Microorganism culture medium recommendation method and system based on gene large model embedding

Abstract

The application discloses a kind of microorganism culture medium recommendation method and system based on gene big model embedding, input target microorganism whole genome sequence into pre-trained genome large language model to obtain genome feature vector, feature compression and denoising processing are carried out by stacked denoising auto-encoder network to obtain low-dimensional microorganism strain feature vector, which is combined with the chemical composition concentration feature vector of candidate culture medium to construct pairing feature set, input deep metric learning network based on bilinear attention mechanism to calculate the adaptation degree score, and output the optimal culture medium recommendation formula in order from high to low according to score.The application extracts deep semantic features using genome large language model, compresses into low-dimensional discriminative representation through supervised dimension reduction network, models multi-order feature interaction using factorization machine, optimizes recommendation quality using ranking learning loss function, and realizes accurate and efficient culture medium formula recommendation.

Description

Technical Field

[0001] This invention relates to the technical field of microbial culture, and in particular to a method and system for recommending microbial culture media based on large gene model embedding. Background Technology

[0002] Microorganisms play a vital role in fields such as industrial biotechnology, pharmaceutical research and development, and environmental remediation. The design of culture medium formulations directly determines the success rate of microbial culture. Traditional methods mainly rely on expert experience and repeated trial and error, which are time-consuming, labor-intensive, and have a low success rate, making it difficult to meet the rapidly growing demand for the culture of novel microorganisms.

[0003] With the development of high-throughput sequencing technology, the cost of whole-genome sequencing of microorganisms has been significantly reduced. Genome sequences contain key physiological characteristics such as microbial metabolic pathways and nutritional requirements, making it possible to predict culture conditions from genotype. However, existing genome analysis methods mainly focus on gene function annotation and metabolic pathway reconstruction, lacking an effective way to directly translate genomic information into specific culture medium formulations.

[0004] In existing technologies, metabolic network reconstruction-based methods infer microbial nutritional requirements by analyzing metabolic enzymes encoded in the genome. However, these methods rely on the completeness of known gene function annotations, resulting in low reliability in predicting microorganisms with genes of unknown function. Furthermore, they only provide qualitative information and cannot determine the optimal concentration ratio of nutrients. Another type of machine learning-based method faces the curse of dimensionality when processing high-dimensional genomic data. Traditional feature engineering methods, such as k-mer frequency statistics, can only capture local statistical characteristics, making it difficult to express deep semantic information and long-range dependencies. Moreover, they often employ simple linear models, failing to effectively model the complex nonlinear interactions between microbial gene features and the chemical composition of the culture medium.

[0005] In recent years, large-scale genome language models based on the Transformer architecture have been able to learn general representations of gene sequences and capture deep features such as evolutionary conservation and functional patterns. However, existing research has mainly applied them to tasks such as gene function prediction, and their potential in culture medium recommendation scenarios has not yet been explored. In addition, the high-dimensional features output by large-scale genome models contain a large amount of redundant information, and traditional unsupervised dimensionality reduction methods may lose key task-relevant feature patterns.

[0006] In matching modeling, existing methods typically employ simple feature concatenation and feedforward neural networks, which struggle to capture the complex interactions between microbial features and culture medium features. In reality, microbial responses to culture medium components are highly nonlinear, with synergistic or antagonistic effects existing between different nutrients, necessitating the modeling of multi-order feature interactions. Furthermore, culture medium recommendation is essentially a ranking problem, but existing methods often use regression or classification loss functions, failing to directly optimize the ranking quality of the recommendation list, resulting in unsatisfactory recommendation results.

[0007] Therefore, there is an urgent need to develop a novel culture medium recommendation method that effectively integrates the characterization capabilities of large genome models, models the complex matching relationship between microorganisms and culture media through an advanced deep learning architecture, and adopts an optimization strategy oriented towards sorting tasks, thereby significantly improving recommendation accuracy. Summary of the Invention

[0008] In view of this, the present invention provides a microbial culture medium recommendation method based on a large gene model embedding. The purpose is to extract deep semantic features of microbial genomes using a pre-trained large genome language model, compress high-dimensional genome embeddings into low-dimensional discriminative representations through a supervised learning feature dimensionality reduction network, explicitly model the multi-order interactions between microbial features and culture medium features using a factorization machine, and optimize the quality of the recommendation list using a ranking learning loss function, thereby achieving accurate and efficient culture medium formulation recommendation.

[0009] To achieve the above objectives, the present invention provides a method for recommending microbial culture media based on large gene model embedding, comprising the following steps: A1: Input the whole genome sequence data of the target microorganism to be recommended culture medium into a pre-trained genome big language model to obtain genome feature vectors; A2: Input the genome feature vector into a stacked denoising autoencoder network for feature compression and denoising. Obtain the low-dimensional microbial strain feature vector by jointly optimizing the mean square error reconstruction loss function and the manifold regularization constraint term. A3: Combine the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a set of paired features between microbial strains and culture medium components. A4: Input the paired feature set into a deep metric learning network based on bilinear attention mechanism, calculate the second-order interaction between strain features and culture medium features and train using triplet boundary sorting loss function to obtain the fit score between each candidate culture medium and the target microorganism. A5: Sort all candidate culture media fit scores from high to low, extract a preset number of records from the front of the sort list, and output the optimal culture media recommended formula for the target microorganism.

[0010] As a further improvement of the present invention: Optionally, step A1 further includes: Obtain the whole genome sequence data of the target microorganism for the recommended culture medium, wherein the whole genome sequence data includes the complete DNA sequence information of the target microorganism; preprocess the whole genome sequence data, wherein the preprocessing includes sequence segmentation, marker encoding and position encoding, specifically: the whole genome sequence data is segmented into gene sequence fragments according to a fixed length, each gene sequence fragment is converted into a numerical marker sequence using a predefined nucleotide vocabulary, and position encoding information is added to each marker sequence to preserve the sequence order characteristics; The preprocessed whole genome sequence is input into a pre-trained genome large language model. The pre-trained genome large language model adopts a multi-layer Transformer architecture, including multiple stacked self-attention layers and feedforward neural network layers. The self-attention layer calculates the attention weights between the positional markers in the preprocessed whole genome sequence through a query-key-value mechanism, which is used to capture the long-range dependencies and contextual semantic information of the gene sequence. The pre-trained genome-wide language model encodes the input labeled sequence layer by layer, and finally extracts feature representations from the output layer of the pre-trained genome-wide language model. The feature representations of all gene sequence fragments in the output layer are aggregated through average pooling to obtain a genome feature vector. The dimension of the genome feature vector is... .

[0011] Optionally, step A2 further includes: A stacked noise reduction autoencoder network is constructed, comprising an encoder part and a decoder part. The network structure of the encoder part is determined, comprising multiple fully connected layers, wherein the number of neurons in each fully connected layer decreases sequentially, thereby progressively compressing the genomic feature vector obtained in step A1 into a low-dimensional microbial strain feature vector. The network structure of the decoder part is determined, comprising multiple fully connected layers, wherein the number of neurons in each fully connected layer increases sequentially, thereby reconstructing the low-dimensional microbial strain feature vector back into a genomic feature vector. The training process of the stacked denoising autoencoder network employs a joint loss function, which includes a mean squared error reconstruction loss term and a manifold regularization constraint term. The formula for calculating the joint loss function is as follows: ; in, Denotes the joint loss function. This represents the mean squared error reconstruction loss term. Represents the manifold regularization weight coefficients. Represents the manifold regularization constraint term; The mean square error reconstruction loss term The formula used to measure the difference between the input genomic feature vector and the feature vector reconstructed by the decoder is as follows: ; in, Indicates the number of training samples. Indicates the first The genomic feature vector of each training sample, Indicates the first The genome feature vectors are reconstructed from training samples using a stacked denoising autoencoder network. Represents the square of the L2 norm; The manifold regularization constraint term To preserve the local geometric structure of the data on the low-dimensional manifold, the feature learning process is constrained by minimizing the weighted sum of squared distances between neighboring samples in the low-dimensional feature space. The calculation formula is as follows: ; in, Indicates the first The training sample and the first The similarity weights of each training sample in the original feature space. Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part; The similarity weight Similarity is calculated based on the Euclidean distance between genomic feature vectors, using a Gaussian kernel function to define the similarity. The calculation formula is as follows: ; in, represents the bandwidth parameter of the Gaussian kernel function; exp represents the natural exponential function. Indicates the first Genomic feature vectors of training samples; By minimizing the joint loss function After training the stacked denoising autoencoder network, the genomic feature vector of the target microorganism obtained in step A1 is input into the trained encoder to obtain a low-dimensional microbial strain feature vector that retains key growth characteristics. The dimension of the low-dimensional microbial strain feature vector is... ,satisfy .

[0012] This step introduces a stacked denoising autoencoder network to perform deep dimensionality reduction on the high-dimensional features output from a large genome model, offering significant advantages over simple linear dimensionality reduction methods. The autoencoder employs an encoder-decoder architecture. The encoder layer progressively compresses high-dimensional genomic features into low-dimensional representations through multiple fully connected layers, while the decoder layer attempts to reconstruct the original features from the low-dimensional representation. This reconstruction constraint ensures the preservation of key information during dimensionality reduction. More importantly, this step designs a joint loss function comprising reconstruction loss and manifold regularization. The reconstruction loss ensures that the reduced features effectively recover the original genomic information, avoiding the loss of important biological features. The manifold regularization constraint, by preserving the local geometry of samples in the low-dimensional space, ensures that genomically similar microorganisms remain close in the reduced feature space. This constraint fully utilizes the inherent manifold structure of microbial genomic data, resulting in learned low-dimensional features that are not only lower in dimensionality and more computationally efficient but also better reflect the true biological similarity relationships between microorganisms.

[0013] Optionally, step A3 further includes: Establish a candidate culture medium library for screening, the candidate culture medium library containing There are 10 candidate culture medium formulations, each containing chemical components and their corresponding concentration information. A chemical component concentration feature vector is extracted from each candidate culture medium formulation. This feature vector includes carbon source concentration, nitrogen source concentration, inorganic salt concentration, trace element concentration, and pH value. All chemical component concentration information is organized into a fixed-dimensional feature vector form to obtain the 10th candidate culture medium formulation. Chemical component concentration feature vector of each candidate culture medium , The dimension of the chemical component concentration feature vector is ; The low-dimensional microbial strain feature vector of the target microorganism obtained in step A2 The chemical composition concentration feature vector of each candidate culture medium in the candidate culture medium library Perform Cartesian product combination, and construct paired feature vectors through vector concatenation operation, the th The formula for calculating each paired feature vector is: ; in, Indicates the target microorganism and the first A paired feature vector consisting of candidate culture media, where [;] represents a vector concatenation operation, and the dimension of the paired feature vector is... ; For all The above splicing operation is performed on each candidate culture medium to obtain a set of paired features between microbial strains and culture medium components. .

[0014] This step, through a systematic feature organization and pairing construction method, effectively integrates microbial features with culture medium features, laying a solid foundation for subsequent intelligent matching and prediction. First, a structured candidate culture medium library is established, standardizing and storing commonly used laboratory culture medium formulations and their chemical composition information. This unified feature representation ensures the comparability between different culture medium formulations, enabling the model to systematically learn the effects of various chemical components on microbial growth. By extracting chemical component concentration feature vectors, this step transforms the complex formulation information of the culture medium into a numerical representation that can be processed by the machine learning model. This retains the quantitative concentration information of each nutrient while facilitating subsequent mathematical operations and pattern recognition by the model.

[0015] Optionally, step A4 further includes: A deep metric learning network based on a bilinear attention mechanism is constructed. The deep metric learning network includes a bilinear interaction layer and a fitness prediction layer. The bilinear interaction layer is used to calculate the second-order interaction between the feature vector of low-dimensional microbial strains and the feature vector of chemical component concentration. By introducing a learnable bilinear weight matrix, the nonlinear matching relationship between strain features and culture medium features is captured. The output calculation formula for the bilinear interaction layer is as follows: ; in, Indicates the target microorganism and the first Bilinear interaction characteristics of candidate culture media This represents the transpose of the feature vector of a low-dimensional microbial strain. This represents the bilinear weight matrix, with dimensions of . ; The fitness prediction layer maps bilinear interaction features to fitness scores, and performs a nonlinear transformation using the sigmoid function. The formula for calculating the compatibility score between the candidate culture medium and the target microorganism is as follows: ; in, Indicates the first The compatibility score between each candidate culture medium and the target microorganism; The deep metric learning network is trained using the triplet boundary ranking loss function. This loss function optimizes the network parameters by constructing triples containing anchor samples, positive samples, and negative samples. The formula for calculating the triplet boundary ranking loss function is as follows: ; in, This represents the boundary sorting loss function for triples. This indicates the number of triples used during training. Indicates the first The compatibility score between the positive sample culture medium and the anchor microorganism in each triplet. The positive sample culture medium refers to a culture medium known to be able to successfully culture the anchor microorganism. Indicates the first The fit score between the negative sample culture medium and the anchor microorganism in each triplet. The negative sample culture medium refers to the culture medium known to be unable to successfully culture the anchor microorganism. Indicates the boundary spacing parameter; It is a function for maximizing the value; By minimizing the triplet boundary sorting loss function Complete the training of the deep metric learning network. After training, combine all the paired feature vectors obtained in step A3. The samples are fed into a pre-trained deep metric learning network to obtain the fit score between each candidate culture medium and the target microorganism. .

[0016] Optionally, step A5 further includes: The compatibility scores between all candidate culture media and the target microorganisms obtained in step A4 The fitness scores are sorted in descending order to obtain the sorted candidate culture medium index sequence; a threshold for the number of recommended culture media is set. Select the first from the sorted candidate culture medium index sequence The candidate culture medium formulations corresponding to each index; output the previous... The candidate culture medium formulations are recommended as the optimal culture medium for the target microorganism. The recommendation results include the chemical composition, component concentration information and corresponding fitness score of each recommended culture medium.

[0017] This invention also discloses a microbial culture medium recommendation system based on large gene model embedding, comprising: Large model embedding module: Input the whole genome sequence data of the target microorganism to be recommended culture medium into the pre-trained genome large language model to obtain genome feature vectors; Dimensionality reduction module: Input the genome feature vector into the stacked denoising autoencoder network for feature compression and denoising. By jointly optimizing the mean square error reconstruction loss function and the manifold regularization constraint, the low-dimensional microbial strain feature vector is obtained. Feature Combination Module: Combines the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a set of paired features between microbial strains and culture medium components by performing a Cartesian product. Fit calculation module: The paired feature set is input into a deep metric learning network based on bilinear attention mechanism. The second-order interaction between strain features and culture medium features is calculated and trained using triplet boundary sorting loss function to obtain the fit score between each candidate culture medium and the target microorganism. Recommendation module: Sort all candidate culture media fit scores in descending order, extract a preset number of records from the front of the sorted list, and output the optimal culture media recommended formula for the target microorganism.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: This invention employs a pre-trained genome-wide language model to extract features from microbial whole-genome sequences, effectively capturing deep semantic information and long-range dependencies such as evolutionary conservation, functional patterns, and regulatory signals within gene sequences. Compared to traditional shallow feature engineering methods like k-mer frequency statistics and gene function counting, the genome-wide model, through unsupervised pre-training on massive amounts of genomic data, learns universal gene sequence representations, exhibiting stronger feature expression capabilities and generalization performance. This deep learning-based feature representation method does not rely on manually designed feature engineering, automatically discovering genomic patterns related to microbial physiological characteristics, providing high-quality input features for subsequent culture medium recommendations, and significantly improving the model's accuracy in predicting microbial nutritional requirements.

[0019] This invention designs a supervised learning feature dimensionality reduction network that effectively compresses the high-dimensional embedding vectors output by large-scale genome models into low-dimensional discriminative representations. This dimensionality reduction network, through end-to-end joint training, retains the most important feature information for the culture medium recommendation task during the dimensionality reduction process, while filtering out redundant and noisy components. Compared to traditional unsupervised dimensionality reduction methods such as principal component analysis, the supervised dimensionality reduction strategy of this invention can learn task-oriented nonlinear feature transformations, ensuring that the compressed low-dimensional features have stronger discriminative power. This not only significantly reduces the computational complexity and number of parameters in subsequent models, improving training and inference efficiency, but also enhances the model's generalization ability, avoids overfitting problems in high-dimensional feature spaces, and enables the recommendation system to better handle novel microorganisms and unseen culture medium formulation combinations. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of a microbial culture medium recommendation method based on large gene model embedding, according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0022] Example 1: Recommendation method for microbial culture media based on large gene model embedding, such as Figure 1 As shown, it includes the following steps: A1: Input the whole genome sequence data of the target microorganism for the recommended culture medium into a pre-trained genome big language model to obtain genome feature vectors, including: The whole genome sequence data of the target microorganism for the recommended culture medium is obtained, including the complete DNA sequence information of the target microorganism. The whole genome sequence data is preprocessed, including sequence segmentation, marker encoding, and position encoding. Specifically, the whole genome sequence data is segmented into gene sequence fragments of fixed length; each gene sequence fragment is converted into a numerical marker sequence using a predefined nucleotide vocabulary; and position encoding information is added to each marker sequence to preserve the sequence order. In this embodiment, the fixed length is set to 512 nucleotide base pairs. In this embodiment, the predefined nucleotide vocabulary contains five basic markers: adenine A (corresponding to a value of 1), thymine T (corresponding to a value of 2), guanine G (corresponding to a value of 3), cytosine C (corresponding to a value of 4), and an unknown base N (corresponding to a value of 0). In this embodiment, the position encoding uses a sine-cosine position encoding method. For the marker at position pos in the gene sequence fragment, its position encoding vector... The Dimensions and the The formulas for calculating dimensions are as follows: ; in, This indicates the position index of the marker within the gene sequence segment. , This represents the dimension index of the position encoding vector. , This represents the total dimension of the positional encoding vector, in this embodiment. sin represents the sine function, and cos represents the cosine function; As an alternative implementation, the position encoding can employ a learnable position embedding method, by constructing a position embedding matrix. ,in Indicates the maximum supported sequence length, for the th sequence in a gene sequence fragment. The tag at the i-th position is extracted directly from the position embedding matrix. The row vector serves as the location encoding vector for that position; The preprocessed whole genome sequence is input into a pre-trained genome language model. The pre-trained genome language model adopts a multi-layer Transformer architecture, including multiple stacked self-attention layers and feedforward neural network layers. The self-attention layers calculate the attention weights between positional markers in the preprocessed whole genome sequence through a query-key-value mechanism, which is used to capture long-range dependencies and contextual semantic information of gene sequences. In this embodiment, the pre-trained genome language model uses the existing open-source genome language model DNABERT-2. The pre-trained genome-wide language model encodes the input labeled sequence layer by layer, and finally extracts feature representations from the output layer of the pre-trained genome-wide language model. The feature representations of all gene sequence fragments in the output layer are aggregated through average pooling to obtain a genome feature vector. The dimension of the genome feature vector is... .

[0023] A2: The genomic feature vector is input into a stacked denoising autoencoder network for feature compression and denoising. By jointly optimizing the mean squared error reconstruction loss function and the manifold regularization constraint, a low-dimensional microbial strain feature vector is obtained, including: A stacked denoising autoencoder network is constructed, comprising an encoder part and a decoder part. The network structure of the encoder part is determined, comprising multiple fully connected layers, with the number of neurons in each fully connected layer decreasing sequentially, progressively compressing the genomic feature vector obtained in step A1 into a low-dimensional microbial strain feature vector. The network structure of the decoder part is also determined, comprising multiple fully connected layers, with the number of neurons in each fully connected layer increasing sequentially, reconstructing the low-dimensional microbial strain feature vector back into a genomic feature vector. In this embodiment, the encoder part comprises four fully connected layers with 512, 256, 128, and [missing information - likely referring to specific neuron counts]. ,in To determine the dimension of the low-dimensional microbial strain feature vector, the ReLU activation function is used between each fully connected layer, and the last layer directly outputs the low-dimensional microbial strain feature vector without using an activation function. In this embodiment, the network structure of the decoder part is symmetrical to that of the encoder part. The decoder includes four fully connected layers with 128, 256, 512 and 768 neurons, respectively. The ReLU activation function is used between each fully connected layer, and the last layer directly outputs the reconstructed genome feature vector without using an activation function. The training process of the stacked denoising autoencoder network employs a joint loss function, which includes a mean squared error reconstruction loss term and a manifold regularization constraint term. The formula for calculating the joint loss function is as follows: ; in, Denotes the joint loss function. This represents the mean squared error reconstruction loss term. Represents the manifold regularization weight coefficients. This represents the manifold regularization constraint term; in this embodiment, ; The mean square error reconstruction loss term The formula used to measure the difference between the input genomic feature vector and the feature vector reconstructed by the decoder is as follows: ; in, Indicates the number of training samples. Indicates the first The genomic feature vector of each training sample, Indicates the first The genome feature vectors are reconstructed from training samples using a stacked denoising autoencoder network. Represents the square of the L2 norm; As an alternative implementation, when the training data contains genomic sequence samples of different quality levels, the mean squared error reconstruction loss term can be weighted, giving greater weight to high-quality samples. The calculation formula is as follows: ; in, Indicates the first The quality weights of each training sample are based on the sample's genome integrity index. Calculation, the calculation formula is: ; The manifold regularization constraint term To preserve the local geometric structure of the data on the low-dimensional manifold, the feature learning process is constrained by minimizing the weighted sum of squared distances between neighboring samples in the low-dimensional feature space. The calculation formula is as follows: ; in, Indicates the first The training sample and the first The similarity weights of each training sample in the original feature space. Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part; The similarity weight Similarity is calculated based on the Euclidean distance between genomic feature vectors, using a Gaussian kernel function to define the similarity. The calculation formula is as follows: ; in, The bandwidth parameter representing the Gaussian kernel function is shown in this embodiment. ;exp represents the natural exponential function. Indicates the first Genomic feature vectors of training samples; By minimizing the joint loss function After training the stacked denoising autoencoder network, the genomic feature vector of the target microorganism obtained in step A1 is input into the trained encoder to obtain a low-dimensional microbial strain feature vector that retains key growth characteristics. The dimension of the low-dimensional microbial strain feature vector is... ,satisfy In this embodiment, the stacked noise reduction autoencoder network is trained using the Adam optimizer, with a learning rate of 0.001, a batch size of 64, and 100 training epochs.

[0024] A3: Combine the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a paired feature set of microbial strains and culture medium components, including: Establish a candidate culture medium library for screening, the candidate culture medium library containing There are 10 candidate culture medium formulations, each containing chemical components and their corresponding concentration information. A chemical component concentration feature vector is extracted from each candidate culture medium formulation. This feature vector includes carbon source concentration, nitrogen source concentration, inorganic salt concentration, trace element concentration, and pH value. All chemical component concentration information is organized into a fixed-dimensional feature vector form to obtain the 10th candidate culture medium formulation. Chemical component concentration feature vector of each candidate culture medium , The dimension of the chemical component concentration feature vector is In this embodiment, the specific composition of the chemical component concentration feature vector is as follows: the carbon source concentration part includes five features: glucose concentration, fructose concentration, sucrose concentration, glycerol concentration, and sodium acetate concentration; the nitrogen source concentration part includes five features: ammonium sulfate concentration, sodium nitrate concentration, urea concentration, peptone concentration, and yeast extract concentration; the inorganic salt concentration part includes four features: potassium dihydrogen phosphate concentration, magnesium sulfate concentration, sodium chloride concentration, and calcium chloride concentration; the trace element concentration part includes four features: ferrous sulfate concentration, zinc sulfate concentration, manganese sulfate concentration, and sodium molybdate concentration; pH value is used as one feature. In this embodiment, the chemical component concentration feature vector is normalized using the minimum-maximum normalization method. The low-dimensional microbial strain feature vector of the target microorganism obtained in step A2 The chemical composition concentration feature vector of each candidate culture medium in the candidate culture medium library Perform Cartesian product combination, and construct paired feature vectors through vector concatenation operation, the th The formula for calculating each paired feature vector is: ; in, Indicates the target microorganism and the first A paired feature vector consisting of candidate culture media, where [;] represents a vector concatenation operation, and the dimension of the paired feature vector is... ; For all The above splicing operation is performed on each candidate culture medium to obtain a set of paired features between microbial strains and culture medium components. .

[0025] A4: The paired feature set is input into a deep metric learning network based on a bilinear attention mechanism. The network is trained by calculating the second-order interaction between strain features and culture medium features and using a triplet boundary ranking loss function to obtain the fit score between each candidate culture medium and the target microorganism, including: A deep metric learning network based on a bilinear attention mechanism is constructed. The deep metric learning network includes a bilinear interaction layer and a fitness prediction layer. The bilinear interaction layer is used to calculate the second-order interaction between the feature vector of low-dimensional microbial strains and the feature vector of chemical component concentration. By introducing a learnable bilinear weight matrix, the nonlinear matching relationship between strain features and culture medium features is captured. The output calculation formula for the bilinear interaction layer is as follows: ; in, Indicates the target microorganism and the first Bilinear interaction characteristics of candidate culture media This represents the transpose of the feature vector of a low-dimensional microbial strain. This represents the bilinear weight matrix, with dimensions of . ; The fitness prediction layer maps bilinear interaction features to fitness scores, and performs a nonlinear transformation using the sigmoid function. The formula for calculating the compatibility score between the candidate culture medium and the target microorganism is as follows: ; in, Indicates the first The compatibility score between each candidate culture medium and the target microorganism; The deep metric learning network is trained using the triplet boundary ranking loss function. This loss function optimizes the network parameters by constructing triples containing anchor samples, positive samples, and negative samples. The formula for calculating the triplet boundary ranking loss function is as follows: ; in, This represents the boundary sorting loss function for triples. This indicates the number of triples used during training. Indicates the first The compatibility score between the positive sample culture medium and the anchor microorganism in each triplet. The positive sample culture medium refers to a culture medium known to be able to successfully culture the anchor microorganism. Indicates the first The fit score between the negative sample culture medium and the anchor microorganism in each triplet. The negative sample culture medium refers to the culture medium known to be unable to successfully culture the anchor microorganism. In this embodiment, the boundary spacing parameter is represented. ; The maximum value function is used. In this embodiment, the triplet is constructed as follows: For each microbial sample in the training dataset as the anchor microorganism, a positive sample culture medium is randomly selected from the known successful culture medium formulations of the microorganism, and a negative sample culture medium is randomly selected from the candidate culture medium library that is not in the list of known successful culture media of the microorganism, thereby forming a triplet. By minimizing the triplet boundary sorting loss function Complete the training of the deep metric learning network. After training, combine all the paired feature vectors obtained in step A3. The samples are fed into a pre-trained deep metric learning network to obtain the fit score between each candidate culture medium and the target microorganism. In this embodiment, the deep metric learning network is trained using the Adam optimizer, with a learning rate of 0.0005, a batch size of 128, and 50 training epochs.

[0026] A5: Sort all candidate culture media fit scores from highest to lowest, extract a preset number of records from the top of the sorted list, and output the recommended optimal culture medium formulation for the target microorganism, including: The compatibility scores between all candidate culture media and the target microorganisms obtained in step A4 The fitness scores are sorted in descending order to obtain the sorted candidate culture medium index sequence; a threshold for the number of recommended culture media is set. Select the first from the sorted candidate culture medium index sequence The candidate culture medium formulations corresponding to each index; output the previous... The candidate culture medium formulations are recommended as the optimal culture medium for the target microorganism. The recommendations include the chemical composition, component concentration information, and corresponding fitness score for each recommended culture medium. In this embodiment, In this embodiment, the recommendation results are output in the form of structured data. Each recommended culture medium formula includes the following information: culture medium number, carbon source type and concentration, nitrogen source type and concentration, inorganic salt type and concentration, trace element type and concentration, pH value, and suitability score.

[0027] Example 2: This invention also discloses a microbial culture medium recommendation system based on large gene model embedding, comprising the following five modules: Large model embedding module: Input the whole genome sequence data of the target microorganism to be recommended culture medium into the pre-trained genome large language model to obtain genome feature vectors; Dimensionality reduction module: Input the genome feature vector into the stacked denoising autoencoder network for feature compression and denoising. By jointly optimizing the mean square error reconstruction loss function and the manifold regularization constraint, the low-dimensional microbial strain feature vector is obtained. Feature Combination Module: Combines the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a set of paired features between microbial strains and culture medium components by performing a Cartesian product. Fit calculation module: The paired feature set is input into a deep metric learning network based on bilinear attention mechanism. The second-order interaction between strain features and culture medium features is calculated and trained using triplet boundary sorting loss function to obtain the fit score between each candidate culture medium and the target microorganism. Recommendation module: Sort all candidate culture media fit scores in descending order, extract a preset number of records from the front of the sorted list, and output the optimal culture media recommended formula for the target microorganism.

[0028] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0029] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0030] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for recommending microbial culture media based on large gene model embedding, characterized in that, Includes the following steps: A1: Input the whole genome sequence data of the target microorganism to be recommended culture medium into a pre-trained genome big language model to obtain genome feature vectors; A2: Input the genome feature vector into a stacked denoising autoencoder network for feature compression and denoising. Obtain the low-dimensional microbial strain feature vector by jointly optimizing the mean square error reconstruction loss function and the manifold regularization constraint term. A3: Combine the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a set of paired features between microbial strains and culture medium components. A4: Input the paired feature set into a deep metric learning network based on bilinear attention mechanism, calculate the second-order interaction between strain features and culture medium features and train using triplet boundary sorting loss function to obtain the fit score between each candidate culture medium and the target microorganism. A5: Sort all candidate culture media fit scores from high to low, extract a preset number of records from the front of the sort list, and output the optimal culture media recommended formula for the target microorganism.

2. The method for recommending microbial culture media based on large gene model embedding according to claim 1, characterized in that, Step A1 includes: Obtain the whole genome sequence data of the target microorganism for the recommended culture medium, wherein the whole genome sequence data includes the complete DNA sequence information of the target microorganism; preprocess the whole genome sequence data, wherein the preprocessing includes sequence segmentation, marker encoding and position encoding, specifically: the whole genome sequence data is segmented into gene sequence fragments according to a fixed length, each gene sequence fragment is converted into a numerical marker sequence using a predefined nucleotide vocabulary, and position encoding information is added to each marker sequence to preserve the sequence order characteristics; The preprocessed whole genome sequence is input into a pre-trained genome large language model. The pre-trained genome large language model adopts a multi-layer Transformer architecture, including multiple stacked self-attention layers and feedforward neural network layers. The self-attention layer calculates the attention weights between the positional markers in the preprocessed whole genome sequence through a query-key-value mechanism, which is used to capture the long-range dependencies and contextual semantic information of the gene sequence. The pre-trained genome-wide language model encodes the input labeled sequence layer by layer, and finally extracts feature representations from the output layer of the pre-trained genome-wide language model. The feature representations of all gene sequence fragments in the output layer are aggregated through average pooling to obtain a genome feature vector. The dimension of the genome feature vector is... .

3. The method for recommending microbial culture media based on large gene model embedding according to claim 2, characterized in that, Step A2 includes: A stacked noise reduction autoencoder network is constructed, comprising an encoder part and a decoder part. The network structure of the encoder part is determined, comprising multiple fully connected layers, wherein the number of neurons in each fully connected layer decreases sequentially, thereby progressively compressing the genomic feature vector obtained in step A1 into a low-dimensional microbial strain feature vector. The network structure of the decoder part is determined, comprising multiple fully connected layers, wherein the number of neurons in each fully connected layer increases sequentially, thereby reconstructing the low-dimensional microbial strain feature vector back into a genomic feature vector. The training process of the stacked denoising autoencoder network employs a joint loss function, which includes a mean squared error reconstruction loss term and a manifold regularization constraint term. The formula for calculating the joint loss function is as follows: ； in, Denotes the joint loss function. This represents the mean squared error reconstruction loss term. Represents the manifold regularization weight coefficients. Represents the manifold regularization constraint term; The mean square error reconstruction loss term The formula used to measure the difference between the input genomic feature vector and the feature vector reconstructed by the decoder is as follows: ； in, Indicates the number of training samples. Indicates the first The genomic feature vector of each training sample, Indicates the first The genome feature vectors are reconstructed from training samples using a stacked denoising autoencoder network. Represents the square of the L2 norm; The manifold regularization constraint term To preserve the local geometric structure of the data on the low-dimensional manifold, the feature learning process is constrained by minimizing the weighted sum of squared distances between neighboring samples in the low-dimensional feature space. The calculation formula is as follows: ； in, Indicates the first The training sample and the first The similarity weights of each training sample in the original feature space. Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part Indicates the first The low-dimensional microbial strain feature vectors obtained from each training sample through the encoder part; The similarity weight Similarity is calculated based on the Euclidean distance between genomic feature vectors, using a Gaussian kernel function to define the similarity. The calculation formula is as follows: ； in, represents the bandwidth parameter of the Gaussian kernel function; exp represents the natural exponential function. Indicates the first Genomic feature vectors of training samples; By minimizing the joint loss function After training the stacked denoising autoencoder network, the genomic feature vector of the target microorganism obtained in step A1 is input into the trained encoder to obtain a low-dimensional microbial strain feature vector that retains key growth characteristics. The dimension of the low-dimensional microbial strain feature vector is... ,satisfy .

4. The method for recommending microbial culture media based on large gene model embedding according to claim 3, characterized in that, Step A3 includes: Establish a candidate culture medium library for screening, the candidate culture medium library containing There are 10 candidate culture medium formulations, each containing chemical components and their corresponding concentration information. A chemical component concentration feature vector is extracted from each candidate culture medium formulation. This feature vector includes carbon source concentration, nitrogen source concentration, inorganic salt concentration, trace element concentration, and pH value. All chemical component concentration information is organized into a fixed-dimensional feature vector form to obtain the 10th candidate culture medium formulation. Chemical component concentration feature vector of each candidate culture medium , The dimension of the chemical component concentration feature vector is ; The low-dimensional microbial strain feature vector of the target microorganism obtained in step A2 The chemical composition concentration feature vector of each candidate culture medium in the candidate culture medium library Perform Cartesian product combination, and construct paired feature vectors through vector concatenation operation, the th The formula for calculating each paired feature vector is: ； in, Indicates the target microorganism and the first A paired feature vector consisting of candidate culture media, where [;] represents a vector concatenation operation, and the dimension of the paired feature vector is... ; For all The above splicing operation is performed on each candidate culture medium to obtain a set of paired features between microbial strains and culture medium components. .

5. The method for recommending microbial culture media based on large gene model embedding according to claim 4, characterized in that, Step A4 includes: A deep metric learning network based on a bilinear attention mechanism is constructed. The deep metric learning network includes a bilinear interaction layer and a fitness prediction layer. The bilinear interaction layer is used to calculate the second-order interaction between the feature vector of low-dimensional microbial strains and the feature vector of chemical component concentration. By introducing a learnable bilinear weight matrix, the nonlinear matching relationship between strain features and culture medium features is captured. The output calculation formula for the bilinear interaction layer is as follows: ； in, Indicates the target microorganism and the first Bilinear interaction characteristics of candidate culture media This represents the transpose of the feature vector of a low-dimensional microbial strain. This represents the bilinear weight matrix, with dimensions of . ; The fitness prediction layer maps bilinear interaction features to fitness scores, and performs a nonlinear transformation using the sigmoid function. The formula for calculating the compatibility score between the candidate culture medium and the target microorganism is as follows: ； in, Indicates the first The compatibility score between each candidate culture medium and the target microorganism; The deep metric learning network is trained using the triplet boundary ranking loss function. This loss function optimizes the network parameters by constructing triples containing anchor samples, positive samples, and negative samples. The formula for calculating the triplet boundary ranking loss function is as follows: ； in, This represents the boundary sorting loss function for triples. This indicates the number of triples used during training. Indicates the first The compatibility score between the positive sample culture medium and the anchor microorganism in each triplet. The positive sample culture medium refers to a culture medium known to be able to successfully culture the anchor microorganism. Indicates the first The fit score between the negative sample culture medium and the anchor microorganism in each triplet. The negative sample culture medium refers to the culture medium known to be unable to successfully culture the anchor microorganism. Indicates the boundary spacing parameter; It is a function for maximizing the value; By minimizing the triplet boundary sorting loss function Complete the training of the deep metric learning network. After training, combine all the paired feature vectors obtained in step A3. The samples are fed into a pre-trained deep metric learning network to obtain the fit score between each candidate culture medium and the target microorganism. .

6. The method for recommending microbial culture media based on large gene model embedding according to claim 5, characterized in that, Step A5 includes: The compatibility scores between all candidate culture media and the target microorganisms obtained in step A4 The fitness scores are sorted in descending order to obtain the sorted candidate culture medium index sequence; a threshold for the number of recommended culture media is set. Select the first from the sorted candidate culture medium index sequence The candidate culture medium formulations corresponding to each index; output the previous... The candidate culture medium formulations are recommended as the optimal culture medium for the target microorganism. The recommendation results include the chemical composition, component concentration information and corresponding fitness score of each recommended culture medium.

7. A microbial culture medium recommendation system based on large gene model embedding, characterized in that, include: Large model embedding module: Input the whole genome sequence data of the target microorganism to be recommended culture medium into the pre-trained genome large language model to obtain genome feature vectors; Dimensionality reduction module: Input the genome feature vector into the stacked denoising autoencoder network for feature compression and denoising. By jointly optimizing the mean square error reconstruction loss function and the manifold regularization constraint, the low-dimensional microbial strain feature vector is obtained. Feature Combination Module: Combines the feature vectors of low-dimensional microbial strains with the chemical component concentration feature vectors of each culture medium in the candidate culture medium library to construct a set of paired features between microbial strains and culture medium components by performing a Cartesian product. Fit calculation module: The paired feature set is input into a deep metric learning network based on bilinear attention mechanism. The second-order interaction between strain features and culture medium features is calculated and trained using triplet boundary sorting loss function to obtain the fit score between each candidate culture medium and the target microorganism. Recommendation module: Sort all candidate culture media fit scores from high to low, extract a preset number of records from the front of the sort list, and output the optimal culture media recommended formula for the target microorganism. To achieve the microbial culture medium recommendation method based on large gene model embedding as described in any one of claims 1-6.

Images