Intelligent iterative optimization method and system for culture medium formula based on deep learning

Abstract

The present application relates to the technical field of culture medium formula data processing, in particular to a culture medium formula intelligent iterative optimization method and system based on deep learning. The method comprises the following steps: obtaining a formula response feedback pair; converting formula component information to obtain a functional state node; obtaining a stage response improvement amount when only the main functional state changes, and a stage response improvement amount when the main functional state and the collaborative functional state change together; determining a component functional relationship edge according to the difference between the two improvement amounts; retaining the positive component functional relationship edge in the component functional relationship edge and determining a component functional collaborative response feature; calculating a stage response offset feature according to the response drop amount; calculating the executable degree of a candidate formula, and optimizing and adjusting all candidate formulas according to the executable degree. The present application can improve the stability, executability and iterative optimization efficiency of candidate formula recommendation.

Description

Technical Field

[0001] This invention relates to the field of culture medium formulation data processing technology, specifically to a method and system for intelligent iterative optimization of culture medium formulations based on deep learning. Background Technology

[0002] Culture medium formulation optimization is widely used in microbial fermentation, cell culture, drug screening, enzyme production, and biomaterial preparation. Culture media typically include carbon sources, nitrogen sources, inorganic salts, trace elements, buffer systems, osmotic pressure regulating components, and other inducing components. These components collectively influence cell growth rate, metabolic stability, accumulation of target products, cell viability, and byproduct formation. In practical experiments, researchers usually adjust the formulation based on the results of the previous culture round and observe changes in the next culture round, gradually approaching an optimal formulation through multiple iterations.

[0003] Currently, commonly used methods for optimizing culture medium formulations include manual adjustment based on experience, single-factor experiments, orthogonal experiments, response surface methodology, and conventional machine learning prediction methods. While these methods can screen out some effective components, culture medium formulations are not simply parameter tables; changes in a single component do not necessarily produce independent, stable effects. For example, increasing the carbon source may lead to enhanced early growth but insufficient later product yield if the nitrogen source cannot effectively absorb it; increasing micronutrients may result in product gain if the target metabolic stage is not activated; and insufficient buffering systems may accelerate carbon source metabolism, potentially causing pH fluctuations and decreased cell viability. Therefore, focusing solely on the correlation between a single component and yield is insufficient to fully grasp the synergistic optimization principles of culture medium formulations.

[0004] In existing techniques for applying deep learning to culture medium formulation optimization, a common approach is to directly input the concentrations of each component as parallel fields into fully connected layers, embedded layers, or ordinary feature encoding layers, and use the final yield as the output target. While this method can learn the correlation between some component concentrations and yields, it is prone to misclassifying surface correlations of single components as effective influencing factors. The key to synergistic responses in culture media lies in the dynamic changes in the relationships between components as the formulation is adjusted. If existing deep learning models fail to perceive these changes before feature extraction, they struggle to determine whether a particular functional supply must change in conjunction with another functional supply to produce a stable improvement in culture performance. Summary of the Invention

[0005] This invention provides a method and system for intelligent iterative optimization of culture medium formulations based on deep learning, in order to solve existing problems.

[0006] The intelligent iterative optimization method for culture medium formulation based on deep learning of the present invention adopts the following technical solution: One embodiment of the present invention provides a method for intelligent iterative optimization of culture medium formulations based on deep learning, the method comprising the following steps: Obtain formulation response feedback pairs, which are obtained by pairing experimental data from two adjacent experiments on the same culture subject. The experimental data include formulation component information, culture stage response information, and formulation constraint information. The formula component information is converted according to its equivalent contribution in the culture function category to obtain the functional state node; Take any functional state node as the main function and any other functional state node as the coordinating function. Obtain the stage response improvement amount when only the main function state changes in the cultivation stage response information, and the stage response improvement amount when both the main function state and the coordinating function state change. Based on the difference between the two improvement amounts, determine the component function relationship edge between the main function and the coordinating function. Retain the positive component functional relationship edges in the component functional relationship edges, and determine the component functional collaborative response characteristics corresponding to each positive component functional relationship edge based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects; The culture stage response information of the formula response feedback alignment experimental data is sorted according to the time sequence of the culture stages, and the response drop of the next culture stage is calculated in turn. The stage response offset characteristics are calculated based on the response drop. For any candidate formulation, statistical analysis is performed on the formulation constraint information to obtain the proportion of formulation constraint satisfaction. Based on the component functional synergistic response characteristics, stage response offset characteristics, and formulation constraint satisfaction proportion of the candidate formulation, the feasibility of the candidate formulation is calculated, and all candidate formulations are optimized and adjusted according to the feasibility.

[0007] Furthermore, the formulation component information is converted according to its equivalent contribution in the culture functional category to obtain functional state nodes, specifically including: For any culture function category, obtain the concentrations of all components in the formulation component information, where the formulation component information includes the concentration of each component; The product of the concentration of each component and the equivalent conversion coefficient corresponding to that component is determined as the component conversion supply amount for that culture functional category. Sum the converted supply quantities of all components to obtain the total converted supply quantity for this culture function category; The ratio of the total converted supply to the reference upper limit of the cultivation function category is normalized to obtain the functional state node corresponding to the cultivation function category.

[0008] Furthermore, based on the difference between the two improvement quantities, the component functional relationship edge between the main function and the synergistic function is determined, specifically including: Calculate the difference between the improvement in stage response when both the main functional state and the collaborative functional state change and the improvement in stage response when the main functional state changes. If the difference is positive, the direction of the component function relationship edge between the main function and the synergistic function is determined to be strengthening; if the difference is negative, the direction of the component function relationship edge is determined to be weakening. The absolute value of this difference is used to determine the strength of the functional relationship edge of the component.

[0009] Furthermore, positive component function relationship edges in the component function relationship edges are retained, specifically including: For any component functional relationship edge, the component functional relationship edge whose direction is strengthening and whose strength is greater than a preset threshold is determined as a positive component functional relationship edge. Preserve the positive component functional relationship edges.

[0010] Furthermore, based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects, the component functional collaborative response characteristics corresponding to each positive component functional relationship edge are determined, specifically including: Multiply the strength of the positive component functional relationship edge by the values ​​of the two functional state nodes it connects to obtain the component functional collaborative response feature corresponding to the positive component functional relationship edge; The strength of the positive component functional relationship edge represents the degree of strength of that positive component functional relationship edge.

[0011] Furthermore, the culture stage response information of the formulation response feedback alignment experimental data is sorted according to the time sequence of the culture stages, and the response decline between the later and earlier culture stages is calculated sequentially. Based on the response decline, the stage response shift characteristics are calculated, specifically including: The response information during the cultivation phase is sorted according to the early, middle, and late stages of the cultivation phase and the outcome. The difference between the response information of the next cultivation stage and the response information of the previous cultivation stage is calculated sequentially to obtain the response fall-off amount. If the response information of the next cultivation stage is less than that of the previous cultivation stage, the response fall-off amount is the difference between the two; otherwise, the response fall-off amount is set to zero. The stage response offset characteristics are obtained by summing up all the response fallback values ​​and normalizing them.

[0012] Furthermore, statistical analysis of the formulation constraint information yields the proportion of formulation constraints satisfied, specifically including: The number of formulation constraints satisfied by candidate formulations and the total number of preset formulation constraints are counted. The ratio of the number of satisfied configuration constraints to the total number of preset configuration constraints is determined as the configuration constraint satisfaction ratio.

[0013] Furthermore, based on the synergistic response characteristics of the components, the stage response shift characteristics, and the proportion of formulation constraint satisfaction of the candidate formulations, the feasibility of the candidate formulations is calculated, specifically including: The exponential decay factor is calculated based on the stage response offset characteristics, where the exponential decay factor is negatively correlated with the stage response offset characteristics; The feasibility of a candidate formulation is determined by multiplying the synergistic response characteristics of its components, the exponential decay factor, and the proportion of formulation constraints.

[0014] Furthermore, all candidate formulations are optimized and adjusted based on feasibility, specifically including: The candidate formulations are sorted according to their feasibility, and the candidate formulation with the highest feasibility is selected as the recommended formulation. The functional state node, component functional relationship edge, stage response offset feature and formulation constraint information of the recommended formulation are output.

[0015] This invention proposes a deep learning-based intelligent iterative optimization system for culture medium formulations, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the deep learning-based intelligent iterative optimization method for culture medium formulations.

[0016] The beneficial effects of the technical solution of the present invention are: This invention converts component concentrations into functional state nodes based on their cultivation effects. Then, it constructs functional relationship edges between components through feedback differences from adjacent experiments, effectively identifying synergistic, weakening, or reversing relationships between functional states. Based on this, feature interactions are performed only on functional states supported by positive relationship edges to obtain synergistic response characteristics of component functions. Simultaneously, stage response shift characteristics are obtained by analyzing the response sequences of the cultivation stages. The feasibility of candidate formulations is calculated by combining formulation constraints with proportional calculations, and finally, the candidate formulations are optimized and adjusted based on their feasibility. This approach reduces the interference of single-component surface correlations on model training, making it easier for the model to identify the true synergistic relationships between carbon and nitrogen source connections, buffer systems and metabolic stability, and trace elements and product stage responses. It avoids recommending infeasible formulations that result in excessively high early growth but insufficient or unfeasible later product development, thus improving the stability, feasibility, and iterative optimization efficiency of candidate formulation recommendations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a deep learning-based intelligent iterative optimization method for culture medium formulations, as provided in one embodiment of the present invention; Figure 2 This is a structural diagram of a deep learning-based intelligent iterative optimization system for culture medium formulations, provided in one embodiment of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the deep learning-based intelligent iterative optimization method for culture medium formulations proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent iterative optimization method for culture medium formulation based on deep learning provided by this invention.

[0022] This invention provides a method and system for intelligent iterative optimization of culture medium formulations based on deep learning. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of a deep learning-based intelligent iterative optimization method for culture medium formulations according to an embodiment of the present invention. The method includes the following steps: S101. Obtain formulation response feedback pairs, wherein the formulation response feedback pairs are obtained by pairing two adjacent experimental data of the same culture object. The experimental data include formulation component information, culture stage response information and formulation constraint information.

[0023] In this embodiment, historical culture experiment records and formula records are first obtained.

[0024] Formulation records include the following information: component name, component concentration, batch number, solvent system, sterilization method, initial pH, osmotic pressure, cost, and formulation limitations. Culture experiment records include the following information: bacterial strain or cell line, inoculum size, culture temperature, rotation speed, aeration rate, culture duration, sampling time, cell density, substrate consumption, metabolites, target product, cell viability, and byproduct indicators.

[0025] Based on the above information, a uniform sample was established according to experimental batches. For batches with multiple sampling points, samples were arranged in ascending order of sampling time. For formulation records showing different units for the same component, these were uniformly converted to mass concentration or molar concentration. For target product detection values ​​from different detection methods, the detection method label was retained, and comparisons were only made between results using the same detection method or calibrated results. Records lacking key component concentrations, incubation stage detection values, or target product endpoint values ​​were not used to form formulation iteration samples but were only saved as background statistical records.

[0026] Based on the above operations, each experimental sample includes at least the following three types of information: The first category is formulation component information, which records the names of the culture medium components used in the sample and the concentration of each component.

[0027] The second category is culture stage response information, which is used to record changes in early growth, mid-stage metabolism, late-stage products, cell viability, and byproduct risk.

[0028] The third category is formulation constraint information, which records the feasibility of the formulation in terms of solubility, sterilization stability, pH range, osmotic pressure range, and cost limitations.

[0029] Grouping by the same bacterial species or cell line, and sorting by experimental round or time, pairing adjacent experimental data for the same culture subject to form a formulation-response feedback pair. This formulation-response feedback pair represents the actual impact of adjusting the formulation function on the response during the culture stage, assuming the culture subjects are the same or comparable.

[0030] For example, suppose an optimization is aimed at the fermentation of *E. coli* to produce recombinant proteins. In the first round of experiments, the formulation consisted of 10 g / L glucose, 5 g / L peptone, 2 g / L dipotassium hydrogen phosphate, 0.5 g / L magnesium sulfate, and 10 mM HEPES buffer. The culture phase responses were: early growth OD600 of 2.5, mid-stage substrate consumption rate of 60%, late-stage target protein yield of 0.8 g / L, and cell viability of 85%. Regarding formulation constraints, solubility, sterilization stability, pH range, osmotic pressure, and cost all met the requirements.

[0031] In the second round of experiments, only the glucose concentration was increased to 15 g / L, while the other components remained unchanged. The culture phase responses were as follows: early growth OD600 increased to 3.8, but mid-stage substrate consumption decreased to 45%, and late-stage target protein yield decreased to 0.5 g / L, with cell viability dropping to 70%. Osmolarity increased from 320 mOsm / kg to 380 mOsm / kg, remaining within the cells' tolerance range, and other formulation constraints were still met.

[0032] The first and second rounds of experiments were paired to form a formulation-response feedback pair. In this feedback pair, only the functional state of the carbon source changed significantly, with the stage response showing early improvement but subsequent decline, and the improvement amount of the stage response being negative.

[0033] Through the above method, the system obtains several sets of formula response feedback pairs. Each set of formula response feedback pairs contains formula component information, culture stage response information, and formulation constraint information, which are used for the construction of functional state nodes and the extraction of relationship-aware features in subsequent steps.

[0034] S102. Convert the formula component information according to its equivalent contribution in the culture function category to obtain the functional state node.

[0035] In this embodiment, the formulation component information is converted according to its equivalent contribution in the culture functional category to obtain the functional state node, specifically including: For any culture function category, obtain the concentrations of all components in the formulation component information, where the formulation component information includes the concentration of each component; The product of the concentration of each component and the equivalent conversion coefficient corresponding to that component is determined as the component conversion supply amount for that culture functional category. Sum the converted supply quantities of all components to obtain the total converted supply quantity for this culture function category; The ratio of the total converted supply to the reference upper limit of the cultivation function category is normalized to obtain the functional state node corresponding to the cultivation function category.

[0036] For example, the information of the formulation components is converted according to their equivalent contribution in the culture function category to obtain the functional state node. Specific components in the culture medium (such as glucose, glycerol, peptone, yeast extract, phosphate, magnesium salts, iron salts, buffers, etc.) contribute to culture function categories (such as carbon source supply, nitrogen source supply, buffer stabilization, trace element activation, osmotic pressure restriction, etc.) in different ways and cannot be directly compared. Different components in the same formulation may contribute to the same culture function category; for example, glucose and glycerol both contribute to carbon source supply. The same component may also contribute to multiple culture function categories; for example, phosphate participates in both buffer stabilization and osmotic pressure regulation. Therefore, it is necessary to first classify the specific components according to the culture function category and then uniformly convert them to the same functional dimension.

[0037] Specifically, it includes the following sub-steps: For any culture function category (i.e., the first (Class function set), first obtain the concentrations of all components in the formulation component information. Let the first... In the first formulation experimental sample, the first The concentration of each component is Based on the function of the components, this component is classified into the first category. Culture function category (i.e., the set of components belonging to this culture function category) For example, glucose and glycerol are classified under the category of carbon source supply in culture, while peptone and yeast extract are classified under the category of nitrogen source supply in culture.

[0038] The product of each component concentration and its corresponding equivalent conversion factor is used to determine the component's equivalent supply amount for that culture function category. Let the first component be... The component is related to the first The equivalent conversion coefficient of the cultivation function category is This coefficient is determined based on the actual contribution of each component to the functional category of the culture: carbon source supply category is converted according to carbon molar equivalent, nitrogen source supply category according to nitrogen molar equivalent, buffer stabilization category according to buffer capacity, and osmotic pressure limitation category according to osmotic pressure contribution. Therefore, the formula for calculating the converted supply amount of each component is: By uniformly converting the original concentrations of different components to the same functional dimension, the contributions of different components to the same culture function category can be added together for comparison. For example, if 1g of glucose provides 1 unit of carbon source equivalent and 1g of glycerol provides 1.2 units of carbon source equivalent, then the carbon source contribution of 5g of glycerol is equivalent to the carbon source contribution of 6g of glucose.

[0039] The total converted supply of all components is obtained by summing the converted supply quantities of all components for this culture function category. The calculation formula is: ; Total converted supply The physical meaning is: this formula in the first... The total supply level of a culture function category reflects the overall degree to which that culture function category is satisfied. For example, the larger the total equivalent supply of carbon sources, the more abundant the available carbon sources provided by the formula.

[0040] Get the Reference upper limit for culture-like functional categories within the allowable range of historical samples or processes. This reference upper limit represents the safety boundary or process limit for the supply of this culture functional category; exceeding this limit may lead to cell growth inhibition or metabolic abnormalities. The total converted supply... Divide by the upper limit of the reference range The supply ratio is obtained and then normalized. And truncate to a comparable interval to obtain the functional state node. The calculation formula is: ; Functional status node The physical meaning is: this formula in the first... The normalization of the supply level of a class of cultivation functions relative to a reference upper limit, typically ranging from 0 to 1. The larger the value, the better the formula is in the [number of]th [year]. The stronger the supply in terms of the functional category of cultivation; The smaller the value, the weaker the supply of that culture function category. Unifying the supply levels of different culture function categories to the same numerical range enables subsequent steps to perform interactive calculations and comparisons of different functional state nodes.

[0041] S103. Take any functional state node as the main function and any other functional state node as the cooperating function, obtain the stage response improvement amount when only the main function state changes in the cultivation stage response information, and the stage response improvement amount when both the main function state and the cooperating function state change together. Based on the difference between the two improvement amounts, determine the component function relationship edge between the main function and the cooperating function.

[0042] In this embodiment, the component functional relationship edge between the main function and the collaborative function is determined based on the difference between the two improvement quantities, specifically including: Calculate the difference between the improvement in stage response when both the main functional state and the collaborative functional state change and the improvement in stage response when the main functional state changes. If the difference is positive, the direction of the component function relationship edge between the main function and the collaborative function is determined to be strengthening; if the difference is negative, the direction of the component function relationship edge is determined to be weakening. The absolute value of this difference is used to determine the strength of the functional relationship edge of the component.

[0043] For example, component-functional relationships cannot be directly learned autonomously by deep learning models because, under ordinary feature encoding, models tend to treat component concentrations as parallel fields, failing to distinguish between surface correlations of single components and true synergistic relationships. Determining whether a synergistic relationship exists between two components requires comparative analysis of two types of feedback results in adjacent experiments: "only the primary function changes" and "both the primary function and synergistic functions change."

[0044] The specific judgment logic is as follows: If the stage response (especially the mid-to-late stage response) when the main function and the collaborative function change together is better than the stage response when only the main function changes, it indicates that the addition of the collaborative function enhances the optimization effect of the main function. At this point, it is determined that the relationship between the two functions is strengthened.

[0045] If the stage response when both the main function and the coordinating function change together is worse than the stage response when only the main function changes, it indicates that the addition of the coordinating function interferes with the normal function of the main function. In this case, the relationship between the two functions is weakened.

[0046] If the direction of response improvement when both the main function and the coordinating function change is opposite to the direction of response improvement when only the main function changes (e.g., the response improves when the main function is adjusted alone, but deteriorates when both are adjusted together; or the response deteriorates when adjusted alone, but improves when both are adjusted together), then the relationship between the two functions is determined to have shifted.

[0047] In this embodiment, a primary functional state is selected from the formula response feedback pair, and then two types of adjacent experiments are identified: the first type is experiments where only the primary functional state changes significantly, and the second type is experiments where both the primary functional state and the candidate synergistic functional state change significantly. The stage response differences between the two types of experiments are then compared, and the type of functional relationship change is determined based on the difference characteristics.

[0048] Therefore, this embodiment specifically includes the following steps: In the formula response feedback pair, the improvement in the stage response when only the main functional state changes significantly is recorded, denoted as the stage when the main functional state changes (assuming it is the first stage). Improvement in the phase response of each functional state node The improvement in stage response when both the main functional state and the collaborative functional state undergo significant changes is recorded as the improvement in stage response when both the main functional state and the collaborative functional state undergo significant changes. Among them, the improvement in stage response indicates the degree of change in the response during the cultivation stage after formula adjustment; a positive value indicates an improved response, and a negative value indicates a worsened response.

[0049] Calculate the improvement in stage response when both the primary function state and the co-function state change. Improvement in stage response when only the main function state changes The difference The calculation formula is: ; Difference The physical meaning is: In addition to adjusting the main function, the incremental benefits brought by adjusting the collaborative functions. If A positive value indicates that the addition of collaborative functionality has brought about additional improvements; if... A negative result indicates that the adjustment of the collaborative function has had a negative impact.

[0050] Based on the difference The sign of the component functional relationship edge determines the direction. If the difference is positive ( If the difference is >0, then the direction of the component function relationship edge between the main function and the synergistic function is determined to be reinforcing, indicating that the common change of the synergistic function enhances the response improvement brought about by the change of the main function, that is, there is a synergistic promoting effect between the two functions; if the difference is negative ( If <0), then the direction of the functional relationship edge of the component is determined to be weakening, indicating that the common change of the synergistic function weakens the response improvement brought about by the change of the main function, that is, there is an antagonistic effect between the two functions.

[0051] Difference absolute value The strength of the functional relationship edge is determined by the difference between the two components. A larger absolute value indicates a stronger influence of the synergistic function on the primary function; a smaller absolute value indicates a weaker influence. To further quantify the strength of the relationship edge, the strength can be normalized to obtain the normalized component functional relationship edge. The calculation formula is: ; in, The sign function is used to preserve direction information; This indicates that the absolute value of the response difference is normalized to make the edge strength of the relationship between different functional pairs comparable. When the value is positive and the absolute value is large, it indicates that the common changes have enhanced the effective response of the main function; When the value is negative, it indicates that the common changes have weakened the main functional response.

[0052] In addition, relational redirection markers can be calculated. This indicates whether the response direction changes relative to a change in only the main function after a common change. The calculation formula is: ; A negative value indicates that the response direction after the combined changes has reversed compared to changes in only the main function. That is, adjusting the main function alone improves the response, while adjusting both functions together worsens the response, or vice versa. This reversal indicator can serve as a risk warning for subsequent formulation adjustments.

[0053] Through the above steps, the component-functional relationship edge is obtained by utilizing the feedback differences between adjacent experiments. The relationship edge includes two dimensions: direction (strengthening or weakening) and strength, which provides a basis for subsequent screening of positive relationship edges and extraction of collaborative response features.

[0054] For example, consider carbon source supply as the primary function and nitrogen source supply as a candidate synergistic function. In the first set of adjacent experiments, increasing only carbon source supply (glucose from 10 g / L to 15 g / L) resulted in improved early growth, but decreased metabolism in the middle stage, reduced product yield in the later stage, and decreased cell viability, with a negative improvement in the stage response. In the second set of adjacent experiments, simultaneously increasing both carbon and nitrogen source supply (glucose from 10 g / L to 15 g / L and peptone from 5 g / L to 8 g / L) resulted in improved early growth, stable metabolism in the middle stage, and improved product yield and cell viability in the later stage, with a positive improvement in the stage response. Comparing the two sets of experiments, the improvement in the stage response under the combined change was better than that under the change of only the primary function, with a positive difference. Therefore, the component functional relationship between carbon source supply and nitrogen source supply is determined to be enhanced.

[0055] Conversely, if increasing both carbon and nitrogen sources simultaneously leads to a further decrease in product yield and activity, the difference is negative, and the relationship weakens. If increasing the carbon source alone results in a positive response (improvement), but increasing both sources together results in a negative response (deterioration), the product of the two improvements is negative, and the relationship reverses. In this case, carbon and nitrogen sources should not be recommended as a synergistic adjustment combination even in subsequent iterations.

[0056] S104. Retain the positive component functional relationship edges in the component functional relationship edges, and determine the component functional collaborative response characteristics corresponding to each positive component functional relationship edge based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects.

[0057] In this embodiment, the positive component functional relationship edges in the retained component functional relationship edges specifically include: For any component functional relationship edge, the component functional relationship edge whose direction is strengthening and whose strength is greater than a preset threshold is determined as a positive component functional relationship edge. Preserve the positive component functional relationship edges.

[0058] Based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects, the component functional collaborative response characteristics corresponding to each positive component functional relationship edge are determined, specifically including: Multiply the strength of the positive component functional relationship edge by the values ​​of the two functional state nodes it connects to obtain the component functional collaborative response feature corresponding to the positive component functional relationship edge; The strength of the positive component functional relationship edge represents the degree of strength of that positive component functional relationship edge.

[0059] For example, after functional relationship edges are formed, the deep learning model no longer freely mixes all component fields, but instead controls feature interactions based on the relationship edges. If there is no feedback evidence to support two functional state nodes, the model does not consider them as key interaction items; if there is a positive relationship edge between two functional state nodes, the model enhances the information transmission between them during the feature extraction stage. Through the above processing, the model learns the impact of changes in functional relationships on the culture response, rather than the parallel correlation between component concentrations and final yield.

[0060] In implementation, the relation-aware feature extraction layer uses functional state nodes as node inputs and component functional relation edges as edge inputs. For the ... For each formulation sample, only positive component functional relationship edges with a strengthening direction and a strength within an effective range participate in synergistic feature extraction. Negative relationship edges (with a weakening direction) are not directly used as synergistic enhancement items, but are used as risk indicators in subsequent candidate screening.

[0061] Specifically, this includes: for any component functional relationship edge, identifying those with a reinforcing direction and a strength greater than a preset threshold as positive component functional relationship edges; retaining these positive component functional relationship edges for subsequent collaborative response feature extraction. The preset threshold can be set based on historical experimental data or process experience. For example, relationship edges with a strength greater than 0.3 can be considered valid positive relationship edges, while relationship edges below this threshold, even if their direction is reinforcing, will not participate in feature interaction to avoid introducing noise from weak correlations.

[0062] The synergistic response characteristics of component functions are jointly formed by functional state nodes and positive component functional relationship edges. Functional state nodes provide the functional supply level, while positive component functional relationship edges define which functional state nodes can form effective interactions. If two functional state nodes themselves have high supply levels, but there is no positive relationship edge between them, they should not be directly considered to have a synergistic response.

[0063] The specific calculation method is as follows: Multiply the strength of the positive component functional relationship edge (i.e., the degree of strength of the relationship edge) by the values ​​of the two functional state nodes it connects to obtain the component functional collaborative response feature corresponding to the positive component functional relationship edge. For the th The nth formula sample, let the nth... The functional status node and the first There are positive component functional relationship edges between each functional state node, and their strength is as follows: (Only the positive part is retained), the values ​​of the two functional state nodes are respectively and Then the component functional synergistic response characteristics corresponding to the relation edge The calculation formula is: ; The physical meaning of this product is that the strength of the cooperative interaction between two functional states depends both on the strength of the relationship between them (whether there is experimental evidence to support it) and on their respective supply levels. Even if there is a strong positive relationship between the two functions, if the supply level of one of the functions is very low, the amount of cooperative interaction between them will decrease accordingly; conversely, even if the supply levels of both functions are high, if there is no positive relationship between them (or the relationship is very weak), the amount of cooperative interaction will also be small.

[0064] All function pairs that satisfy the positive reinforcement condition form a set of positive relation edges. The amount of collaborative interaction for each functional pair within the set is normalized and summarized to obtain the first... Component functional synergistic response characteristics of individual formulation samples The calculation formula is: ; Component functional synergistic response characteristics The physical meaning is: the overall synergistic effect among all functional pairs with positive synergistic relationships in the formula. The larger the value, the more effective the synergistic effect of the formula across multiple functional dimensions, which is beneficial for the stable optimization of the cultivation process.

[0065] In the relation-aware feature extraction layer, the deep learning model uses functional state nodes as node inputs and component functional relation edges as edge inputs. For functional state node pairs with positive relation edges, the model enhances the information transfer weights between them during feature extraction; for functional state node pairs without positive relation edges or with negative relation edges, the model suppresses or blocks feature interactions between them. This mechanism can be implemented through graph neural networks or attention mechanisms, enabling the model to control feature interactions based on relation edges and learn the impact of changes in functional relations on the training response.

[0066] Through the above steps, the interaction objects are first defined by the set of relational edges, and then the interaction strength is given by the functional state node. This avoids directly combining all component concentrations in parallel, thereby reducing the interference of single-component surface correlation on model training.

[0067] S105. Sort the culture stage response information of the formula response feedback alignment experimental data according to the time sequence of the culture stages, and calculate the response drop between the next culture stage and the previous culture stage in turn. Calculate the stage response offset characteristics based on the response drop.

[0068] In this embodiment, the culture stage response information of the formula response feedback alignment experimental data is sorted according to the time sequence of the culture stages, and the response decline between the next culture stage and the previous culture stage is calculated sequentially. The stage response offset characteristics are calculated based on the response decline, specifically including: The response information during the cultivation phase is sorted according to the early, middle, and late stages of the cultivation phase and the outcome. The difference between the response information of the next cultivation stage and the response information of the previous cultivation stage is calculated sequentially to obtain the response fall-off amount. If the response information of the next cultivation stage is less than that of the previous cultivation stage, the response fall-off amount is the difference between the two; otherwise, the response fall-off amount is set to zero. The stage response offset characteristics are obtained by summing up all the response fallback values ​​and normalizing them.

[0069] For example, during the model training phase, the component functional synergistic response features calculated above are used... and stage response offset characteristics As input features, these features, along with the final culture response targets (such as target product yield, cell viability, etc.), are used to construct training samples. The model employs a relation-aware feature extraction layer, using graph neural networks or attention mechanisms to achieve feature interaction and learn the impact of changes in functional relationships on the culture response.

[0070] After the model training is completed, for any candidate formulation, the system first extracts its component functional synergistic response features according to steps S102 to S105. and stage response offset characteristics Then, combined with the configuration constraints to meet the proportions Calculate the feasibility in accordance with the method described in S106. Each round of newly generated experimental data is merged with historical data, and the functional relationship edges are recalculated, enabling the model to adapt to the dynamic changes in the correlation between culture medium components round by round, thereby achieving continuous optimization of the formulation.

[0071] Optimizing culture medium formulations cannot be based solely on the final product or early growth stages. Some formulations may rapidly increase early cell density but lead to metabolic instability in the mid-stages or insufficient late-stage products; others may increase the target product but result in decreased cell viability or increased byproducts. These situations indicate that formulation changes have caused a stage response shift, meaning that improvements in the early stages have failed to translate smoothly into later results. Stage response shift characteristics are used to quantify the risk of this "high start, low finish" phenomenon; the greater the shift, the higher the risk of the candidate formulation being a direction for continued optimization.

[0072] The response values ​​for each stage in the culture process are arranged chronologically according to the culture timeline. In this embodiment, early growth, mid-stage metabolism, late-stage products, and cell viability are ordered sequentially to form a stage response sequence. Each stage response value is pre-normalized in the same direction, so that a larger value indicates a more favorable performance at that stage. For example, early growth is normalized using OD600 value, mid-stage metabolism using substrate consumption rate, late-stage products using target protein yield, and cell viability using percentage of live cells.

[0073] Let the first The first formula sample in the first The normalized response value for each cultivation stage is ,in The corresponding stages are early growth, mid-stage metabolism, late-stage products, and cell viability, respectively. The stage response sequence is as follows:

[0074] The response difference between each subsequent culture stage and the previous culture stage is calculated sequentially to obtain the response drop-off. Specifically, for the first culture stage... stage( If the response value in the next stage Less than the response value of the previous stage This indicates a phase of decline, and the amount of the decline in response is the difference between the two. If the response value of the later stage is greater than or equal to the response value of the previous stage, it indicates that the stage response is stable or continuously improving, and the response drop is set to zero. This approach avoids misjudging normal increases or stable performance as risks.

[0075] The calculation formula is: the pullback amount equals The physical meaning of this calculation method is: only the portion of the response that decreases is counted, excluding the portion that increases or remains unchanged. The larger the decline, the more severe the degradation of this stage compared to the previous stage.

[0076] The total fallback is obtained by summing the positive fallback amounts of all adjacent stages, and then normalizing the sum to obtain the stage response offset characteristics. The calculation formula is: ; Stage response offset characteristics The physical meaning is: the degree of overall response decay of the formula from the early to the later stages of cultivation. The larger the value, the more likely the early improvements in the formulation will not be sustained in later products or activity levels, and the higher the risk of using it as a direction for continuous optimization. The smaller the value, the better the response at each stage can be carried over and transmitted, and the better the continuity of the formula optimization effect.

[0077] Assume a formulation has normalized response values ​​of 0.9 for early growth, 0.8 for mid-stage metabolism, 0.6 for late-stage products, and 0.5 for cell viability. Calculate the decline from early to mid-stage as follows: The decline from the middle to the later stage was The survival rate dropped significantly in the later stages. The total decline was After normalization The large value indicates that the formula has a significant "high start, low finish" problem and is not suitable as a direction for continuous optimization.

[0078] The normalized response values ​​for the other formulation at the four stages were: early growth 0.5, mid-stage metabolism 0.6, late-stage products 0.7, and cell viability 0.8. All stages showed an upward trend with no decline, and the total decline was 0. =0 indicates that the optimization effect of the formula can be sustained well, which is an ideal direction for iteration.

[0079] S106. For any candidate formulation, perform statistical analysis on the formulation constraint information to obtain the formulation constraint satisfaction ratio, and calculate the feasibility of the candidate formulation based on the component functional synergistic response characteristics, stage response offset characteristics and formulation constraint satisfaction ratio, and optimize and adjust all candidate formulations according to the feasibility.

[0080] In this embodiment, statistical analysis of the preparation constraint information is performed to obtain the proportion of preparation constraints satisfied, specifically including: The number of formulation constraints satisfied by candidate formulations and the total number of preset formulation constraints are counted. The ratio of the number of satisfied configuration constraints to the total number of preset configuration constraints is determined as the configuration constraint satisfaction ratio.

[0081] Based on the synergistic response characteristics of the components, the stage response shift characteristics, and the proportion of formulation constraint satisfaction of the candidate formulations, the feasibility of the candidate formulations is calculated, specifically including: The exponential decay factor is calculated based on the stage response offset characteristics, where the exponential decay factor is negatively correlated with the stage response offset characteristics; The feasibility of a candidate formulation is determined by multiplying the synergistic response characteristics of its components, the exponential decay factor, and the proportion of formulation constraints.

[0082] All candidate formulations were optimized and adjusted based on feasibility, specifically including: The candidate formulations are sorted according to their feasibility, and the candidate formulation with the highest feasibility is selected as the recommended formulation. The functional state node, component functional relationship edge, stage response offset feature and formulation constraint information of the recommended formulation are output.

[0083] For example, the suitability of a candidate formulation for the next round of experiments cannot be determined solely by its high synergistic response. If the synergistic response is high but the stage response shift is significant, it may only indicate an increase in early growth without improvement in later product yield and activity. Conversely, if the synergistic response is high but the formulation exceeds the limits of solubility, pH, osmotic pressure, or cost, it cannot be used as a practical experimental formulation. Therefore, it is necessary to consider the synergistic response characteristics of component functions, the stage response shift characteristics, and formulation constraints in sequence to comprehensively assess the feasibility of the candidate formulation.

[0084] This embodiment first uses the synergistic response characteristics of component functions as the basic score for formulation improvement, then uses the stage response offset characteristics to attenuate it, and finally uses the formulation constraint to limit it by the proportion. The resulting feasibility is not a parallel stacking of multiple heterogeneous fields, but rather it first judges whether the functional relationship has improved, then judges whether the improvement can continue to the later stages, and finally judges whether the formulation can be actually formulated.

[0085] Formulation constraints include actual production conditions such as solubility, sterilization stability, pH range, osmotic pressure range, and cost limitations. Statistical analysis is performed on the formulation constraints of candidate formulations to count the number of formulation constraints satisfied by each candidate formulation. and the total number of configuration constraints that must be satisfied. (That is, the total number of all constraints that need to be examined, i.e., the total number of preset configuration constraints, or the total number of configuration constraints that should be satisfied). Calculate the ratio of satisfied configuration constraints to the total number of configuration constraints that should be satisfied to obtain the configuration constraint satisfaction ratio. The calculation formula is: ; Configuration constraints satisfy proportion The physical meaning is: the degree to which the candidate formulation is feasible in actual preparation. The closer the value is to 1, the easier the formula is to prepare under actual experimental conditions; The smaller the value, the more formulation limitations exist. Even if the theoretically high synergistic response is high, it is difficult to implement as a practical experimental formulation.

[0086] Based on stage response offset characteristics Calculate the exponential decay factor. Stage response offset characteristics. The larger the value, the greater the risk of the formulation "starting high and ending low", and the recommendation strength of the candidate formulation should be reduced. The smaller the value, the better the response at each stage, and the less the decay. The formula for calculating the exponential decay factor is: Exponential decay factor equals... The physical meaning of this factor is: to decay the baseline score in the form of an exponential function. When When =0, No decay occurs; when hour, The attenuation is approximately 40%; when hour, The attenuation is approximately 63%; when The larger the value, the more significant the decay. The reason for using exponential decay instead of nonlinear decay is that when the risk of stage deviation is high, the feasibility of the formulation as a direction for continuous optimization should be quickly suppressed to avoid high-risk formulations being mistakenly recommended.

[0087] Synergistic response characteristics of component functions in candidate formulations Exponential decay factor and the proportion of configuration constraints to be satisfied Multiplying these three factors together yields the feasibility of the candidate formulation. The calculation formula is: ; Feasibility The physical meaning is: the overall score of the candidate formulation after taking into account functional synergy, stage response continuity and actual formulation feasibility. The larger the value, the better the formulation performs in terms of functional synergy, phase continuity, and practical feasibility, making it suitable as a candidate formulation for the next round of experiments. The smaller the value, the more likely it is to indicate a significant defect in at least one dimension.

[0088] For multiple candidate formulations, calculate the feasibility of each formulation to obtain a set of feasibility values. The next step will be to rank and optimize all candidate formulations based on their feasibility.

[0089] The system is based on the feasibility of each candidate formulation. The system prioritizes candidate formulations with high feasibility and no rigid formulation restrictions. Specifically, the candidate formulation with the highest feasibility is selected as the recommended formulation. If a candidate formulation has high synergistic response characteristics of its components but large stage response deviation characteristics, the system will not directly recommend the formulation. Instead, it will retain the functional relationship edge and the source of the stage deviation, suggesting that the adjustment magnitude of the components causing the later decline should be reduced in the next round of experiments. For example, if a formulation has a high synergistic response in carbon source supply, but the stage response deviation characteristics show a significant decline in later product and activity, the system will not blindly recommend the formulation. Instead, it will prompt the experimenter to pay attention to and reduce the carbon source adjustment magnitude that leads to the later decline.

[0090] When generating candidate recipes, the following rules apply: Rule 1: If the component function relationship between the carbon source supply function and the nitrogen source supply function is positive (indicating enhancement), and the stage response offset characteristic is below a preset threshold, then simultaneous fine-tuning of the carbon source and nitrogen source is allowed within the process range. This is because a positive relationship indicates a synergistic promoting effect between the two, and simultaneous adjustment can produce better cultivation results.

[0091] Rule 2: If the phase response shift characteristic exceeds a preset threshold after increasing the carbon source, buffer stability and osmotic pressure limitation should be checked first, and the carbon source should not be increased directly. This is because a large phase shift indicates that the response declines after increasing the carbon source, and the problem may not lie with the carbon source itself, but with the buffer system or osmotic pressure failing to accommodate the accelerated carbon source metabolism.

[0092] Rule 3: If a positive component function relationship is formed between the activation function of trace elements and the response of subsequent products, then trace element candidate modifiers will be generated without exceeding the toxicity or precipitation limits. This is because the positive relationship between trace elements and subsequent products indicates that a reasonable supply of trace elements is beneficial to product accumulation, but attention must be paid to the toxicity threshold and precipitation risk of trace elements.

[0093] Rule 4: If a candidate formulation fails to meet any of the constraints of solubility, sterilization stability, pH range, osmotic pressure range, or cost limitations, it will not be included in the next round of experimental recommendations and will only be recorded as an unpreparable candidate. This is because even if the theoretically synergistic response is high, it has no practical value if it cannot be prepared under actual experimental conditions.

[0094] The recommended formulation, functional state nodes, component functional relationship edges, stage response offset characteristics, and formulation constraint information are output together for experimental personnel to refer to and execute.

[0095] After operators or experimenters perform the next round of cultivation according to the recommended formula, the new experimental results are added to the historical sample database and form a new formula response feedback pair with the results of the previous round. Through this cycle, the system can progressively correct the functional relationship edges of the components, allowing the deep learning feature extraction process to continuously adapt to the dynamic changes in the correlation between culture medium components. As the number of iterations increases, the accuracy of the functional relationship edges gradually improves, and the recommendation quality of candidate formulas is continuously optimized.

[0096] This invention transforms components into functional state nodes and constructs functional relationship edges based on feedback from adjacent experiments, enabling deep learning models to learn the real synergistic relationships between components. It combines stage response offsets and formulation constraints to calculate the feasibility of the formulation, and continuously corrects the relationship edges through iterative rounds, effectively improving the stability, feasibility, and adaptive optimization capability of culture medium formulation optimization recommendations.

[0097] This invention also proposes a deep learning-based intelligent iterative optimization system for culture medium formulations; please refer to [link / reference]. Figure 2 The diagram shows a structural diagram of a deep learning-based intelligent iterative optimization system for culture medium formulations provided in an embodiment of the present invention. The system includes: a data acquisition module 101, a data processing module 102, and an optimization adjustment module 103.

[0098] The data acquisition module 101 is used to acquire formulation response feedback pairs, wherein the formulation response feedback pairs are obtained by pairing the experimental data of two adjacent experiments on the same culture object, and the experimental data includes formulation component information, culture stage response information and formulation constraint information. Data processing module 102 is used to convert the formula component information according to its equivalent contribution in the culture function category to obtain the functional state node; Take any functional state node as the main function and any other functional state node as the coordinating function. Obtain the stage response improvement amount when only the main function state changes in the cultivation stage response information, and the stage response improvement amount when both the main function state and the coordinating function state change. Based on the difference between the two improvement amounts, determine the component function relationship edge between the main function and the coordinating function. Retain the positive component functional relationship edges in the component functional relationship edges, and determine the component functional collaborative response characteristics corresponding to each positive component functional relationship edge based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects; The culture stage response information of the formula response feedback alignment experimental data is sorted according to the time sequence of the culture stages, and the response drop of the next culture stage is calculated in turn. The stage response offset characteristics are calculated based on the response drop. The optimization and adjustment module 103 is used to perform statistical analysis on the formulation constraint information for any candidate formulation to obtain the formulation constraint satisfaction ratio, and calculate the feasibility of the candidate formulation based on the component functional synergistic response characteristics, stage response offset characteristics and formulation constraint satisfaction ratio of the candidate formulation, and optimize and adjust all candidate formulations according to the feasibility.

[0099] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the intelligent iterative optimization system for culture medium formulation based on deep learning and the intelligent iterative optimization method for culture medium formulation based on deep learning provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiments and will not be repeated here.

[0100] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent iterative optimization of culture medium formulations based on deep learning, characterized in that, include: Obtain formulation response feedback pairs, which are obtained by pairing experimental data from two adjacent experiments on the same culture subject. The experimental data include formulation component information, culture stage response information, and formulation constraint information. The formula component information is converted according to its equivalent contribution in the culture function category to obtain the functional state node; Take any functional state node as the main function and any other functional state node as the coordinating function. Obtain the stage response improvement amount when only the main function state changes in the cultivation stage response information, and the stage response improvement amount when both the main function state and the coordinating function state change. Based on the difference between the two improvement amounts, determine the component function relationship edge between the main function and the coordinating function. Retain the positive component functional relationship edges in the component functional relationship edges, and determine the component functional collaborative response characteristics corresponding to each positive component functional relationship edge based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects; The culture stage response information of the formula response feedback alignment experimental data is sorted according to the time sequence of the culture stages, and the response drop of the next culture stage is calculated in turn. The stage response offset characteristics are calculated based on the response drop. For any candidate formulation, statistical analysis is performed on the formulation constraint information to obtain the proportion of formulation constraint satisfaction. Based on the component functional synergistic response characteristics, stage response offset characteristics, and formulation constraint satisfaction proportion of the candidate formulation, the feasibility of the candidate formulation is calculated, and all candidate formulations are optimized and adjusted according to the feasibility.

2. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The process of converting the formulation component information according to its equivalent contribution in the culture functional category to obtain the functional state node specifically includes: For any culture function category, obtain the concentrations of all components in the formulation component information, where the formulation component information includes the concentration of each component; The product of the concentration of each component and the equivalent conversion coefficient corresponding to that component is determined as the component conversion supply amount for that culture functional category. Sum the converted supply quantities of all components to obtain the total converted supply quantity for this culture function category; The ratio of the total converted supply to the reference upper limit of the cultivation function category is normalized to obtain the functional state node corresponding to the cultivation function category.

3. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The step of determining the component functional relationship edge between the main function and the synergistic function based on the difference between the two improvement quantities specifically includes: Calculate the difference between the improvement in stage response when both the main functional state and the collaborative functional state change and the improvement in stage response when the main functional state changes. If the difference is positive, the direction of the component function relationship edge between the main function and the synergistic function is determined to be strengthening; if the difference is negative, the direction of the component function relationship edge is determined to be weakening. The absolute value of this difference is used to determine the strength of the functional relationship edge of the component.

4. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The positive component functional relationship edges in the retained component functional relationship edges specifically include: For any component functional relationship edge, the component functional relationship edge whose direction is strengthening and whose strength is greater than a preset threshold is determined as a positive component functional relationship edge. Preserve the positive component functional relationship edges.

5. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The step of determining the component functional collaborative response characteristics corresponding to each positive component functional relationship edge based on the strength of the positive component functional relationship edge and the values ​​of the two functional state nodes it connects specifically includes: Multiply the strength of the positive component functional relationship edge by the values ​​of the two functional state nodes it connects to obtain the component functional collaborative response feature corresponding to the positive component functional relationship edge; The strength of the positive component functional relationship edge represents the degree of strength of that positive component functional relationship edge.

6. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The process of sorting the culture stage response information of the formulation response feedback alignment experimental data according to the time sequence of the culture stages, and calculating the response decline between the later and previous culture stages in turn, and calculating the stage response shift characteristics based on the response decline, specifically includes: The response information during the cultivation phase is sorted according to the early, middle, and late stages of the cultivation phase and the outcome. The difference between the response information of the next cultivation stage and the response information of the previous cultivation stage is calculated sequentially to obtain the response fall-off amount. If the response information of the next cultivation stage is less than that of the previous cultivation stage, the response fall-off amount is the difference between the two; otherwise, the response fall-off amount is set to zero. The stage response offset characteristics are obtained by summing up all the response fallback values ​​and normalizing them.

7. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The statistical analysis of the formulation constraint information to obtain the proportion of formulation constraint satisfaction specifically includes: The number of formulation constraints satisfied by candidate formulations and the total number of preset formulation constraints are counted. The ratio of the number of satisfied configuration constraints to the total number of preset configuration constraints is determined as the configuration constraint satisfaction ratio.

8. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The feasibility of candidate formulations is calculated based on the synergistic response characteristics of component functions, stage response offset characteristics, and the proportion of formulation constraint satisfaction. Specifically, this includes: The exponential decay factor is calculated based on the stage response offset characteristics, where the exponential decay factor is negatively correlated with the stage response offset characteristics; The feasibility of a candidate formulation is determined by multiplying the synergistic response characteristics of its components, the exponential decay factor, and the proportion of formulation constraints.

9. The intelligent iterative optimization method for culture medium formulation based on deep learning according to claim 1, characterized in that, The optimization and adjustment of all candidate formulations based on feasibility specifically includes: The candidate formulations are sorted according to their feasibility, and the candidate formulation with the highest feasibility is selected as the recommended formulation. The functional state node, component functional relationship edge, stage response offset feature and formulation constraint information of the recommended formulation are output.

10. A deep learning-based intelligent iterative optimization system for culture medium formulations, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent iterative optimization method for culture medium formulation based on deep learning as described in any one of claims 1-9.

Images