Strain-specific medium prediction method, medium, and computer device

Abstract

The application discloses a strain-specific culture medium prediction method, a medium and a computer device, wherein the method comprises the following steps: obtaining strain gene function annotation data and coding into a high-dimensional feature vector; performing principal component analysis dimension reduction on the high-dimensional feature vector; obtaining candidate culture medium component information and coding into a binary vector; splicing the low-dimensional strain characteristic representation and the culture medium component vector to form a joint input feature vector; inputting the joint input feature vector into a pre-trained deep residual network model; and outputting a growth compatibility probability score and sorting the candidate culture medium. The application integrates high-throughput genome annotation data and known culture medium formula to construct a data-driven intelligent prediction model, which is used for efficiently recommending the optimal culture medium composition suitable for a specific microbial strain, and significantly improves the prediction accuracy and screening speed of microbial separation and culture.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, specifically to a method, medium, and computer device for predicting strain-specific culture media that integrates microbial genome functional characteristics and deep residual learning. Background Technology

[0002] Microorganisms play an irreplaceable role in fields such as pharmaceutical development, food fermentation, environmental remediation, and synthetic biology. However, more than 99% of microorganisms in nature still cannot be successfully cultured under laboratory conditions. This phenomenon, known as the "Great Plate Count Anomaly," severely restricts the development and utilization of microbial resources.

[0003] Traditional culture medium optimization relies on empirical trial-and-error methods, such as the One-Factor Rotation Approach (OFAT) and Response Surface Methodology (RSM). These methods are time-consuming, costly, and struggle to capture the nonlinear interactions between nutrients, proving almost ineffective, especially for uncultured microorganisms with complex nutritional requirements. In recent years, machine learning techniques have been introduced into this field, attempting to predict suitable culture conditions using historical experimental data. Existing studies have employed random forests, support vector machines, or shallow neural networks for modeling, achieving some progress in specific scenarios. However, the aforementioned predictive methods incorporating machine learning generally suffer from the following bottlenecks: (1) High-dimensional sparsity problem is prominent: the dimension of genome feature vectors based on functional annotations such as KEGG Orthology (KO) can reach tens of thousands, which can easily lead to the "curse of dimensionality", resulting in overfitting and decreased generalization ability of the model.

[0004] (2) Lack of effective modeling of complex relationships: Microbial growth is the result of a highly nonlinear coupling between the metabolic potential of strains and the supply of nutrients in the environment. Traditional models are unable to fully explore the deep relationship between the two.

[0005] (3) There is a lack of systematic dimensional reduction comparison and biological interpretability support. Most studies simply use PCA without demonstrating its superiority. Summary of the Invention

[0006] The present invention provides a strain-specific culture medium prediction method, medium and computer equipment that can effectively deal with high-dimensional genomic data and accurately model strain-culture medium matching relationships, and can solve at least one of the above-mentioned technical problems.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for predicting strain-specific culture media includes the following steps: S1. Obtain gene function annotation data of the microbial strain to be tested and encode it into a high-dimensional binary feature vector; S2. Perform principal component analysis to reduce the dimensionality of the high-dimensional binary feature vector, retain the principal components with a cumulative variance contribution rate of more than 95%, and obtain the low-dimensional strain feature representation. S3. Obtain the component information of all candidate culture media and encode it into a binary component vector, where each dimension of the binary component vector represents whether a certain culture media component exists or not. S4. The low-dimensional strain feature representation obtained in S2 is concatenated with the culture medium component vector obtained in S3 to form a joint input feature vector; S5. Input the joint input feature vector into the pre-trained deep residual network model. The deep residual network model contains multiple fully connected residual blocks with skip connections, which are used to learn the complex nonlinear relationship between the strain and the culture medium. S6. The deep residual network model outputs prediction results, which indicate whether the current microbial strain to be tested can grow in the candidate culture medium. The output prediction result value is activated by the Sigmoid function and used as a probability score for growth compatibility. S7. Sort multiple candidate culture media according to probability scores, and recommend the Top-K culture media with the best probability to support the growth of the current microbial strain under test.

[0008] Further, in step S1, the gene function annotation data includes information on the presence or absence of KEGG Orthology identifiers. The gene function annotation data is encoded into a high-dimensional binary feature vector, where each dimension corresponds to a KEGG Orthology identifier. A value of 1 indicates that the KEGG Orthology identifier exists in the strain genome, and a value of 0 indicates that it does not exist.

[0009] Furthermore, in S1, the gene function annotation data is sourced from the KEGG database, and the high-dimensional binary feature vector has a dimension of 9747, covering the key metabolic function features of the strain.

[0010] Furthermore, in S2, the low-dimensional strain feature representation after principal component analysis has a dimension of 1598. This is achieved by retaining principal components with a cumulative variance contribution rate of over 95%. This compresses the feature space while maximizing the preservation of key biological information on functional and metabolic differences between strains, thereby mitigating the curse of dimensionality.

[0011] Furthermore, in step S3, the component information of the candidate culture medium comes from the MediaDive database, and the binary component vector has a dimension of 491, with each dimension corresponding to a unique culture medium component, which includes at least a carbon source, a nitrogen source, vitamins, inorganic salts, and growth factors.

[0012] Furthermore, in S5, the deep residual network model includes three residual blocks, each consisting of two fully connected layers, a ReLU activation function, and a Dropout layer, and is equipped with a skip connection structure to alleviate the gradient vanishing problem in deep networks and enhance the learning ability of high-dimensional sparse biological data.

[0013] Furthermore, in S5, the deep residual network model is trained using the maximum margin loss function. By distinguishing between compatible and incompatible strain-culture medium paired samples, the model's discriminative ability in the embedding space is optimized. The maximum margin loss function introduces positive and negative sample pairs and sets a marginal hyperparameter γ to enhance classification robustness.

[0014] Furthermore, in step S7, Top-K accuracy is selected as the evaluation index of the deep residual network model. For a certain strain, the strain information is paired with all n candidate culture media combinations one by one to generate corresponding n data points, which are input into the trained deep residual network model to obtain the corresponding n m-dimensional output vectors. The L2 norm distance between the n m-dimensional output vectors and the m-dimensional all-1 vectors is calculated respectively, and the top K candidate culture media with the smallest distance are selected as the Top-K recommendation results.

[0015] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the strain-specific culture medium prediction method described above.

[0016] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the strain-specific culture medium prediction method described above.

[0017] The beneficial effects of this invention are reflected in: 1. Principal component analysis (PCA) is used to effectively process high-dimensional genomic data, preserve key metabolic features, avoid the curse of dimensionality, and improve the model's generalization ability.

[0018] 2. By using deep residual networks to model the complex nonlinear relationship between bacterial strains and culture media, the accuracy and stability of predictions are significantly improved.

[0019] 3. Achieve end-to-end intelligent matching from gene function to culture conditions, providing a reliable tool for the isolation and industrial application of difficult-to-culture microorganisms, reducing the consumption of experimental resources and shortening the screening cycle. Attached Figure Description

[0020] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0021] Figure 1 This is a schematic diagram of the overall process of the strain-specific culture medium prediction method according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the KO dataset structure according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the deep residual network model structure according to an embodiment of the present invention.

[0024] Figure 4 This is a graph showing the comparison results of different models in Top-K accuracy in embodiments of the present invention.

[0025] Figure 5 This is a graph showing the comparison of the standard deviations of different models in Top-K accuracy in embodiments of the present invention.

[0026] Figure 6 This is a graph showing the comparison of Top-K accuracy between PCA dimensionality reduction and other dimensionality reduction methods in this embodiment of the invention.

[0027] Figure 7 This is a graph showing the comparison of the standard deviation of PCA dimensionality reduction with other dimensionality reduction methods in Top-K accuracy in this embodiment of the invention.

[0028] Figure 8 This is a structural block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0029] 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 a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that the meaning of "and / or" throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, solution B, or a solution that simultaneously satisfies A and B. Furthermore, "multiple" refers to two or more. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0031] See Figure 1 This invention provides a method for predicting strain-specific culture media, based on a strain-specific culture media prediction system that integrates genomic functional features and deep residual learning. The system architecture includes a data processing module, a feature engineering module, and a residual network prediction module, aiming to efficiently and accurately match the gene functional features of a strain with compatible culture medium components through a computational model. Specifically, this prediction method includes the following steps: Step 1: Data Acquisition and Preprocessing 1.1 Acquisition and Encoding of Strain Gene Function Annotation Data The protein sequence files (FAA format) of the microbial strains were downloaded from the NCBI database, and functional annotation was performed using the KofamScan tool to obtain the KEGG Orthology (KO) number. The KO information of each strain was encoded as a 9747-dimensional binary vector (capable of covering the key metabolic functional features of the strain), where each dimension corresponds to a KO number. A value of 1 indicates that the KO exists in the strain, and a value of 0 indicates that it does not exist.

[0032] 1.2 Acquisition and Coding of Culture Medium Component Information Culture medium formulation information, including component names and concentrations, was crawled from the MediaDive database. To simplify modeling, the culture medium components were represented as 491-dimensional binary vectors, with each dimension corresponding to a single component. A value of 1 indicates that the component is present in the culture medium, and a value of 0 indicates that it is not present.

[0033] 1.3 Construction of positive and negative samples and dataset structure Known culturable bacterial strains paired with culture media were used as positive samples (9140 pairs in total), and unculturable pairs were randomly generated as negative samples (the number of pairs was the same as the number of positive samples), forming a balanced dataset. The dataset structure is as follows. Figure 2 As shown, this illustrates the relationship between the strain KO information and the components of the culture medium.

[0034] Step 2: PCA Dimensionality Reduction and Curse of Dimension Analysis To alleviate the curse of dimensionality caused by high-dimensional data, the original joint input feature space of this invention is composed of 9747-dimensional strain KO features and 491-dimensional culture medium component features, resulting in a total dimension of 10238. Therefore, this invention performs principal component analysis (PCA) dimensionality reduction on the 9747-dimensional KO feature vector of the strains. The specific steps include: 2.1 Standardize the KO data so that its mean is 0 and its variance is 1; 2.2 Calculate the covariance matrix of the standardized KO eigenvectors; 2.3 Calculate the eigenvalues ​​and eigenvectors of the covariance matrix; 2.4 Select principal components with a cumulative variance contribution rate of over 95% to obtain the dimensionality-reduced strain characteristic representation (1598 dimensions).

[0035] Furthermore, regarding the "dimensional catastrophe theory analysis": In high-dimensional space, the distance between data points increases on the order of the square root. Two points are independently and uniformly sampled within a unit hypercube. and The expected value of its Euclidean distance is:

[0036] in, The expected value is represented by d, the Euclidean distance is represented by x, and the y-axis is represented by x and y, respectively. Randomly selecting two points in a 10238-dimensional space results in a distance of approximately 41.3, leading to an extremely sparse sample distribution in the high-dimensional space.

[0037] Furthermore, regarding the "PCA dimensionality reduction implementation": The KO data were standardized to have a mean of 0 and a variance of 1. The covariance matrix was calculated and eigenvalue decomposition was performed. Principal components with a cumulative variance contribution rate of more than 95% were selected. Finally, the KO features were reduced from 9747 dimensions to 1598 dimensions, which significantly reduced the feature dimensionality while retaining most of the biological information.

[0038] Step 3: Construction and Training of Deep Residual Network Model 3.1 Model Architecture Design The deep residual network (ResNet) model structure used in this invention is as follows: Figure 3 As shown, the specific architecture is as follows: Input layer: The dimensionality-reduced strain features (1598 dimensions) and the culture medium component vector (491 dimensions) are concatenated into a 2089-dimensional joint input vector; Hidden layers: First, the input is mapped to 1024 dimensions through a fully connected layer, followed by 3 residual blocks. Each residual block contains two fully connected layers, a ReLU activation function, and a Dropout layer (dropout rate of 0.5), and uses a skip connection structure. Output layer: Finally, a 50-dimensional fully connected layer outputs a feature vector, which is then converted into growth compatibility probability by a sigmoid activation function; Loss function: The maximum margin loss function is used, and its calculation expression is as follows:

[0039] in: The margin hyperparameter serves two purposes: firstly, it prevents the overall loss from becoming negative too early, ensuring that the loss function continuously provides effective gradients; secondly, it defines the minimum margin between positive and negative samples in the vector space. The Euclidean distance (also known as the L2 norm distance) between two vectors is:

[0040] The model output represents positive examples. These are positive examples; The model output represents the negative examples. These are negative examples; express A dimensional vector of all 1s express A vector consisting entirely of zeros; The design of this loss function is similar to that of the TransE algorithm loss function. By minimizing the maximum margin function, the output of the positive examples is made as close as possible to... A vector of all 1s, the output of negative examples should be as close as possible to... The zero-dimensional vector effectively distinguishes between compatible and incompatible strain-culture medium paired samples.

[0041] 3.2 Training Algorithm The model training employs the maximum margin loss minimization algorithm, optimizing model parameters through batch sampling and gradient descent.

[0042] Step 4: Experimental Results and Model Performance Validation 4.1 Evaluation Index System Top-K accuracy was used as the primary evaluation metric. In the model evaluation phase, for a given strain, its information was paired with all 880 possible culture medium combinations, generating 880 data points, which were then input into the trained deep residual network model, resulting in 880 m-dimensional output vectors. By calculating the L2 norm distance between these 880 m-dimensional output vectors and the m-dimensional all-1 vector, the top K candidate culture media with the smallest distances were selected as the Top-K recommendations.

[0043] The Top-K accuracy is calculated as follows: if the Top-K recommendations for a strain include at least one experimentally verified culture medium that can cultivate the strain, then the prediction for that strain is considered correct. The Top-K accuracy of the overall dataset is obtained by averaging the prediction accuracy of all strains. For example, assuming there are 1000 strains, if 100 strains have correct Top-1 predictions, the Top-1 accuracy is 10%; if 200 strains have Top-5 predictions that include at least one correct culture medium, the Top-5 accuracy is 20%. In this invention, K=1, 5, 10, 20, and 100 are used for accuracy statistics in experiments.

[0044] Experimental section: I. Experimental Results of Model Comparison To verify the superiority of deep residual networks (ResNet) in this invention, this embodiment compares it with multilayer perceptrons (MLP), multi-head attention models, and Transformer models. Top-K accuracy (K=1, 5, 10, 20, 100) is used as the evaluation metric in the experiments.

[0045] like Figure 4 As shown, under the same input features, the ResNet model of this invention performs best in all metrics. Specific data demonstrates: Accuracy advantage: The ResNet model achieves a Top-5 accuracy of 66.07%, significantly higher than the 54.69% of the multilayer perceptron and the 60.52% of the Transformer model. In the Top-10 and Top-20 metrics, ResNet also maintains a lead of approximately 4%-8%.

[0046] Stability advantages: such as Figure 5 As shown, by comparing the standard deviations of each model in multiple experiments, the ResNet model has the smallest standard deviations among all the compared models for Top-10 (2.83%), Top-20 (2.18%), and Top-100 (0.70%). This indicates that ResNet not only has high prediction accuracy but also exhibits strong robustness in processing high-dimensional sparse biological data and is not easily affected by random factors.

[0047] II. Comparison Experiment of Dimensionality Reduction Methods To address the sparsity problem of high-dimensional genomic features (KO features), this embodiment compares the performance differences between PCA dimensionality reduction and pathway dimensionality reduction, random projection, t-SNE, and UMAP.

[0048] like Figures 6-7 As shown, after using PCA to reduce the dimensionality of the features, the model performance was significantly improved. The remarkable effectiveness of PCA: When using PCA to preserve 95% of the cumulative variance (reduced to 1598 dimensions), the model's Top-5 accuracy soared to 77.17%. Compared to no dimensionality reduction or other dimensionality reduction methods: Compared to pathway dimensionality reduction (Top-5 is 57.15%), the PCA method improves by about 20 percentage points, demonstrating that feature extraction based on statistical data characteristics has a better information retention capacity than path mapping based solely on biological prior knowledge. Compared to random projection (65.74%) and nonlinear methods (t-SNE 59.07% and UMAP 54.08%), PCA demonstrates excellent adaptability in processing high-dimensional biological data.

[0049] Extremely low standard deviation: in terms of stability ( Figure 7 The PCA-reduced model has extremely low standard deviations in the Top-20 (0.97%) and Top-100 (0.55%) metrics, which is far superior to path reduction (3.65% for Top-20) and other manifold learning methods.

[0050] III. Experimental Conclusions Based on the above experiments, the configuration of PCA dimensionality reduction (1598 dimensions) combined with a 3-layer residual network achieves an accuracy of 77.17% in the Top-5 recommendation task, with minimal model variance. This combination constitutes the optimal embodiment recommended by this invention.

[0051] IV. Optimal Parameter Configuration Experiments have verified that the optimal parameter configuration is as follows: Number of residual blocks: 3 Hidden layer dimension: 1024 Dropout rate: 0.5 Output dimension: 50 dimensions PCA dimensionality reduction: 1598 dimensions (retaining 95% variance information) This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the strain-specific culture medium prediction method described above.

[0052] See Figure 8 The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the strain-specific culture medium prediction method described above.

[0053] This invention also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform the steps of the strain-specific culture medium prediction method described above.

[0054] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above-mentioned strain-specific culture medium prediction method.

[0055] It should be noted that those skilled in the art will understand that all or part of the steps implemented in the embodiments of the present invention can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. When implemented in hardware, it can be implemented entirely or partially by purchasing standard parts or modifications. When implemented in software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid state disks (SSDs)).

[0056] In summary, this invention primarily addresses the technical challenges of low screening efficiency in traditional culture media and the difficulty in culturing most microorganisms. It provides a strain-specific culture medium prediction system that integrates genomic functional characteristics and deep residual learning. Specifically, this method uses principal component analysis (PCA) to reduce the dimensionality of high-dimensional functional gene profiles and combines this with a ResNet residual network to construct an end-to-end relational learning model. This enables high-precision prediction of whether microorganisms can grow in specific culture media, providing a scientific basis for the design of novel culture media and the domestication of difficult-to-culture microorganisms. Experimental verification shows that this method can effectively support the growth prediction of difficult-to-culture microorganisms and is applicable to fields such as microbial resource development, synthetic biology, and industrial fermentation, demonstrating promising application prospects and economic benefits.

[0057] It should be understood that the examples and embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various modifications or changes based on them. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. A method for predicting strain-specific culture media, characterized in that, Includes the following steps: S1. Obtain gene function annotation data of the microbial strain to be tested and encode it into a high-dimensional binary feature vector; The gene function annotation data includes information on the presence or absence of KEGG Orthology identifiers. The gene function annotation data is encoded into a high-dimensional binary feature vector, where each dimension corresponds to a KEGG Orthology identifier. A value of 1 indicates that the KEGG Orthology identifier exists in the strain genome, and a value of 0 indicates that it does not exist. S2. Perform principal component analysis to reduce the dimensionality of the high-dimensional binary feature vector, retain the principal components with a cumulative variance contribution rate of more than 95%, and obtain the low-dimensional strain feature representation. S3. Obtain the component information of all candidate culture media and encode it into a binary component vector. Each dimension of the binary component vector represents whether a certain culture media component exists or not. A value of 1 indicates that the component exists in the culture media, and a value of 0 indicates that it does not exist. S4. Take the known culturable strain-culture medium pairings as positive samples, and randomly generate an equal number of unculturable pairs as negative samples. Concatenate the low-dimensional strain feature representation obtained in S2 with the culture medium component vector obtained in S3 to form a joint input feature vector. S5. Input the joint input feature vector into the pre-trained deep residual network model. The deep residual network model contains multiple fully connected residual blocks with skip connections, which are used to learn the complex nonlinear relationship between the strain and the culture medium. The deep residual network model is trained using the maximum margin loss function. By distinguishing between compatible and incompatible strain-culture medium pairs, the model's discriminative ability in the embedding space is optimized. The maximum margin loss function introduces positive and negative sample pairs and sets a marginal hyperparameter γ to enhance classification robustness. The maximum interval loss function is calculated as follows: in: For the interval hyperparameter, The L2 norm distance between two vectors. The model output represents positive examples. As a positive sample, The model output represents the negative examples. For negative samples, express A dimensional vector of all 1s express A vector consisting entirely of zeros; This loss function minimizes the maximum margin function, making the output of positive samples close to... A vector of all 1s, the output of negative samples is close to A zero-dimensional vector is used to distinguish between compatible and incompatible strain-culture medium paired samples; S6. The deep residual network model outputs prediction results, which indicate whether the current microbial strain to be tested can grow in the candidate culture medium. The output prediction result value is activated by the Sigmoid function and used as a probability score for growth compatibility. S7. Sort multiple candidate culture media according to probability scores and recommend the Top-K culture media with the best probability to support the growth of the current microbial strain under test. Top-K accuracy was selected as the evaluation index of the deep residual network model. For a certain strain, the strain information was paired with all n candidate culture media combinations one by one to generate the corresponding n data points, which were then input into the trained deep residual network model to obtain the corresponding n m-dimensional output vectors. The L2 norm distance between the n m-dimensional output vectors and the m-dimensional all-1 vectors was calculated, and the top K candidate culture media with the smallest distance were selected as the Top-K recommendation results.

2. The method for predicting strain-specific culture media as described in claim 1, characterized in that, In S1, the gene function annotation data comes from the KEGG database, and the high-dimensional binary feature vector has a dimension of 9747, covering the key metabolic function features of the strain.

3. The method for predicting strain-specific culture media as described in claim 1, characterized in that, In S2, the low-dimensional strain feature representation after principal component analysis has a dimension of 1598. This is achieved by retaining principal components with a cumulative variance contribution rate of over 95%. This compresses the feature space while maximizing the preservation of key biological information on functional and metabolic differences between strains, thus mitigating the curse of dimensionality.

4. The method for predicting strain-specific culture media as described in claim 1, characterized in that, In step S3, the component information of the candidate culture medium comes from the MediaDive database. The binary component vector has 491 dimensions, and each dimension corresponds to a unique culture medium component. The culture medium component includes at least a carbon source, a nitrogen source, vitamins, inorganic salts, and growth factors.

5. The method for predicting strain-specific culture media as described in claim 1, characterized in that, In S5, the deep residual network model contains three residual blocks. Each residual block consists of two fully connected layers, a ReLU activation function, and a Dropout layer. It also has a skip connection structure to alleviate the gradient vanishing problem in deep networks and enhance the learning ability of high-dimensional sparse biological data.

6. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the strain-specific culture medium prediction method as described in any one of claims 1-5.

7. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the strain-specific culture medium prediction method as described in any one of claims 1-5.

Images