Construction of protein expression system optimization model based on XGBoost algorithm
By constructing a preferred model of protein expression system based on XGBoost algorithm, the uncertainty problem of protein expression system selection is solved, the protein expression efficiency and protein solubility are improved, and the R&D cost is reduced.
Patent Information
- Application Number
- CN202510551248.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-12
AI Technical Summary
The lack of clear methods in the prior art to select the best protein expression system, resulting in waste of resources and protein insoluble problems, affecting protein activity and production efficiency.
A preferred model of protein expression system based on the XGBoost algorithm was constructed. By collecting and processing protein sequence data of E. coli, Pichia cerevisiae, Fattia meadow and Chinese hamster ovarian cells, feature engineering and model optimization were carried out to establish an evaluation tool for the adaptability of target proteins and expression systems.
It improves protein expression efficiency, reduces the cost of protein drugs and vaccine research and development, provides guidance on the selection of protein expression systems, and ensures protein soluble and biological activity.
Smart Images
Figure CN120472983A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to bioinformatics, protein engineering and machine learning, and specifically to the development and application of a protein expression system optimization method based on the XGBoost algorithm. Background Art
[0002] Currently, various protein expression systems have been widely used in industrial and medical fields, including in vaccine preparation, drug and agricultural product research and development industries. Currently, researchers mainly use prokaryotic cells and eukaryotic cells to express proteins. Prokaryotic expression systems are widely used due to their rapid proliferation, low nutritional requirements, high expression levels, ease of transformation and low equipment investment. However, due to their limited ability to post-translationally modify proteins, this may result in the target protein not being able to fold correctly or forming inclusion bodies. While eukaryotic expression systems can perform more complex post-translational modifications and produce more complex proteins, they have risks such as long production cycles, reduced virulence, and easy contamination. Therefore, in practical applications, they cannot meet experimental and production needs well. How to rationally select a biological expression system for the target protein is an important link in protein engineering.
[0003] In practice, there's no clear basis for selecting the optimal expression system, so protein expression currently relies on trial-and-error experiments, resulting in significant waste of resources. In theory, an expression system typically exhibits good compatibility with its intended target protein, effectively avoiding the problems associated with blindly selecting an expression system. Furthermore, when the target protein's properties closely match those of the proteins within the expression system, the system will be more favorable for its expression and help maintain protein homeostasis in solution. Maintaining protein solubility is a fundamental aspect of protein homeostasis and a prerequisite for maintaining protein activity, profoundly impacting the expression and purification of therapeutic proteins. Solubility is primarily determined by the primary sequence and environmental conditions, encompassing properties such as the protein's amino acid composition, charge, and polarity, as well as its ability to adapt and interact with its surroundings. Producing soluble and functional proteins is a major goal in biotechnology. Due to the inability to achieve satisfactory yields from natural sources, expression systems are often used to produce target proteins. However, proteins expressed in these systems often face the challenge of insolubility. Although protein aggregate treatment can sometimes yield higher-purity proteins, protein refolding is a significant technical challenge and often results in loss of biological activity. Therefore, an expression system that matches the protein's characteristics and downstream needs should be selected to ensure protein yield while minimizing the formation of insoluble protein.
[0004] With the increase of network data and the development of computer science, machine learning (ML), as a branch of artificial intelligence, has been widely used in e-commerce, driving, biochemistry, biomedicine, and pharmaceutical industries. In the field of protein drug and vaccine research, because ML methods take up less resources and have outstanding performance in classification and regression, they can be involved in every step of drug and vaccine production. Therefore, ML methods have shown great potential in the production and discovery of protein drugs. In theory, during protein expression, matching the various properties of the target protein with the properties of the protein in the expression system can better promote the correct folding of the protein, ensure its biological activity and solubility, and ensure protein production and its homeostasis in the microenvironment of the expressing cell. Although machine learning methods have been widely used in the field of protein expression to analyze protein sequence solubility, aggregation characteristics and stability, there are no related research reports on the application of machine learning algorithms to the selection of protein expression systems.
[0005] As the first step in the protein expression process, it is extremely important to select a suitable expression system for the target protein characteristics. Protein activity is an important factor affecting subsequent protein-related research, and is closely related to protein solubility. Selecting the right protein expression system can not only improve the solubility and yield of the active protein, but also contribute to the optimization of the subsequent expression process. Therefore, there is an urgent need for a novel, high-precision expression system selection tool. The present invention constructs a target protein-expression system adaptation optimization model based on the XGBoost algorithm in the field of machine learning for the first time. This invention will provide researchers with a simple and easy-to-use protein expression system selection recommendation tool, and help improve future protein expression efficiency and reduce protein drug and vaccine research and development costs. Summary of the Invention
[0006] To address the lack of system selection strategies in existing protein expression systems, the present invention uses the XGBoost algorithm to learn protein-related data from Escherichia coli (E. coli), Pichia pastoris (Pichia P), Spodoptera frugiperda (sf9), and Chinese hamster ovary (CHO) cells. Through feature engineering, model tuning, and generalization ability testing on test data, an XGBoost-based protein expression system optimization model is constructed that can evaluate the compatibility of target proteins and expression systems.
[0007] In order to achieve the above object, the technical solution adopted by the present invention is:
[0008] (1) Dataset collection and processing: First, all protein sequence data contained in E. coli (5241 sequences), Pichia P (8851 sequences), sf9 (26958 sequences), and CHO (23884 sequences) were collected. Bioinformatics and programming languages were used to align the data format and remove redundant content to obtain pure protein sequence information for analysis. Feature descriptors such as protein sequence composition, autocorrelation description information, and sequence order coupling number were deeply evaluated and data was collected. The data were deduplicated and classified using R language, and the data obtained from the two software were merged. The data were divided into training set, validation set, and test set, and each data set was normalized and subjected to principal component analysis (PCA).
[0009] (2) Feature Importance Analysis: In the current study, the importance of features can be evaluated using the Boruta algorithm.
[0010] (3) XGBoost model construction and optimization: The XGBoost algorithm was used to initially learn the hidden patterns in the data, with 10 10-fold CV runs. Hyperparameters were then optimized using the competition method, with a "smote" sampling method and 10 10-fold CV runs. The model was constructed using the optimal hyperparameters, and performance was evaluated on the validation set. ROC / PRC-AUC plots were plotted to assess model stability.
[0011]
[0012] (4) Evaluation of the model’s generalization ability on the test set
[0013] The optimized model is used to further verify its accuracy on the TestS set.
[0014] Beneficial effects of the present invention:
[0015] To address the lack of protein expression system selection options and the resulting waste of resources, this paper combines bioinformatics, artificial intelligence, and bioengineering techniques. Using various protein sequence data from E. coli, Pichia P, sf9, and CHO cells themselves, it constructs a model based on the XGBoost algorithm to recommend expression systems based on target protein-expression system compatibility. The construction and application of this predictive model can provide guidance for future protein expression, identify protein optimization strategies, improve protein expression efficiency, and reduce the cost of developing protein drugs and vaccines. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 Schematic diagram of the process of the present invention.
[0017] Figure 2Data processing, feature importance analysis, and training data organization.
[0018] Figure 3 The top 50 significantly important features.
[0019] Figure 4 ROC plot and AUC value of the model.
[0020] Figure 5 PRC plot and AUC value of the model. DETAILED DESCRIPTION
[0021] The specific embodiments of the present invention are further described in detail below with reference to the examples. Figure 1 Schematic diagram of the overall method of the present invention.
[0022] Example 1. Acquisition and processing of expression system characteristic data
[0023] (1) Dataset collection and processing
[0024] like Figure 2 As shown, by searching the UniProt database (https: / / www.uniprot.org / ), we collected sequence information for 5241 E. coli proteins, 8851 Pichia P proteins, 26958 Sf9 cell proteins, and 23884 CHO cell proteins. Using Propy and Peptide, we identified 19,949 protein-sequence-related feature descriptors, which served as the original protein feature database for model building. After removing features with zero values greater than 80% and / or correlation coefficients greater than 75%, we generated a 64,934 × 491 database. This database was randomly divided into the original training set, validation set, and test set in a ratio of 7:1.5:1.5, and data normalization was performed.
[0025] (2) Feature Importance Analysis
[0026] Figure 2 The Boruta algorithm was used to analyze the feature importance by calculating the Z-score, and it was finally determined that 491 features were important features, of which 138 features had significant contribution and importance ( Figure 3 The top 50 significant features were ranked and the model was built using all the important features.
[0027] Example 2: Model preliminary verification, optimization and performance testing
[0028] (1) Preliminary training of XGBoost algorithm
[0029] The XGBoost algorithm was used to initially build models for the constructed training set, and the ROC_AUC values of these models were verified to understand the learning effect of the XGBoost algorithm on the training set data. The algorithm achieved an ROC_AUC value of 0.905±0.0005 on the training data, further optimizing the model.
[0030] (2) XGBoost algorithm hyperparameter optimization
[0031] The XGBoost hyperparameters were optimized using a competitive method, showing that the optimal hyperparameter combination for the model was min_n = 7, tree_depth = 9, learn_rate = 0.0424, loss_reduction = 0.0000344, sample_size = 0.801, and stop_iter = 4. The accuracy reached 0.784 ± 0.000778, and the AUC value reached 0.943 ± 0.00036, indicating satisfactory prediction performance. The model was named BTXO and its performance on an independent validation set was calculated (Table 1). The model not only had satisfactory overall performance (all performance indicators were greater than 0.65), but also had good performance in identifying various protein categories.
[0032] Table 1. Overall performance of the model and performance for each category on the validation set.
[0033]
[0034] In addition, the ROC curve and AUC value ( Figure 4 ) also confirmed that the BTXO model has good prediction robustness, and the AUC values for E. coli, Pichia.P, Sf9, and CHO were 0.9706, 0.9684, 0.9123, and 0.9199, respectively; Since the data used in this patent has obvious imbalance, the PRC curve and its AUC value were also drawn and evaluated, and the results showed that ( Figure 5 ) The AUC values of the PRC curves for each category are all higher than 0.80, and the model can handle imbalanced data well.
[0035] (2) Model performance on the test set
[0036] The model's generalization ability was further verified on the test set. As shown in Table 2, the model achieved a high overall accuracy of 0.7907 on the unfamiliar dataset and was able to effectively identify proteins of various categories. Although the prediction performance on E. coli was slightly worse, it did not affect the overall prediction performance. These results demonstrate that the model has good generalization ability.
[0037] Table 2. Model performance on the test set.
[0038]
Claims
1. A protein expression system adaptability analysis tool based on the XGBoost algorithm, characterized in that: The following steps are involved: S1: Collect information on all protein sequences of Escherichia coli (E. coli), Pichia pastoris (Pichia P), Spodoptera frugiperda (sf9) cells, and Chinese hamster ovary (CHO) cells, and use bioinformatics tools to obtain these protein sequences and structural feature descriptors to generate a raw protein feature database. After data preprocessing, an expression system protein characterization database is obtained, which is divided into training set, test set, and validation set. S2: Use the Boruta algorithm to analyze the importance of feature data for all features and calculate the average Gini coefficient. Extract important features for the next step of modeling. S3: Use the XGBoost algorithm to learn important feature data in the training set, build models and optimize hyperparameters to evaluate the degree to which the algorithm learns hidden data patterns in different expression systems. At the same time, verify the performance of the constructed model on the validation data and evaluate different model performance indicators. S4: Use the model to evaluate the test set to measure the generalization ability of predictions on unfamiliar data.
2. The prediction method according to claim 1, characterized in that The specific method of step (1) is as follows: all protein sequences for E. coli, Pichia P, sf9, and CHO are collected. Different feature descriptors of the collected protein sequences are evaluated using peptides and Propy3, and an original expression system feature database is constructed. Incomplete samples, features with a zero score of more than 80%, and features with a correlation coefficient greater than 75% are deleted, and finally an expression system feature training database is constructed. The database is divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5, and the data is standardized.
3. The protein expression system adaptability analysis tool according to claim 1, wherein The step S2 uses the Boruta algorithm to calculate feature importance analysis for the feature data in the training set described in claim 2. Where X is the importance score of the real feature in XGBoost (mean average precision reduction); μ is the average importance score of all shadow features; σ is the standard deviation of the importance scores of all shadow features. Finally, the training set, validation set, and test set are extracted according to the important features obtained to construct the training set, validation set, and test set for model training.
4. The protein expression system adaptability analysis tool according to claim 1, wherein In step S3, an XGBoost model is constructed, algorithm learning modeling is performed on the constructed training set, and the constructed validation set is used to validate the model. The construction of the XGBoost model includes the following steps: a. Set up 10 10-fold cross validation (CV) on the training set containing important feature data, preliminarily build the XGBoost model, and check the performance. b. Optimize the XGBoost hyperparameters using the competition method and screen them using a 100-hyperparameter grid. When training the model, perform 10 10-fold cross-validation (CV) cycles to obtain the hyperparameter combinations with good model performance on the validation set. Plot the ROC and PRC curves and calculate the AUC value under the curve.
5. The protein expression system adaptability analysis tool according to claim 1, wherein: The step S4 uses a test set to verify the model constructed in claim 4 and evaluate its generalization effect on an unfamiliar dataset.
6. A protein expression system adaptability analysis model based on the XGBoost algorithm, characterized in that: include: Database composition: The original protein feature database is generated, and the important feature database (including training set, validation set and test set) is obtained through data preprocessing and feature importance analysis. Model composition: Use a training set consisting of important features to perform algorithm learning, parameter optimization, and model construction, and use a validation set to test the constructed models. Prediction result composition: Use test data containing important features to input into the optimized XGBoost model to evaluate the compatibility of the target protein and expression system and recommend an expression system.
7. The prediction method according to any one of claims 1 to 6 is preferably applied to a protein expression system. The prediction method according to any one of claims 1 to 6 is applied to protein expression system screening and expression scheme design.