A batch prediction model training method for hydrolytic protease enzyme cleavage sites
By using a self-built database, a T5 pre-trained model, and multi-model ensemble optimization, the accuracy and efficiency issues of predicting cleavage sites for hydrolytic proteases were resolved, achieving efficient and accurate cleavage site prediction while reducing experimental costs and time.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG GUANZHAN NUTRITION & HEALTH TECHNOLOGY CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for predicting cleavage sites of hydrolytic proteases have shortcomings in terms of accuracy, efficiency, and generalization ability. Traditional experimental methods are time-consuming and costly, sequence alignment-based methods are highly dependent, and machine learning methods rely on large amounts of high-quality data and are difficult to select and optimize, resulting in unstable prediction results.
By building a self-built database of hydrolytic protease cleavage sites, using the T5 pre-trained model to extract amino acid sequence features and reduce dimensionality, and combining the ensemble optimization and cross-validation of multiple machine learning models, the best-performing model was selected for prediction.
It improves the accuracy and generalization ability of predicting cleavage sites of hydrolytic proteases, reduces experimental costs and time, and enhances the stability and reliability of prediction results.
Smart Images
Figure CN122117017A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrolytic proteases, and particularly to a method for training a batch prediction model of hydrolytic protease cleavage sites. Background Technology
[0002] Hydrolase proteases are enzymes that catalyze protein hydrolysis reactions and are widely used in biopharmaceuticals, food processing, and environmental remediation. Hydrolases act on substrate proteins through specific cleavage sites, breaking them down into small peptides or amino acids, thereby regulating protein function. This enzymatic cleavage process is crucial for optimizing biocatalytic processes and developing enzyme applications; therefore, accurately predicting the cleavage sites of hydrolases has become a key issue in current protein function research. However, existing methods for predicting cleavage sites have many shortcomings in terms of accuracy, efficiency, and generalization ability, hindering their widespread application in practice.
[0003] Traditional methods for predicting cleavage sites of hydrolytic proteases mainly rely on laboratory biological experimental techniques, such as enzyme digestion experiments and mass spectrometry analysis. These methods can provide relatively accurate cleavage site data, but they generally suffer from problems such as long processing times and high experimental costs. Enzyme digestion experiments not only require a large amount of reagents, equipment, and human resources, but the results of each experiment can also be affected by various factors such as operating conditions and equipment precision, leading to instability and non-reproducibility of the results.
[0004] Taking mass spectrometry as an example, although mass spectrometry can effectively identify enzyme cleavage sites in proteins, its ability to analyze complex protein samples is limited, especially when dealing with large-scale proteomics data. The resolution and throughput of mass spectrometry still cannot meet the needs of efficient and comprehensive prediction. In addition, traditional biological experimental methods usually rely on a single experimental protocol, which makes it difficult to take into account the characteristics of different enzymes and complex experimental conditions when dealing with diverse enzyme cleavage sites, thus limiting its application in high-throughput screening.
[0005] To overcome the limitations of traditional experimental methods, researchers have attempted to predict the cleavage sites of hydrolytic proteases using sequence alignment-based methods. This method typically infers potential cleavage sites by comparing the amino acid sequence of the target protein with a database of known cleavage site sequences. Common sequence alignment tools include BLAST and HMMER, which search for potential cleavage sites by calculating the similarity between sequences.
[0006] However, sequence alignment-based methods also have significant limitations. First, these methods are heavily reliant on databases; if the information on restriction enzyme sites in the database is incomplete or lacks detailed descriptions of specific enzymes, the accuracy of predictions will be greatly reduced. Second, the prediction of restriction enzyme sites depends not only on the composition of the amino acid sequence but also on the protein's three-dimensional structure, spatial conformation, and the physicochemical properties of the substrate. Traditional sequence alignment methods often ignore these structural factors that influence the selection of restriction enzyme sites, relying solely on the similarity of amino acid sequences for prediction, thus easily leading to large errors. Furthermore, sequence alignment methods have low accuracy in predicting cross-species or cross-enzyme sites, especially when dealing with insufficiently studied restriction enzyme sites, where traditional methods often perform poorly.
[0007] In recent years, with the continuous advancement of computing technology, machine learning has gradually become a mainstream tool in the field of bioinformatics, especially in protein function prediction and enzyme cleavage site identification. Machine learning methods improve the accuracy and efficiency of predictions by automatically learning features and building models from large amounts of data. For the prediction of enzyme cleavage sites of hydrolytic proteases, machine learning methods can accurately identify cleavage sites by extracting more feature information (such as amino acid sequences, protein structure, and environmental factors) without relying on traditional sequence alignment.
[0008] However, existing machine learning methods still face several technical bottlenecks. First, the quality and quantity of data remain key factors limiting model performance. Although several protease hydrolysis databases exist, their size and coverage are still limited, especially in research on certain less common enzymes or specific proteins, where the lack of relevant data poses a significant challenge to model training. Furthermore, real-world datasets often exhibit an imbalance between positive and negative samples, with positive samples (i.e., restriction enzyme sites) far outnumbering negative samples (non-restriction sites). This imbalance can cause machine learning models to be biased towards the negative class during training, thus affecting the overall performance of the model.
[0009] Secondly, feature extraction plays a crucial role in machine learning models. Existing feature extraction methods largely rely on manual design, typically based on the physicochemical properties of amino acids. While simple and effective, this approach struggles to comprehensively capture all potential factors influencing enzyme cleavage sites within protein molecules. With the rise of deep learning, feature learning methods based on deep neural networks can automatically extract more complex features from data. However, these methods demand significant computational resources and large datasets, and with limited data, they are prone to overfitting, failing to effectively generalize to new datasets.
[0010] Even though machine learning methods have shown good performance in predicting cleavage sites of hydrolytic proteases, model selection and optimization remain challenging. Currently, many different algorithms and models (such as logistic regression, decision trees, support vector machines, random forests, and deep neural networks) are available. Choosing the most suitable model from among these options and fine-tuning its parameters still requires extensive experimentation and experience. Furthermore, the prediction results of different models often vary significantly, especially when dealing with complex and diverse cleavage sites; the accuracy and generalization ability of existing models still need further improvement.
[0011] To improve the generalization ability of models and avoid overfitting and underfitting, methods such as cross-validation are widely used in model evaluation. However, while these methods can improve the stability and robustness of models, they still struggle to address the problem of insufficient generalization ability when faced with insufficient data or high sample diversity. Constructing a general model that performs well under various conditions remains a significant challenge in current research.
[0012] In summary, although existing methods for predicting cleavage sites of hydrolytic proteases have improved the efficiency and accuracy of prediction to some extent, they still have many shortcomings. Traditional biological experimental methods are not only costly and inefficient, but also suffer from poor stability and reproducibility of results; sequence alignment-based prediction methods are limited by database constraints and struggle to adapt to the complexity of different enzymes; while machine learning methods can improve prediction accuracy, their reliance on large amounts of high-quality data and the optimization of feature extraction and model selection remain significant challenges. Therefore, overcoming these technical bottlenecks and developing an efficient, accurate, and generalizable model for predicting cleavage sites of hydrolytic proteases has become an urgent technical challenge. Summary of the Invention
[0013] In view of this, the present invention aims to propose a batch prediction model training method for hydrolytic protease cleavage sites to solve the problems in the prior art.
[0014] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention proposes a batch prediction model training method for hydrolytic protease cleavage sites, the method comprising the following steps: S1. Obtain hydrolytic protease data, cleavage site data, and substrate data from a self-built hydrolytic protease cleavage site database; S2, use a pre-trained model to convert the amino acid sequence data obtained in S1 into a high-dimensional feature vector, and then perform dimensionality reduction processing on the high-dimensional feature vector to obtain a low-dimensional feature vector; S3 uses multiple machine learning models to train low-dimensional feature vectors multiple times, and selects the best-performing model based on its accuracy on the validation set. S4. Cross-validation is used to evaluate the selected models to confirm their generalization ability. S5 will use the evaluated model to predict the cleavage sites of hydrolytic proteases on the target protein fragment.
[0015] Furthermore, in S1, the substrate data includes information on the restriction sites of different hydrolytic proteases on the substrate. Amino acid fragments that respond to the restriction sites are extracted as positive data, and amino acid fragments that respond to the non-restriction sites are extracted as negative data. The positive and negative samples are balanced, and the amino acid sequences of the positive data and the Fuxiang data are applied to S2.
[0016] Furthermore, the positive data includes an 8-amino acid sequence at the P4-P4 site.
[0017] Furthermore, in S2, the pre-trained model is a T5 pre-trained model based on the Transformer architecture, and the dimensionality reduction process is achieved through principal component analysis (PCA).
[0018] Furthermore, the T5 pre-trained model converts each amino acid in the amino acid fragment into a high-dimensional vector with an embedding vector dimension of 1024, generating an 8×1024 matrix. Principal component analysis (PCA) reduces the dimensionality of the 8×1024 matrix to an 8×8 matrix.
[0019] Furthermore, the flatten operation is used to flatten the dimension-reduced matrix into a vector, forming a feature vector that can be used to train a machine learning model.
[0020] Furthermore, S3 includes multiple machine learning models, including logistic regression, decision trees, random forests, support vector machines, Gaussian Naive Bayes, gradient boosting trees, and the K-nearest neighbors algorithm.
[0021] Furthermore, in S4, cross-validation comprehensively evaluates the model's performance on different data subsets by dividing the dataset into multiple subsets, with each subset serving as the validation set in turn and the remaining subsets serving as the training set.
[0022] Compared with the prior art, the present invention has the following advantages: In this invention, the accuracy and generalization ability of predicting the cleavage sites of hydrolytic proteases are effectively improved by using techniques such as building a self-built database of hydrolytic protease cleavage sites, extracting amino acid sequence features based on the T5 pre-trained model, multi-model integration optimization, and cross-validation. Attached Figure Description
[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the present invention; Figure 2 Selecting data graphs for different models in this invention; Figure 3 This is the cross-validation diagram in this invention; Figure 4 This is a graph showing the final model accuracy data in this invention; Figure 5 This is a map showing the predicted enzyme cleavage sites in this invention. Detailed Implementation
[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0025] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "back," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0026] Furthermore, in the description of this invention, unless otherwise explicitly defined, the terms "installation," "connection," "linking," and "connector" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention in light of the specific circumstances.
[0027] The following will refer to the appendix. Figures 1 to 5 The present invention will be described in detail with reference to the embodiments.
[0028] Overall, this invention proposes a batch prediction model training method for protease cleavage sites, comprising the following steps: S1. Obtain hydrolytic protease data, cleavage site data, and substrate data from a self-built hydrolytic protease cleavage site database; S2 uses a pre-trained model to convert the amino acid sequence data obtained in S1 into a high-dimensional feature vector, and then performs dimensionality reduction on the high-dimensional feature vector to obtain a low-dimensional feature vector. S3 uses multiple machine learning models to train low-dimensional feature vectors multiple times, and selects the best-performing model based on its accuracy on the validation set. S4. Cross-validation is used to evaluate the selected models to confirm their generalization ability. S5 will use the evaluated model to predict the cleavage sites of hydrolytic proteases on the target protein fragment.
[0029] It should be noted that in S1, the substrate data includes information on the restriction sites of different hydrolytic proteases on the substrate. Amino acid fragments that respond to the restriction sites are extracted as positive data, and amino acid fragments that respond to the non-restriction sites are extracted as negative data. The positive and negative samples are then balanced. The amino acid sequences of the positive data and the Fuxiang data are applied to S2. The positive data includes 8 amino acid sequences, which are P4-P4 sites.
[0030] In this embodiment, during the data preprocessing stage, hydrolytic protease data, restriction enzyme site data, and substrate data are first extracted from the database. Specifically, the substrate data is obtained through extensive experimental research and includes information on the restriction enzyme sites of different hydrolytic proteases on the substrate. Positive and negative data are extracted from the substrate to ensure the representativeness and balance of the dataset.
[0031] In detail, during preprocessing, the amino acid sequences of the hydrolytic protease and the substrate are first aligned, and eight representative amino acid sequences (i.e., P4-P4 sites) are extracted. These amino acid sequences represent the substrate's response sites (i.e., cleavage sites). These amino acid sequences are then used as input features (X), with a label of 1 (indicating a valid cleavage site). This process generates the positive dataset. Conversely, negative data (i.e., amino acid sequences that are not cleavage sites) are randomly selected from the substrate sequences and used as input features (X), with a label of 0. To ensure dataset balance, the number of positive and negative data points is roughly equal, thus avoiding data bias during model training.
[0032] In this way, the preprocessed dataset includes samples labeled 1 and 0, which can be used to train and validate different prediction models. The dataset undergoes multiple cleaning and denoising processes to ensure its quality and accuracy, providing a solid data foundation for subsequent feature extraction and model training.
[0033] In S2, the pre-trained model is the T5 pre-trained model based on the Transformer architecture. Dimensionality reduction is achieved through Principal Component Analysis (PCA). The T5 pre-trained model converts each amino acid in the amino acid fragment into a high-dimensional embedding vector of dimension 1024, generating an 8×1024 matrix. PCA then reduces the 8×1024 matrix to an 8×8 matrix. Finally, the flatten operation is used to flatten the dimensionality-reduced matrix into a single vector, forming a feature vector that can be used to train the machine learning model.
[0034] In predicting the cleavage sites of proteases, the amino acid sequence of a protein is the most fundamental input feature. However, the raw amino acid sequence information is too complex and contains a lot of redundant information. Directly using this data for training may lead to inefficient model training and even affect the prediction results.
[0035] To address this issue, this embodiment employs a pre-trained T5 large model (prot_t5_xl_uniref50) for feature embedding. T5 is a deep learning model based on the Transformer architecture, which has achieved excellent performance in biological sequence analysis. By utilizing the embedder module of the T5 model, each amino acid sequence is converted into a high-dimensional vector, with each amino acid corresponding to an embedding vector of dimension 1024. For each 8-amino acid sequence, this embedding process generates an 8×1024 matrix, describing the complex features of the amino acid sequence.
[0036] Based on the above settings, since high-dimensional data increases computational overhead, Principal Component Analysis (PCA) is further used to reduce the dimensionality of the matrix. Through PCA dimensionality reduction, the 8×1024 matrix is compressed into an 8×8 matrix, effectively preserving the main features of the data while significantly reducing computational resource consumption. Finally, the flatten operation is used to flatten the dimensionality-reduced matrix into a vector, forming a feature vector that can be used to train a machine learning model.
[0037] This process transforms each amino acid sequence into a 64-bit vector, which fully preserves the high-dimensional information of the amino acid sequence and provides effective input features for subsequent model training.
[0038] In S3 above, multiple machine learning models include logistic regression, decision tree, random forest, support vector machine, Gaussian Naive Bayes, gradient boosting tree, and K-nearest neighbor algorithm.
[0039] These models have different advantages and disadvantages and are suitable for datasets with different feature distributions. Therefore, by comparing multiple models, we can find the most suitable one.
[0040] During training, the dataset was divided in an 8:2 ratio, with 80% used for model training and 20% for model validation. For each model, we performed multiple training runs and calculated its accuracy on the validation set to determine its predictive performance. In this way, we selected the best-performing model.
[0041] To further ensure the selected model has good generalization ability, a 10x cross-validation method was used for evaluation. Cross-validation divides the dataset into multiple subsets, with each subset serving as the validation set in turn, and the remaining subsets serving as the training set, thus comprehensively evaluating the model's performance on different data subsets. Through cross-validation, we can reduce the risk of overfitting and ensure the model's predictive accuracy on unknown data.
[0042] For example, among the 58 trained models, 52 achieved an accuracy exceeding 80% after evaluation. These models performed well in the task of predicting cleavage sites of hydrolyzed proteases and can be directly applied to practical predictions. The remaining 6 models, however, did not achieve an accuracy of 80% and require further parameter tuning or data processing to improve their performance.
[0043] It's important to further explain that once the optimal model is determined and its generalization ability is ensured through cross-validation, the next step is to apply these trained models to actual protein fragment prediction tasks. These models are applied to protein datasets to predict the cleavage sites of hydrolytic proteases. Through prediction, the predicted cleavage sites of different hydrolytic proteases on the same protein fragment can be obtained.
[0044] At this stage, the model can output possible restriction enzyme sites based on the input protein fragment. These predictions provide guidance for experimental researchers, helping them select appropriate hydrolytic proteases, thereby reducing workload and cost and improving experimental efficiency.
[0045] The reliability and accuracy of the prediction results depend on the training quality of the model and the richness of the dataset. By continuously optimizing the model and expanding the dataset, this invention can provide continuous support for the prediction of cleavage sites of hydrolytic proteases.
[0046] In summary, this invention effectively improves the accuracy and generalization ability of predicting hydrolytic protease cleavage sites through techniques such as building a self-built hydrolytic protease cleavage site database, extracting amino acid sequence features based on the T5 pre-trained model, multi-model integration optimization, and cross-validation.
[0047] It should be added that other pre-trained models (such as BERT, GPT, etc.) can be used for amino acid sequence embedding; in the model selection process, in addition to common machine learning models, deep learning models (such as CNN, RNN) can also be considered for feature learning; feature dimensionality reduction can use other methods, such as t-SNE or LLE, to further optimize feature representation and computational efficiency; to optimize model parameters, methods such as grid search or Bayesian optimization can be used to further improve the model's prediction accuracy. The above design changes can be flexibly adjusted and applied according to actual needs.
[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for training a batch prediction model of hydrolytic protease cleavage sites, characterized in that, The method includes the following steps: S1. Obtain hydrolytic protease data, cleavage site data, and substrate data from a self-built hydrolytic protease cleavage site database; S2, use a pre-trained model to convert the amino acid sequence data obtained in S1 into a high-dimensional feature vector, and then perform dimensionality reduction processing on the high-dimensional feature vector to obtain a low-dimensional feature vector; S3 uses multiple machine learning models to train low-dimensional feature vectors multiple times, and selects the best-performing model based on its accuracy on the validation set. S4. Cross-validation is used to evaluate the selected models to confirm their generalization ability. S5 will use the evaluated model to predict the cleavage sites of hydrolytic proteases on the target protein fragment.
2. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 1, characterized in that: In S1, the substrate data includes information on the restriction sites of different hydrolytic proteases on the substrate. Amino acid fragments that respond to the restriction sites are extracted as positive data, and amino acid fragments that respond to the non-restriction sites are extracted as negative data. The positive and negative samples are balanced, and the amino acid sequences of the positive data and the Fuxiang data are applied to S2.
3. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 2, characterized in that: The positive data includes an 8-amino acid sequence, which is the P4-P4 site.
4. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 1, characterized in that: In S2, the pre-trained model is a T5 pre-trained model based on the Transformer architecture, and the dimensionality reduction process is achieved through principal component analysis (PCA).
5. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 4, characterized in that: The T5 pre-trained model converts each amino acid in an amino acid fragment into a high-dimensional vector with an embedding vector dimension of 1024, generating an 8×1024 matrix. Principal component analysis (PCA) reduces the dimensionality of the 8×1024 matrix to an 8×8 matrix.
6. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 5, characterized in that: The flatten operation is used to flatten the dimension-reduced matrix into a vector, forming a feature vector that can be used to train a machine learning model.
7. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 1, characterized in that: In S3, multiple machine learning models include logistic regression, decision trees, random forests, support vector machines, Gaussian Naive Bayes, gradient boosting trees, and the K-nearest neighbors algorithm.
8. The method for training a batch prediction model of hydrolytic protease cleavage sites according to claim 1, characterized in that: In S4, cross-validation divides the dataset into multiple subsets, with each subset serving as the validation set in turn and the remaining subsets serving as the training set, thereby comprehensively evaluating the model's performance on different data subsets.