Plant secondary metabolic pathway prediction method based on deep transfer learning

Through deep transfer learning and graph transformer neural network, combined with the deep learning model of convolutional neural network, the problem of unspecific prediction of plant secondary metabolic pathways is solved, and high accuracy and reliability prediction of plant secondary metabolic pathways is achieved. The limitation of insufficient database information supports the accuracy of metabolomics research.

CN119028463APending Publication Date: 2024-11-26DALIAN INSTITUTE OF CHEMICAL PHYSICS CHINESE ACADEMY OF SCIENCES +1

Patent Information

Application Number
CN202310590497.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-23
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing prediction methods for plant secondary metabolic pathways are not specific enough, and it is difficult to effectively predict plant secondary metabolic pathways. Due to insufficient database information, there are deviations in the prediction results, which affects the accuracy of metabolomics research.

Method used

Using a deep transfer learning method, a deep learning model is constructed, and the source data set and target data set are constructed for pre-training and fine-tuning, combined with graph transformer neural network and convolutional neural network, the structural characteristics of the compound are extracted and solved through transfer learning. The problem of insufficient data volume can achieve more detailed prediction of metabolic pathways.

Benefits of technology

Accurate prediction of plant secondary metabolic pathways is achieved, the reliability and accuracy of predictions are improved, the problem of insufficient database information is overcome, and the metabolic pathway enrichment analysis in metabolomics research is provided.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028463A_ABST
    Figure CN119028463A_ABST
Patent Text Reader

Abstract

The invention discloses a plant secondary metabolic pathway prediction method based on deep transfer learning, and establishes a plant secondary metabolic pathway prediction method based on deep learning and transfer learning to realize prediction of a plant secondary metabolic pathway of a compound with known structure information. The method comprises the following steps: firstly, carrying out deep learning model pre-training on a metabolic pathway data set (source data set) with richer information; then, through transfer learning, carrying out fine adjustment on the target data set according to the pre-training parameters, and constructing a prediction model of the plant secondary metabolic pathway; and finally, inputting the simplified molecule linear input standard character string of a compound to be predicted into the model, and obtaining predicted plant secondary metabolic pathway information. The method effectively overcomes the problem of insufficient plant secondary metabolism pathway data information at present, and has the advantages of high prediction accuracy, high model training efficiency and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to a method for predicting plant secondary metabolic pathways based on deep transfer learning, and belongs to the technical field of plant secondary metabolic pathway prediction. Background Art

[0002] Plant secondary metabolites are intermediates and products produced by secondary metabolism, mainly including nitrogen-containing organic matter, terpenoids and phenolic compounds. Unlike primary metabolites, secondary metabolites are not essential for the growth and survival of organisms. However, plant secondary metabolites are not only involved in stress resistance and disease resistance, but also the active ingredients of many herbal medicines. Research on plant secondary metabolism can not only guide the breeding of high-quality crops with ideal traits, but also help discover new drugs for the treatment of human diseases.

[0003] Metabolomics (Nicholson, JK et al., Metabonomics: a platform for studying drug toxicity and gene function. Nature Reviews Drug Discovery 2002, 1, 153-161.) is a powerful tool for studying the production and function of secondary metabolites. The goal of metabolomics is to analyze the overall changes of metabolites (molecular weight <1500Da) in biological systems. Differential metabolite analysis based on metabolomics research can effectively discover potential biomarkers. Changes in related metabolic pathways reflect biological mechanisms and are of guiding significance for disease diagnosis, drug development, and crop breeding.

[0004] However, the current understanding of plant secondary metabolic pathways is limited, and some pathways are not yet included in existing databases due to outdated database information and the need for continuous maintenance. Therefore, pathway enrichment analysis based on metabolomics data may be biased, and further research is needed to fully explore the diversity of plant secondary metabolic pathways. However, it is costly to conduct biological experiments to study these pathways. Therefore, it is crucial to develop new methods and technologies to enhance our understanding of plant secondary metabolism and its potential applications.

[0005] Based on traditional machine learning algorithms and deep learning algorithms, some metabolic pathway prediction methods have been reported in the literature.

[0006] Document 1 (A metabolic pathway prediction method, system, terminal device and readable storage medium, application number CN202011509783.8, application date 2020-12-18) developed a method for predicting the metabolic generation reaction of metabolites based on deep learning. By repeatedly generating substrate molecules that produce product molecules, the entire metabolic pathway is predicted. However, the metabolic pathway obtained by this method tends to be reaction flow, making it difficult to perform metabolic pathway enrichment analysis.

[0007] Reference 2 (Cai YD, Qian Z, Lu L, et al. Prediction of compounds'biological function (metabolic pathways) based on functional group composition [J]. Molecular diversity, 2008, 12: 131-137.) reported a method based on the nearest neighbor algorithm in traditional machine learning to predict 11 metabolic pathway categories in the Kyoto Encyclopedia of Genes and Genomes database. However, this method can only predict known compounds in one metabolic pathway category, ignoring the fact that the same compound can participate in many metabolic pathways.

[0008] Document 3 (Baranwal M, Magner A, Elvati P, et al. A deep learning architecture for metabolic pathway prediction [J]. Bioinformatics, 2020, 36(8): 2547-2553.) and document 4 (Du BX, Zhao PC, Zhu B, et al. MLGL-MP: a Multi-Label Graph Learning framework enhanced by pathway interdependence for Metabolic Pathway prediction [J].

[0009] Bioinformatics, 2022, 38(Supplement_1):i325-i332.) built a metabolic pathway prediction model based on graph transformer neural network, which can achieve multi-label prediction tasks. However, like reference 2, they all predict 11 metabolic pathway categories in the Kyoto Encyclopedia of Genes and Genomes database, which is not specific enough and it is difficult to predict more detailed information on plant secondary metabolic pathways.

[0010] Therefore, the current prediction methods are not specific enough for the prediction categories of metabolic pathways, and cannot predict plant secondary metabolic pathways, which is not conducive to the subsequent interpretation of biological functions in plant metabolomics. To solve this problem, the present invention collects the structural information of plant secondary metabolites and the secondary metabolic pathway information of these metabolites, and provides a plant secondary metabolic pathway prediction method based on deep learning and transfer learning methods. Summary of the invention

[0011] According to one aspect of the present application, a method for predicting plant secondary metabolic pathways based on deep transfer learning is provided, which can effectively overcome the problem of insufficient information in the current plant secondary metabolic pathway database, has the advantages of reliable prediction, high accuracy, and high model training efficiency, and can realize the effective prediction of plant secondary metabolic pathways.

[0012] To achieve the above object, the present invention provides the following technical solution: a method for predicting plant secondary metabolic pathways based on deep transfer learning, comprising the following steps:

[0013] S1: Construct the source dataset, namely the metabolic pathway dataset, for pre-training the model;

[0014] Construct a target dataset, i.e., a plant secondary metabolic pathway dataset, to fine-tune the trained model;

[0015] S2: Build a deep learning pre-training model and use the source dataset for training for metabolic pathway prediction;

[0016] Use cross-validation to select the deep learning model parameters with the highest accuracy as the initial parameters for transfer learning;

[0017] S3: Build a deep transfer learning model for metabolic pathway prediction;

[0018] Using the target data set to fine-tune the deep learning pre-training model to obtain a prediction model;

[0019] Use cross-validation to select the deep transfer learning model parameters with the highest accuracy as the final prediction model parameters;

[0020] S4: Obtain the SMILES string of the compound to be predicted, input it into the final prediction model, and obtain the predicted plant secondary metabolic pathway information.

[0021] Optionally, the source data set is constructed based on the Kyoto Encyclopedia of Genes and Genomes database (KEGG) (https: / / www.genome.jp / kegg / ) according to the acquired metabolic pathway category information and the compound structure information contained in the pathway;

[0022] The target dataset is based on the Plant Metabolism Network Database (PWN)

[0023] (https: / / plantcyc.org / ), which is constructed based on the obtained plant secondary metabolic pathway category information and the structural information of the compounds contained in the pathway.

[0024] Optionally, the metabolic pathway category labels in the source data set and the target data set are one-hot encoded;

[0025] In the one-hot encoding, "1" represents that the compound is involved in the pathway, and "0" represents that the compound is not involved in the pathway.

[0026] Optionally, the compound structures in the source data set and the target data set are represented using SMILES strings.

[0027] Optionally, the deep learning pre-trained model is input with a SMILES string in the source data set and output as a metabolic pathway category label in the KEGG database.

[0028] Optionally, the SMILES string in the target data set is used to input the deep learning pre-training model, and the output is a plant secondary metabolic pathway category label.

[0029] Optionally, methods for obtaining the SMILES string of the compound to be predicted include, but are not limited to, querying public structure databases such as Pubchem, and using open source chemical informatics toolkits such as CDK and RDKit for query or conversion.

[0030] Optionally, the deep learning pre-training model is a modular neural network, including a first module and a second module, each module learns separately, extracts molecular structure features, and obtains 1 feature vector respectively, and then merges and learns.

[0031] Optionally, the combined learning step includes:

[0032] A1: concatenate the two feature vectors learned by the first module and the second module respectively to obtain a concatenated feature vector;

[0033] A2: Use a convolutional neural network to learn the result feature vector.

[0034] Optionally, the first module is a graph transformer neural network, and the second module is a convolutional neural network.

[0035] Optionally, the learning rate is lowered, a fine-tuning strategy is adopted, the deep learning pre-trained model parameters are used as initialization parameters, the convolutional neural network parameters of the second module are frozen, the other parts of the deep transfer learning model are retrained using the target data set, and a prediction model is constructed.

[0036] Optionally, the loss function of the deep learning pre-training model uses an ASL function.

[0037] The beneficial effects of this application include:

[0038] 1) The present application provides a method for predicting plant secondary metabolic pathways based on deep transfer learning, which constructs a deep learning model for predicting plant secondary metabolic pathways, realizes more detailed metabolic pathway prediction, effectively overcomes the problem of insufficient information in the current plant secondary metabolic pathway database, and provides assistance for metabolic pathway enrichment analysis in metabolomics research;

[0039] 2) The present application provides a method for predicting plant secondary metabolic pathways based on deep transfer learning. The modular deep learning model using graph transformer neural network and convolutional neural network is constructed to effectively extract the structural characteristics of compounds and effectively solve the problem of insufficient data through transfer learning;

[0040] 3) The present application provides a method for predicting plant secondary metabolic pathways based on deep transfer learning. The constructed plant secondary metabolic pathway prediction model has the advantages of reliable prediction, high accuracy, and high model training efficiency, and can realize the effective prediction of plant secondary metabolic pathways. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of the plant secondary metabolic pathway prediction steps provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present application is described in detail below with reference to embodiments, but the present application is not limited to these embodiments.

[0043] In order to verify the effectiveness and feasibility of the present invention, 577 compounds that were not involved in the model training and updated in January 2023 in the Plant Metabolism Network (PWN) (https: / / plantcyc.org / ) database were screened, and plant secondary metabolic pathways were predicted for these compounds.

[0044] Example

[0045] Figure 1 A flow chart of a method for predicting plant secondary metabolic pathways provided in this application, the method comprising the following steps:

[0046] Step S1: The target dataset and the source dataset need to be constructed.

[0047] The target dataset is constructed based on the PWN database, and the information of 18 plant secondary metabolic pathway categories and 2028 compounds contained therein are obtained. The download time is October 2022. Among them, 11 pathway category information is represented by 18 digital labels, '0', '1', '2', ..., '17', and 2028 compounds are represented by simplified molecular linear input specification (SMILES) strings. The distribution of compound labels in each pathway is shown in Table 1:

[0048] Table 1

[0049]

[0050]

[0051] The source dataset is based on the Kyoto Encyclopedia of Genes and Genomes (KEGG) database.

[0052] (https: / / www.genome.jp / kegg / ), and obtained 11 metabolic pathway category information and the 5288 compound information contained therein, which was downloaded in December 2022. Among them, the 11 pathway category information is represented by 11 digital labels, '0', '1', '2', ..., '10', and the 5288 compounds are represented by simplified molecular linear input specifications (SMILES) strings. The distribution of compound labels in each pathway is shown in Table 2:

[0053] Table 2

[0054]

[0055] Step S2: Construct a deep learning pre-trained model for metabolic pathway prediction, and obtain the pre-trained model by training with the source data set. A modular deep learning model is constructed using a graph transformer neural network and a convolutional neural network to fully extract molecular structure features. The two modules are learned separately to extract molecular structure features and obtain two feature vectors. Then, the two feature vectors obtained by module 1 and module 2 are concatenated to obtain a concatenated feature vector. A convolutional neural network is used to learn this concatenated feature vector. Among them, the model inputs the SMILES string in the source data set and outputs the metabolic pathway category label in the KEGG database. Specifically, in module 1, molecules are converted into graph structures, atoms are nodes, and chemical bonds are edges. Each graph passes through a layer of 8-head attention graph transformer network, a layer of 1-head attention graph transformer network, a layer of global pooling layer, a layer of fully connected layer, and finally outputs a 2000-dimensional vector. In module 2, the molecule is first input into the pre-trained public Molecular Transformer Embeddings neural network to obtain the molecular compiler embedding, and then passes through a 1D convolution layer (kernel size 3, stride 2), a 1D pooling layer (kernel size 3, stride 2), a fully connected layer, and finally outputs a 2000-dimensional vector. Then, the 2000-dimensional vectors obtained by module 1 and module 2 are concatenated to obtain a 4000-dimensional feature vector, which then passes through a 1D convolution layer (kernel size 3, stride 2), a 1D pooling layer (kernel size 3, stride 2), a fully connected layer, a temporary retreat layer (probability 0.2), a fully connected layer, and finally outputs an 11-dimensional vector. These 11 numbers are recorded as 0 if less than 0.5, indicating that they are not on this pathway label; and recorded as 1 if greater than 0.5, indicating that they are on this pathway label. The Adam optimizer was used for training, with an initial learning rate of 0.0003; the loss function was ASL (parameters gamma_neg = 2, gamma_pos = 1.5) and the number of training times was 300; an early stopping strategy was adopted (parameters patience = 10, delta = 5); cross-validation was used to select the deep learning model with the highest accuracy, and the parameters were saved as the initial parameters for the next step of transfer learning.Here, the model in the present invention is compared with model 1 (based on random forest) and model 2 (based on graph neural network) in document 3 (Baranwal M, Magner A, Elvati P, et al. A deep learning architecture for metabolic pathway prediction [J]. Bioinformatics, 2020, 36 (8): 2547-2553.), and the model in document 4 (Du BX, Zhao PC, Zhu B, et al. MLGL-MP: a Multi-Label Graph Learning framework enhanced by pathway interdependence for Metabolic Pathway prediction [J]. Bioinformatics, 2022, 38 (Supplement_1): i325-i332.) (based on graph neural network). The results show that the model in the present invention performs best in the four indicators of accuracy, precision, recovery rate, and F1 score, and there are statistical differences to varying degrees, as shown in Table 3:

[0056] Table 3

[0057]

[0058]

[0059] The values ​​in the above table represent the mean ± standard deviation obtained by 5-fold cross validation. The symbols *, ** and *** indicate that the performance of the model of the present invention in the t-test is significantly better than the corresponding model when the p value is less than 0.05, 0.01 and 0.001 respectively.

[0060] Step S3: Construct a transfer learning model for metabolic pathway prediction, and use the target dataset to fine-tune the pre-trained model constructed in the second step. The input of the model is the SMILES string in the target dataset, and the output is the plant secondary metabolic pathway category label. Adopt a fine-tuning strategy, use the pre-trained parameters as initialization parameters, freeze the initial parameters of the 1D convolution layer (kernel size is 3, step size is 2) and 1D pooling layer (kernel size is 3, step size is 2) in module 2, adjust the learning rate value to 0.0001, retrain, and construct a plant secondary metabolic pathway prediction model. Use cross-validation to save the model with the highest accuracy as the final plant secondary metabolic pathway prediction model.

[0061] Step S4: Input the SMILES strings of the 577 compounds to be predicted into the plant secondary metabolic pathway prediction model constructed in the third step to obtain the predicted plant secondary metabolic pathway information. Here, the 577 compounds used do not have the same SMILES strings as the 2028 compounds in the target data set in the first step of building the model. The accuracy of the predicted plant secondary metabolic pathway information is 97.82%, which proves the accuracy of the model prediction of the present invention.

[0062] The above results show that the deep learning model constructed by the present invention effectively extracts molecular structure features, and realizes the multi-label prediction task of plant secondary metabolic pathways through transfer learning, and the results are accurate and reliable.

[0063] The above are only a few embodiments of the present application and do not constitute any form of limitation to the present application. Although the present application is disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the profession, without departing from the scope of the technical solution of the present application, using the technical content disclosed above to make slight changes or modifications are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for predicting plant secondary metabolic pathways based on deep transfer learning, characterized in that: The steps include: S1: Construction of source dataset, namely metabolic pathway dataset; Construct the target dataset, i.e., the plant secondary metabolism pathway dataset; S2: Build a deep learning pre-training model and use the source dataset for training; Use cross-validation to select the deep learning model parameters with the highest accuracy as the initial parameters for transfer learning; S3: Build a deep transfer learning model; Fine-tune the deep learning pre-training model using the target data set to obtain a prediction model; Use cross-validation to select the deep transfer learning model parameters with the highest accuracy as the final prediction model parameters; S4: Obtain the SMILES string of the compound to be predicted, input it into the final prediction model, and obtain the predicted plant secondary metabolic pathway information.

2. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: The source data set is constructed based on the KEGG database according to the acquired metabolic pathway category information and the compound structure information contained in the pathway; The target data set is constructed based on the plant metabolic network PWN database according to the acquired plant secondary metabolic pathway category information and the compound structure information contained in the pathway.

3. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: The metabolic pathway category labels in the source data set and the target data set are encoded using one-hot encoding; In the one-hot encoding, "1" represents that the compound is involved in the pathway, and "0" represents that the compound is not involved in the pathway.

4. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: The compound structures in the source data set and the target data set are represented by SMILES strings; Preferably, the deep learning pre-training model is input with a SMILES string in the source data set, and outputs a metabolic pathway category label in the KEGG database; Preferably, the SMILES string in the target data set is used to input the deep learning pre-training model, and the output is a plant secondary metabolic pathway category label.

5. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: Methods for obtaining the SMILES string of the compound to be predicted include querying the Pubchem public structure database and using open source chemical informatics toolkits such as CDK and RDKit for query or conversion.

6. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: The deep learning pre-training model is a modular neural network, including a first module and a second module. Each module learns separately to extract molecular structure features, and after obtaining a feature vector respectively, they are merged for learning.

7. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 6, characterized in that: The first module is a graph transformer neural network, and the second module is a convolutional neural network.

8. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 7, characterized in that: The learning rate is reduced, the deep learning pre-trained model parameters are used as initialization parameters, the convolutional neural network parameters of the second module are frozen, the other parts of the deep transfer learning model are retrained using the target data set, and a prediction model is constructed.

9. A method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 6, characterized in that: The combined learning step includes: A1: concatenate the two feature vectors learned by the first module and the second module respectively to obtain a concatenated feature vector; A2: Use a convolutional neural network to learn the result feature vector.

10. The method for predicting plant secondary metabolic pathways based on deep transfer learning according to claim 1, characterized in that: The deep learning pre-training model uses the ASL function as the loss function.

Citation Information

Patent Citations

  • Metabolic pathway prediction method and system, terminal equipment and readable storage medium

    CN112530516A

  • Method for identifying hot spot residues on interface of protein and RNA (Ribonucleic Acid) compound based on transfer learning

    CN112927753A

  • Ionic liquid type antibiotic drug property prediction method based on transfer learning and graph neural network and high-throughput screening platform

    CN114220497A

  • Metabolic pathway prediction method based on label correlation and graph representation learning

    CN114927173A

  • Drug molecule skeleton replacing and screening method based on deep transfer learning model

    CN115881244A

Cited By

  • Method and system for regulating and controlling nitrogen circulation of constructed wetland by using root exudates

    CN120887552A

  • Method and system for predicting enzyme turnover number by fusing multi-modal characteristics of enzymatic reaction

    CN121483405A

  • A method and system for predicting enzyme turnover based on multimodal characteristics of enzyme-catalyzed reactions

    CN121483405B