graph neural network adopts a graph attention network architecture, which is configured to calculate the importance weights of neighbor nodes to the central node in the heterogeneous graph through a multi-head attention mechanism, and perform link prediction based on the aggregated node features to predict potential, unrecorded reaction associations between the metabolite nodes and the enzyme nodes, and output metabolic pathway design schemes.

[0014] Preferably, the simulation module includes a hybrid modeling engine, which is constructed as a physical information neural network architecture, including a parallel or serially coupled mechanism model layer and a data-driven layer; the mechanism model layer embeds a genome-scale metabolic network model and stoichiometry matrix of the target microorganism, used to describe the material conservation relationship of the cell metabolic network; the data-driven layer adopts a deep neural network or a recurrent neural network.

[0015] Preferably, the data-driven layer receives fermentation environment parameters and gene expression level data as input, and outputs estimated values ​​of metabolic reaction kinetic parameters or flux boundary constraints; the mechanism model layer uses the estimated values ​​of kinetic parameters or flux boundary constraints output by the data-driven layer to constrain the genome-wide metabolic network model and performs flux balance analysis to output the predicted results of genome-wide metabolic flux distribution.

[0016] Preferably, the training process of the physical information neural network employs a hybrid loss function, which includes: Data-driven loss used to minimize the error between model-predicted flux and experimentally measured flux; Physical constraint regularization terms are used to penalize predictions that violate biophysical laws. These physical constraints include mass conservation constraint residuals based on the stoichiometric matrix and thermodynamic feasibility constraints based on Gibbs free energy.

[0017] Preferably, it also includes an analysis module configured to calculate the additive interpretation value of each input feature in the simulation results, and to make modification suggestions for the generation of metabolic enzymes based on the sign and magnitude of the additive interpretation value.

[0018] A machine learning-based intelligent optimization method for metabolic engineering DBTL cycles includes the following steps: Design steps: Input the target product information into the pathway design module, use graph neural network to predict heterologous synthesis pathways and potential enzyme candidates, and generate candidate strain design schemes; Learning steps: Use the simulation module to simulate the metabolic flow of candidate strains and predict their phenotypes under different environments; Explanation steps: Calculate the Shapley additive interpretation value of each input feature in the simulation results using the analysis module, identify bottleneck enzymes and competitive branches that limit product synthesis, and generate optimized design schemes based on knockout or overexpression. Construction steps: Construct genetically engineered strains according to the optimized design scheme; Control steps: The constructed strain is inoculated into the bioreactor. During fermentation, the regulation module dynamically adjusts the feeding rate and environmental parameters based on real-time status feedback provided by the spectral sensor, using deep reinforcement learning to maximize the yield of the target product.

[0019] Compared with the prior art, the advantages of this invention are: By using physical information hybrid modeling and biochemical mechanisms to constrain the search space of neural networks, the overfitting problem of purely data-driven models with small sample sizes is effectively solved. At the same time, it makes up for the inaccuracy of parameters in purely mechanistic models and significantly improves the prediction accuracy of complex metabolic phenotypes.

[0020] By introducing Shapley additive interpretation value analysis, the complex model weights are transformed into a gene importance ranking that biologists can understand, directly guiding the direction of strain modification and avoiding blind trial and error.

[0021] The deep reinforcement learning-based regulation module enables the fermentation control system to move beyond simply maintaining a fixed setpoint. Instead, it can dynamically adjust culture conditions (such as staged feeding) based on the real-time metabolic state of the cells (e.g., through Raman spectroscopy), thereby optimizing the entire process trajectory and maximizing spatiotemporal yield. Attached Figure Description

[0022] Figure 1 This is a system architecture diagram of the metabolic engineering DBTL cycle intelligent optimization system based on machine learning proposed in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] The system of this invention is built on a distributed computing architecture, which mainly includes the following three layers: Data acquisition and management layer: This includes sensor arrays (pH, dissolved oxygen, temperature, Raman spectroscopy probes) connected to the bioreactor and interfaces to external biological databases.

[0025] Core algorithm computation layer: Deployed on high-performance computing clusters or cloud servers, it includes pathway design modules, simulation modules, analysis modules, and tuning modules.

[0026] Execution and Control Layer: This layer includes edge computing nodes and bioreactor hardware controllers (PLCs), which are responsible for receiving control commands from the algorithm layer in real time and driving the actuators (peristaltic pumps, gas flow meters, and stirring motors).

[0027] Based on the aforementioned distributed computing architecture, refer to Figure 1 The present invention describes the machine learning-based metabolic engineering DBTL cycle intelligent optimization system, which includes a data acquisition module, a data processing module, a pathway design module, a hybrid modeling module, an analysis module, and a regulation module.

[0028] The data acquisition module is used to collect spectral data of the culture medium inside the bioreactor using a spectral sensor.

[0029] The spectral sensor uses an immersion Raman probe excited at 785 nm, which is directly inserted into the fermenter liquid to avoid the risk of sampling contamination.

[0030] The data processing module is used to receive the culture medium spectral data collected by the data acquisition module and estimate the state variables of the fermentation process in real time based on the data.

[0031] The data processing module connects to the CCD detector of the Raman spectrometer via a Gigabit Ethernet interface or a USB 3.0 interface. The spectral acquisition integration time is set to 100-3000 ms (automatically adjusted based on signal strength), and sampling is triggered every 5 minutes.

[0032] The received raw Raman spectral data often contains fluorescence background, cosmic ray noise, and equipment thermal noise. The data processing module performs the following preprocessing: Cosmic ray removal: The Laplacian edge detection algorithm is used to identify and remove extreme peaks of single pixels, and interpolation is performed using the mean of neighboring pixels for repair.

[0033] Baseline correction: The broadband fluorescence background of the spectrum was fitted using the asymmetric least squares method and subtracted from the original spectrum, while retaining the characteristic Raman peaks.

[0034] Smoothing and Differentiation: The spectrum was smoothed by applying a Savitzky-Golay filter (window width = 15, polynomial order = 3) and the first derivative spectrum was calculated to enhance the resolution of overlapping peaks and eliminate additive baseline drift.

[0035] Standardization: Perform standard normal variable transformation on the full spectrum to eliminate proportional errors caused by sample scattering and optical path differences.

[0036] The preprocessed spectral data is input into a pre-trained deep learning model for state variable calculation. This embodiment employs a hybrid convolutional long short-term memory network (CNN-LSTM) model, specifically: Feature extraction: The CNN layer automatically extracts the position and intensity pattern of feature peaks in the spectrum and generates high-dimensional feature vectors.

[0037] Temporal modeling: The LSTM layer receives the spectral feature sequences of the current and past T time points to capture the dynamic evolution of the fermentation process (such as cell growth inertia).

[0038] Variable calculation: The fully connected output layer simultaneously regresses and predicts the main state variables at the current time step: Glucose concentration: based on vibrational peak characteristics of COC and CH bonds.

[0039] Biomass concentration: based on protein amide band and nucleic acid peak characteristics.

[0040] Target product concentration: based on the characteristics of product-specific functional groups (such as the carboxyl group of succinic acid or the lactone ring of artemisinic acid).

[0041] The pathway design module is used to generate metabolic pathway design schemes using graph neural networks.

[0042] The graph neural network dynamically learns the contribution of different substructures within a molecule to specific metabolic functions (such as enzyme binding sites and reactive sites) through an attention mechanism.

[0043] The graph building unit first extracts metabolic response information from the KEGG and MetaCyc databases to construct a directional isomorphic graph.

[0044] The nodes in this directed heterogeneous graph mainly include metabolite nodes representing metabolic substrates, intermediates, and products, and enzyme nodes representing catalytic enzymes or encoding genes.

[0045] The edges in this directed isomorphic graph mainly include substrate-enzyme linkage edges, enzyme-product linkage edges, and regulatory edges, which represent the direction of biochemical reactions and gene regulatory relationships.

[0046] For metabolite nodes, the system calls the RDKit toolkit to convert the SMILES molecular formula into a molecular fingerprint and inputs it into the pre-trained ChemBERTa model to extract a 128-dimensional structural feature vector.

[0047] For enzyme nodes, the system obtains their amino acid sequences and uses the ESM-1b protein language model to extract 256-dimensional sequence and functional feature vectors.

[0048] Based on the constructed directed heterogeneous graph and the feature vectors of each node, the graph attention network predicts and generates metabolic pathways.

[0049] First, the features of the input nodes Perform linear transformations to enhance feature representation capabilities: ; In the formula, It is a learnable weight matrix.

[0050] Next, compute nodes Its neighboring nodes Attention coefficient between This coefficient measures the nodes For nodes Importance: ; In the formula, This represents a vector concatenation operation. It is the weight vector of a single-layer feedforward neural network. The LeakyReLU activation function is used to introduce non-linearity, enabling the model to handle complex negative correlations.

[0051] To ensure the comparability of attention coefficients across different nodes, a Softmax function is used to normalize them, yielding the final attention weights. : ; In the formula, M is a node. The set of neighboring nodes.

[0052] Finally, the calculated weights are used to aggregate the features of neighboring nodes to generate nodes. Update feature vector : ; In the formula, It is a non-linear activation function (such as ReLU).

[0053] Through multilayer aggregation, the model learns long-range dependencies (such as feedback inhibition) in metabolic networks. Finally, a fully connected layer is used for link prediction to predict whether there is a potential enzymatic catalytic pathway between a given precursor and the target product.

[0054] The simulation module is used to simulate the metabolic pathway design scheme generated by the pathway design module in order to simulate and predict the metabolic phenotype of the strain.

[0055] The module includes a hybrid modeling engine built as a physical information neural network architecture. This architecture not only leverages the powerful nonlinear fitting capabilities of deep learning to process high-dimensional omics data, but also addresses the problems of poor generalization ability and violation of biological common sense in traditional "black box" models with small sample data by embedding rigid biochemical physical laws (mass conservation and thermodynamic laws).

[0056] The physical information neural network architecture consists of two coupled parts: a data-driven dynamics prediction subnetwork and a physical mechanism constraint layer. The specific structure is as follows: Data-driven dynamics prediction subnet: Input layer: Receives two types of feature vectors: Environmental state vector: includes real-time temperature, pH, dissolved oxygen concentration, substrate concentration (e.g., glucose), and inducer concentration within the fermenter.

[0057] Biological state vectors include gene expression levels (transcriptome data) or enzyme protein abundance (proteome data) of key metabolic enzymes.

[0058] Hidden layer: Employs multilayer perceptrons or long short-term memory networks to capture the nonlinear effects of environmental and gene regulation on enzyme activity.

[0059] Output layer: Outputs the apparent kinetic parameters of key reactions in the metabolic network. In this embodiment, the output is the upper and lower bounds of reaction flux or the effective turnover number of the enzyme.

[0060] Physical mechanism constraint layer: This layer is a differentiable operator layer composed of mathematical equations describing cellular metabolic networks.

[0061] This layer contains a built-in stoichiometric matrix of the target microorganism. ,in For the amount of metabolites, This represents the amount of reaction.

[0062] This layer receives parameters from the kinetic prediction subnet and, in conjunction with the Michaelis equation or the law of mass action, calculates the preliminary metabolic flux distribution.

[0063] To train the aforementioned network, this system constructs a composite loss function that includes a data fitting term and a physical regularization term. The function is minimized using the backpropagation algorithm: ; in, This is a data-driven loss term used to quantify the error between the model-predicted flux and the experimentally measured flux: ; In the formula, The number of experimental samples, For the first Predicted flux for each experimental sample For the first The experimental throughput of a sample.

[0064] As a mass conservation constraint, based on the metabolic quasi-steady-state assumption, the rate of formation of intracellular intermediate metabolites should be equal to the rate of consumption. This penalty violates the predictions of Kirchhoff's flow law. ; In the formula, Stoichiometric matrix and Let be the dimension of the stoichiometric matrix. Indicates the first Metabolites in each reactant stoichiometric coefficients The model predicts the first The flux of each reaction.

[0065] By including this term, if the predicted flux leads to the creation or disappearance of certain intermediate metabolites out of thin air (i.e., If this happens, a huge penalty value will be generated.

[0066] As a thermodynamic feasibility constraint, this embodiment introduces a penalty term based on Gibbs free energy to avoid the model predicting results that violate thermodynamics.

[0067] First, the reaction is determined by querying the database. The standard Gibbs free energy change.

[0068] Subsequently, the reaction was calculated based on the standard Gibbs free energy change. Actual Gibbs free energy change: ; In the formula, For the reaction The actual Gibbs free energy change, For the reaction Standard Gibbs free energy change, Let be the ideal gas constant. Absolute temperature Metabolites Intracellular concentration.

[0069] Finally, a thermodynamic loss is constructed, and the ReLU function is used to penalize reactions whose flux direction is opposite to the direction of the free energy driving force: ; when and When the signs are opposite (i.e., the reaction follows the thermodynamic direction), the product is negative, the ReLU output is 0, and there is no penalty; when the signs are the same (violating thermodynamics), a positive loss is generated, and the gradient backpropagates to correct the network parameters.

[0070] , and These are the corresponding weights, used to balance the magnitudes of the various loss terms.

[0071] In a preferred embodiment, to address the scarcity of high-quality biological data, the hybrid modeling engine also integrates an active learning strategy, specifically: During the prediction phase, Monte Carlo Dropout is used to enable random deactivation during network inference and perform multiple forward propagations. The variance of the predicted flux is calculated and used as a measure of the model's uncertainty.

[0072] These highly uncertain experimental conditions are recommended for automated wet laboratory platforms or researchers. After completing a new fermentation experiment, the newly acquired data is incorporated into the training set to fine-tune the physical information neural network model, thereby achieving continuous self-evolution of the model's accuracy.

[0073] Through the above implementation methods, the hybrid modeling engine can accurately predict the distribution of metabolic flux at the whole genome scale with only a small amount of labeled experimental data, and effectively reduce prediction errors.

[0074] The analysis module is configured to calculate the Shapley additive interpretation value of each input feature in the simulation results, and to rank the importance of metabolic enzymes according to the sign and magnitude of the Shapley additive interpretation value, thereby identifying key targets for gene editing.

[0075] For the target variable output by the above physical information neural network model, the Kernel SHAP algorithm is used to calculate the Shapley additive explanation value of each input feature (such as the expression level or flux of a specific enzyme), which represents the average marginal contribution of the input feature to the product synthesis rate.

[0076] Then, specific modification suggestions are generated based on the Shapley additive interpretation value: Positive key target: If the SHAP value of an enzyme is significantly positive and its current expression level is not saturated, it is recommended to overexpress the gene.

[0077] Negative bottleneck targets: If the SHAP value of an enzyme is significantly negative, it indicates that the branch has diverted the carbon source or produced an inhibitor. It is recommended to knock out the gene or use CRISPRi for transcriptional repression.

[0078] The adjustment module is used to utilize a deep reinforcement learning agent to output control actions based on a preset reward function according to the state variables in order to adjust the operating parameters of the bioreactor.

[0079] Among them, the deep reinforcement learning agent includes: State space: including the current time Biomass substrate concentration Product concentration Byproduct concentration pH value, DO value, and the action at the previous moment. Among them, biomass substrate concentration Product concentration Byproduct concentration Obtained from the data processing module.

[0080] Action space: a continuous variable vector, including feed pump flow rate, stirring speed, aeration rate, and inducer addition rate.

[0081] reward function Defined as the incremental economic benefit at each step: ; In the formula, , and For the weights of each item, The number of reward products increased. Penalty for byproduct accumulation (the squared term represents a severe penalty for high concentrations of byproducts). The punishment action involves violent fluctuations to protect the equipment.

[0082] In view of the characteristics of continuous action space, high-dimensional state and extremely high stability requirements in the fermentation process, this implementation method uses the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to train the agent, which includes a policy network and a value network.

[0083] The policy network receives the state as input and outputs the action corresponding to that state.

[0084] The value network receives the state and corresponding action as input and outputs the Q-value (expected return).

[0085] The TD3 algorithm uses two value networks to reduce the overestimation bias of the Q value.

[0086] In actual training, deep reinforcement learning is first pre-trained in a "digital twin" simulation environment built by the simulation module to learn basic control strategies, avoiding exploration in a real reactor. Then, the pre-trained model is deployed to a real controller, utilizing online learning or small-batch offline fine-tuning to adapt to the differences in the real environment.

[0087] A machine learning-based intelligent optimization method for metabolic engineering DBTL cycles includes the following steps: Design steps: Input the target product information into the pathway design module, use graph neural network to predict heterologous synthesis pathways and potential enzyme candidates, and generate candidate strain design schemes; Learning steps: Use the simulation module to simulate the metabolic flow of candidate strains and predict their phenotypes under different environments; Explanation steps: Calculate the Shapley additive interpretation value of each input feature in the simulation results using the analysis module, identify bottleneck enzymes and competitive branches that limit product synthesis, and generate optimized design schemes based on knockout or overexpression. Construction steps: Construct genetically engineered strains according to the optimized design scheme; Control steps: The constructed strain is inoculated into the bioreactor. During fermentation, the regulation module dynamically adjusts the feeding rate and environmental parameters based on real-time status feedback provided by the spectral sensor, using deep reinforcement learning to maximize the yield of the target product.

[0088] Example 1 This embodiment further illustrates the scheme through optimization of succinic acid production by *E. coli*. Specifically, it includes: Design and Prediction: The pathway design module was used to analyze the E. coli metabolic network. The analysis module, through sapride additive interpretation value analysis, indicated that: pflB (pyruvate-formate lyase) and ldhA (lactate dehydrogenase) have extremely high negative sapride explanatory values, indicating that they are major pathways for carbon flux loss and should be knocked out.

[0089] ppc (PEP carboxylase) has a positive sapride additive explanatory value and overexpression is recommended.

[0090] The physical information neural network predicts that knocking out only pflB and ldhA will lead to NADH accumulation and inhibit cell growth. The model suggests introducing exogenous fdh (formate dehydrogenase) through its link prediction function. This not only eliminates the byproduct formic acid but also regenerates NAD+ into NADH, thereby balancing the redox potential. This is a synergistic effect that is easily overlooked in traditional rational design.

[0091] Strain construction: Based on the above design, engineered strains were constructed: Knocking out ptsG is intended to slow down the rate of glucose uptake and reduce extravasation metabolism.

[0092] Fermentation control: The strain was inoculated into a 5L fermenter. The deep reinforcement learning agent of the regulation module monitored glucose and acetic acid concentrations in real time using Raman spectroscopy.

[0093] During this process, the agent learned a non-linear "pulse-like" feeding strategy. When Raman spectroscopy detected that the acetic acid concentration began to rise due to overflow metabolism (>2 g / L), the agent automatically reduced the feeding rate, forcing the cells to utilize acetic acid; when the acetic acid concentration decreased, it quickly resumed high feeding to maintain high productivity.

[0094] Ultimately, compared to traditional constant-rate feeding or exponential feeding, the fermentation cycle under intelligent control was shortened by 20%, and the succinic acid yield increased from 65 g / L to 98.2 g / L. Due to the minimal byproducts, the subsequent separation and purification costs were reduced by 40%.

[0095] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0096] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A machine learning-based intelligent optimization system for metabolic engineering DBTL cycles, characterized in that, It includes a data acquisition module, a data processing module, a pathway design module, a hybrid modeling module, and a regulation module; The data acquisition module is used to acquire spectral data of the culture medium in the bioreactor using a spectral sensor. The data processing module is used to receive the culture medium spectral data collected by the data acquisition module, and to estimate the state variables of the fermentation process in real time based on the data. The pathway design module is used to generate metabolic pathway design schemes using graph neural networks, and includes a graph construction unit and a feature embedding unit. The simulation module is used to simulate the metabolic pathway design scheme in order to simulate and predict the metabolic phenotype of the strain. The adjustment module is used to utilize a deep reinforcement learning agent to output control actions based on a preset reward function according to the state variables in order to adjust the operating parameters of the bioreactor.

2. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, The graph construction unit is used to construct a heterogeneous graph containing metabolite nodes and enzyme nodes. The heterogeneous graph includes: metabolite nodes representing metabolic substrates, intermediates and products, and enzyme nodes representing catalytic enzymes or encoding genes. The edges in the heterogeneous graph include substrate-enzyme connection edges, enzyme-product connection edges representing the direction of biochemical reactions, and regulatory edges representing gene regulatory relationships.

3. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, The feature embedding unit is used to: extract the molecular fingerprint or structural feature vector of the metabolite node using a pre-trained molecular characterization model; extract the amino acid sequence feature or functional feature vector of the enzyme node using a pre-trained protein language model; and use the structural feature vector and the sequence feature vector as node input features of the graph neural network.

4. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, The graph neural network adopts a graph attention network architecture, which is configured to calculate the importance weights of neighbor nodes to the central node in the heterogeneous graph through a multi-head attention mechanism, and perform link prediction based on the aggregated node features to predict potential, unrecorded reaction associations between the metabolite nodes and the enzyme nodes, and output metabolic pathway design schemes.

5. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, The simulation module includes a hybrid modeling engine, which is constructed as a physical information neural network architecture, comprising a parallel or serially coupled mechanism model layer and a data-driven layer. The mechanism model layer embeds a genome-scale metabolic network model and stoichiometry matrix of the target microorganism to describe the material conservation relationships of the cell metabolic network. The data-driven layer employs a deep neural network or a recurrent neural network.

6. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 5, characterized in that, The data-driven layer receives fermentation environment parameters and gene expression level data as input, and outputs estimated values ​​of metabolic reaction kinetic parameters or flux boundary constraints. The mechanism model layer uses the estimated values ​​of kinetic parameters or flux boundary constraints output by the data-driven layer to constrain the genome-scale metabolic network model and performs flux balance analysis to output the predicted results of genome-wide metabolic flux distribution.

7. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, The training process of the physical information neural network employs a hybrid loss function, which includes: Data-driven loss used to minimize the error between model-predicted flux and experimentally measured flux; Physical constraint regularization terms are used to penalize predictions that violate biophysical laws. These physical constraints include mass conservation constraint residuals based on the stoichiometric matrix and thermodynamic feasibility constraints based on Gibbs free energy.

8. The intelligent optimization system for metabolic engineering DBTL cycles based on machine learning according to claim 1, characterized in that, It also includes an analysis module configured to calculate the Shapley additive interpretation value of each input feature in the simulation results, and to make modification suggestions for the generation of metabolic enzymes based on the sign and magnitude of the Shapley additive interpretation value.

9. A machine learning-based intelligent optimization method for metabolic engineering DBTL cycles, based on the machine learning-based intelligent optimization system for metabolic engineering DBTL cycles as described in claims 1-8, characterized in that, Includes the following steps: Design steps: Input the target product information into the pathway design module, use graph neural network to predict heterologous synthesis pathways and potential enzyme candidates, and generate candidate strain design schemes; Learning steps: Use the simulation module to simulate the metabolic flow of candidate strains and predict their phenotypes under different environments; Explanation steps: Calculate the Shapley additive interpretation value of each input feature in the simulation results using the analysis module, identify bottleneck enzymes and competitive branches that limit product synthesis, and generate optimized design schemes based on knockout or overexpression. Construction steps: Construct genetically engineered strains according to the optimized design scheme; Control steps: The constructed strain is inoculated into the bioreactor. During fermentation, the regulation module dynamically adjusts the feeding rate and environmental parameters based on real-time status feedback provided by the spectral sensor, using deep reinforcement learning to maximize the yield of the target product.

Images