Polypeptide biological activity regression prediction method and device

Abstract

The invention discloses a regression prediction method and device for the biological activity of polypeptide, and the method comprises the steps: collecting the biological activity numerical information of polypeptide, and constructing a polypeptide biological activity data set; the method comprises the following steps: performing feature extraction on polypeptides in a polypeptide biological activity data set by using an Evolutionary Scale Modeling-2 protein language model, and dividing the polypeptides into a training set and a test set; constructing a long short-term memory network model, and optimizing hyper-parameters of the long short-term memory network model through a bionic meta-heuristic algorithm; training the hyper-parameter optimized long-short-term memory network model by adopting a training set to obtain a polypeptide biological activity regression prediction model; evaluating the biological activity regression prediction model by adopting the test set; and substituting to-be-detected polypeptide sequences into the biological activity regression prediction model in batches to obtain the biological activity value of the polypeptide. According to the polypeptide biological activity regression prediction method and device, the specific biological activity numerical value information of the polypeptide can be rapidly and accurately predicted, and meanwhile the high fitting degree and the low root-mean-square error are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioactive peptide screening, and in particular to a method and device for regression prediction of the bioactivity of a polypeptide. Background Art

[0002] Bioactive peptides have an important impact in inhibiting chronic and degenerative diseases such as cardiovascular disease and diabetes. Proteins extracted from animals and plants are valuable sources of bioactive peptides, which have specific physiological and biochemical activities in addition to providing essential nutritional benefits.

[0003] In recent years, scientists have shown increasing interest in extracting bioactive peptides directly from food. For example, some scientists have extracted a water-soluble enzyme extract, EERB, from rice bran, which is effective in inhibiting cardiovascular disease and oxidative damage. In addition, peptides extracted from egg white, such as FRADHPFL and RADHPFL, play an important role in vasodilation and anti-hypertension. These bioactive peptides can play beneficial activities in functional foods, and ingredients derived from food have higher food safety. Therefore, consumer preference for active peptides is shifting towards food sources.

[0004] However, the current screening and identification of biological peptides mostly rely on traditional chemical methods, such as separation from protein hydrolysates or fermentation products, and direct acquisition through chemical synthesis. At the same time, the determination of the biological activity of peptides requires a large number of wet experiments. Due to the high cost of wet experiments, scientists have begun to turn to artificial intelligence (AI) technologies such as deep learning to guide the screening and activity determination of biological peptides.

[0005] Metaheuristic optimization algorithms are advanced optimization algorithms used to solve complex combinatorial optimization problems. Compared with traditional optimization algorithms, metaheuristic algorithms are more abstract and general. They do not rely on specific knowledge of the problem domain, but provide a general framework for searching the solution space of the problem to find a near-optimal solution. Metaheuristic algorithms have the advantages of high efficiency, strong versatility, and the ability to solve complex problems.

[0006] Evolutionary Scale Modeling-2 (ESM-2) is a protein language model based on the BERT architecture, developed by the Meta Foundation Artificial Intelligence Research Protein Team (FAIR) in 2022 based on a very large protein dataset of millions of sequences. It can internalize the encrypted information in the protein sequence and provide a more comprehensive approach to peptide representation research.

[0007] Deep learning is a branch of artificial intelligence that uses statistical and computer science methods to enable computer systems to extract patterns, laws, and knowledge from large amounts of data and use this knowledge to perform tasks such as regression prediction, enabling computer systems to learn from data and make decisions or predictions based on the learned knowledge.

[0008] In addition, existing patents rarely use the protein language model ESM-2 and metaheuristic algorithms to optimize bioactive peptide related models, and the bioactive peptide regression models in existing literature and patents perform poorly. The present invention can make up for these shortcomings, quickly and accurately obtain the specific biological activity prediction values ​​of bioactive peptides, and ensure a high degree of fitting and a low root mean square error, providing a reference for the rapid detection of bioactive peptides and future research. Summary of the invention

[0009] The present invention provides a method and device for predicting the biological activity of a polypeptide by regression, which can quickly and accurately predict the specific biological activity numerical information of the polypeptide, while ensuring that the model has a high degree of fit and a low root mean square error.

[0010] The technical solution of the present invention is as follows:

[0011] A method for regression prediction of the biological activity of a polypeptide comprises the following steps:

[0012] (1) Collecting the biological activity numerical information of peptides and constructing a peptide biological activity dataset; using the Evolutionary Scale Modeling-2 (ESM-2) protein language model to extract features of peptides in the peptide biological activity dataset and divide it into a training set and a test set;

[0013] (2) Construct a long short-term memory network model and optimize the hyperparameters of the long short-term memory network model through a bionic meta-heuristic algorithm;

[0014] (3) Using the training set to train the long short-term memory network model after optimizing the hyperparameters, a regression prediction model for the biological activity of the peptide is obtained;

[0015] (4) Substitute the peptide sequences to be tested into the biological activity regression prediction model in batches to obtain the biological activity value of the peptide.

[0016] In step (1), the biological activity numerical information of the polypeptide is the antioxidant property or bitterness threshold of the polypeptide.

[0017] The antioxidant activity was expressed using the TEAC value, with the unit being μmol TE / μmol peptide; the bitterness threshold was expressed in mmol / L.

[0018] In step (1), the Evolutionary Scale Modeling-2 (ESM-2) protein language model is used to extract features from the peptides in the peptide biological activity dataset, including:

[0019] The protein sequence data of the peptides in the Excel file were read using the Pandas library. The format of the protein sequence data of each peptide was converted into a format acceptable to the Evolutionary Scale Modeling-2 protein language model. The Evolutionary Scale Modeling-2 protein language model was then loaded to calculate the feature representation of the protein sequence data of each peptide, generating 320 feature vectors for each peptide.

[0020] In step (2), the long short-term memory network model includes three LSTM layers and three dense layers connected in sequence;

[0021] The number of neurons in the three LSTM layers are 128, 64, and 32 respectively, and the LSTM layer returns the hidden state values ​​of all time steps;

[0022] Among the three dense layers, the number of neurons in the last dense layer is 1, and the number of neurons in the first two dense layers is optimized by the bionic meta-heuristic algorithm.

[0023] The bionic meta-heuristic algorithm is a whale optimization algorithm (WOA), a sparrow search algorithm (SSA) or a Harris Hawks Optimization algorithm (HHO).

[0024] Preferably, in step (3), the evaluation indicators for evaluating the bioactivity regression prediction model are absolute error (AE), root mean square error (RMSE), goodness of fit (R 2 ) and explained variance (EVS).

[0025] The present invention adopts the combination of R 2 , root mean square error, absolute error, explained variance and other indicators are analyzed, and both the training set and the test set are evaluated to prevent overfitting.

[0026] The present invention also provides a device for regression prediction of the biological activity of a polypeptide, comprising an activity prediction module, wherein the activity prediction module comprises one or more biological activity regression prediction models constructed by the above method;

[0027] Substitute the peptide sequences to be tested into the selected biological activity regression prediction model in batches to obtain the corresponding biological activity data values ​​of the peptides.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] (1) The present invention can quickly, efficiently and accurately predict the specific biological activity values ​​of food-borne bioactive peptides, reducing the research and development cycle and cost, and providing assistance for the future development of food-borne bioactive peptides;

[0030] (2) The present invention combines the meta-heuristic algorithm with the deep learning model to improve the traditional hyperparameter optimization method, effectively avoiding the waste of computing resources and low efficiency caused by the traditional gradient descent method to optimize hyperparameters, and improving the efficiency while ensuring the accuracy of the development model;

[0031] (3) The present invention uses the latest protein language model Evolutionary Scale Modeling-2 for peptide embedding representation to generate peptide descriptors. ESM-2 based on the BERT architecture can internalize protein sequence information and accurately describe peptides from word vector features, position encoding features, residue-residue interaction features and window features. Compared with traditional peptide embedding methods such as sequence embedding and physical and chemical property embedding, it is more comprehensive;

[0032] (4) The present invention also constructs a regression prediction device for the biological activity of polypeptides. Users do not need to have any knowledge of ESM-2 and deep learning. They only need to input the polypeptide sequence to predict its specific biological activity value. It also supports large-scale prediction, which greatly facilitates future research in the field of food-borne bioactive peptides. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of the regression prediction method of bioactive peptides according to an embodiment of the present invention;

[0034] Figure 2 The training process and performance comparison of various bionic meta-heuristic algorithms in the embodiments of the present invention are shown in Figure (a), where WOA is a whale optimization algorithm, SSA is a sparrow search algorithm, and HHO is a Harris Hawk algorithm. Figure (b) shows the training process.

[0035] Figure 3 The prediction performance of the model after the optimization of the sparrow search algorithm in the embodiment of the present invention on the antioxidant tripeptide data set;

[0036] Figure 4 The training process and prediction performance of the optimal model structure deployed on the bitter peptide dataset according to an embodiment of the present invention, (a) shows the training process, and (b) shows the prediction performance. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention is clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Example:

[0039] like Figure 1 As shown in the figure, a regression prediction method for bioactive peptides based on bionic metaheuristic algorithm optimization and protein language model is proposed:

[0040] S1. Collect antioxidant tripeptide benchmark data sets from the public optimal antioxidant tripeptide regression model and collect bitter peptide data sets from the public bitter peptide classification model;

[0041] The benchmark data set of antioxidant tripeptides is entirely from the optimal antioxidant peptide regression model published in the literature, with a total of 130 tripeptides (reference: Du, Z.; Wang, D.; Li, Y., Comprehensive Evaluation and Comparison of Machine Learning Methods in QSAR Modeling of Antioxidant Tripeptides. Acs Omega 2022, 7 (29), 25760-25771.), and their antioxidant bioactivity indicators are determined by ABTS experiment and expressed using TEAC value, with the unit of μmol TE / μmol peptide; the data set of bitter peptides is entirely from the optimal bitter peptide classification prediction model published in the literature, with a total of 556 peptides (reference: Zhiyong, C., Zhiwei, Z., Tianxing, Z., Xueke, Z., Yin, Z., Hengli, M., Wenli, W., & Yuan, L. (2023). A Taste Peptides-Meta system including an umami / bitter classification model Umami_YYDS, a Taste PeptidesDB database and an open-source package Auto_Taste_ML. Food Chemistry, 405 (Part B), 134812.), their regression values ​​are bitterness thresholds in mmol / L. This is to make a fair comparison of model performance and prevent intentionally designed datasets from causing models to have good performance but poor generalization ability.

[0042] S2. Use the Evolutionary Scale Modeling-2 (ESM-2) protein language model to extract features from the dataset, generating 320 feature representations for each peptide. The feature-represented dataset is divided into a training set and a test set in a ratio of 8:2. The specific method is as follows:

[0043] S2-1. Obtain the Evolutionary Scale Modeling-2 protein language model developed by the FAIR team of Meta. This model is a pre-trained language model based on the BERT architecture that can internalize protein sequence information.

[0044] S2-2. Use the pre-trained version esm2_t6_8M_UR50D() in ESM-2, which has 6 representation layers, and select the last representation layer. Use the Pandas library to read the protein sequence data in the Excel file, convert each protein sequence format into a format acceptable to the ESM-2 protein language model, then load the Evolutionary ScaleModeling-2 protein language model, calculate the feature representation, and generate 320 feature vectors for each peptide; after feature representation, use the train_test_split function in the scikit-learn library to split the data set into a training set and a test set in a ratio of 8:2;

[0045] S3. Use three bionic meta-heuristic algorithms, Whale Optimization Algorithm (WOA), Sparrow Search Algorithm (SSA) and Harris Hawks Optimization (HHO), to optimize the hyperparameters of the deep learning LSTM model and determine the optimal hyperparameters of the model.

[0046] The specific method of using the bionic meta-heuristic algorithm optimization includes the following steps:

[0047] S3-1. Use the Keras library in Tensorflow to build a Long Short-Term Memory (LSTM) model. The built LSTM model has 6 layers. The first three layers are LSTM layers. The number of neurons in the layers is 128, 64, and 32 respectively. All LSTM layers return the hidden state values ​​of all time steps. The last three layers are dense layers. Except for the last layer with 1 neuron, the number of neurons in the other two layers is optimized by the bionic meta-heuristic algorithm. The model learning rate is 0.001.

[0048] S3-2, deploy the whale optimization algorithm, sparrow search algorithm, and Harris Eagle algorithm to S3-1 to build an LSTM model, train it with the antioxidant tripeptide dataset, optimize the three hyperparameters of the number of neurons in the two dense layers and the number of model training epochs, use the root mean square error as the standard to select the best optimization algorithm, and store the optimal hyperparameters.

[0049] refer to Figure 2 ,The four indicators of the sparrow search algorithm are all optimal on the test set, so the sparrow search algorithm is used as the optimal bionic meta-heuristic algorithm to optimize the model structure.

[0050] The number of neurons in the first dense layer obtained by the sparrow search algorithm is 50, the number of neurons in the second dense layer is 25, and the number of model training epochs is 96.

[0051] S4. After the optimal hyperparameters of the model are determined, the antioxidant dataset is used for training. The trained LSTM model is used to predict the antioxidant tripeptide dataset. The model is evaluated using multiple evaluation indicators and compared with the existing optimal model. The multiple evaluation indicators are absolute error (AE), root mean square error (RMSE), goodness of fit (R 2 ) and explained variance (EVS). These evaluation indicators are directly called and calculated through the metrics library in scikit-learn, and are calculated for the training set and the test set respectively.

[0052] Refer to Table 1, which shows the performance of the model structure optimized by the bionic metaheuristic algorithm of the present invention on the antioxidant tripeptide data set and its performance comparison with the current optimal model.

[0053] Table 1

[0054]

[0055] It can be seen from Table 1 that the performance of the model of the present invention is better than the current optimal model, and the fit index R 2 The value is 11.66% higher than the optimal model, and the root mean square error is 50.00% lower. 2 The closer it is to 1, the closer the root mean square error is to 0, indicating that the corresponding model regression prediction ability is stronger.

[0056] refer to Figure 3 , the predicted values ​​and actual values ​​on the test set are basically near the optimal prediction line, which also shows that the model performs well.

[0057] S5. Deploy the optimal model to the bitter peptide dataset after feature representation and evaluate its performance on the bitter peptide dataset. The specific method is as follows:

[0058] S5-1, read bitter peptides and their corresponding characteristic method representation tables, and convert them into corresponding tensor arrays;

[0059] S5-2, deploy the LSTM model with optimal hyperparameter settings in step S4 to the tensor array in step S5-1, perform model training, and use the evaluation index in step S4 to evaluate the model training set and test set respectively; its R 2 The R on the test set is 0.9204, and the root mean square error is 0.2649. 2 The value is 0.8385, and the root mean square error is 0.4414.

[0060] refer to Figure 4 , the predicted values ​​and actual values ​​on the test set are basically near the optimal prediction line, indicating that the model also performs well in predicting the biological activity of bitter peptides, has good generalization ability, and can predict a variety of biological activities.

[0061] S6, substituting the batch polypeptide sequences into the model, and the model outputting the specific activity value of the selected biological activity;

[0062] The model will output the specific antioxidant value or bitterness threshold of the peptide based on the characteristics of the input peptide, and has the ability to predict the active functions of a variety of bioactive peptides.

[0063] The batch peptide sequences were substituted into the model, and the peptides with a predicted antioxidant probability greater than 80% were considered as potential antioxidant peptides.

[0064] S7. Develop local applications based on the above optimal model.

[0065] A user-friendly local bioactive peptide prediction program was developed based on the Python Tkinter library, and the optimal LSTM model and ESM-2 peptide embedding method that performed well on the antioxidant tripeptide and bitter peptide datasets were deployed. The local application supports single sequence prediction and large-scale processing, allowing users to upload a single sequence or xlsx file without having to understand how to use ESM-2 embedding or machine learning knowledge to quickly and accurately predict the specific antioxidant tripeptide activity value or bitterness threshold of the peptide.

[0066] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for regression prediction of the biological activity of a polypeptide, characterized in that: The following steps are involved: (1) Collecting the biological activity numerical information of peptides and constructing a peptide biological activity dataset; using the EvolutionaryScale Modeling-2 protein language model to extract features of the peptides in the peptide biological activity dataset and dividing it into a training set and a test set; (2) Construct a long short-term memory network model and optimize the hyperparameters of the long short-term memory network model through a bionic meta-heuristic algorithm; (3) Using the training set to train the long short-term memory network model after optimizing the hyperparameters, a regression prediction model for the biological activity of the peptide is obtained; and using the test set to evaluate the regression prediction model for the biological activity; (4) Substitute the peptide sequences to be tested into the biological activity regression prediction model in batches to obtain the biological activity value of the peptide.

2. The method for regression prediction of the biological activity of a polypeptide according to claim 1, characterized in that: In step (1), the biological activity numerical information of the polypeptide is the antioxidant property or bitterness threshold of the polypeptide.

3. The method for regression prediction of the biological activity of a polypeptide according to claim 2, characterized in that: The antioxidant activity was expressed using the TEAC value, with the unit being μmol TE / μmol peptide; the bitterness threshold was expressed in mmol / L.

4. The method for regression prediction of the biological activity of a polypeptide according to claim 1, characterized in that: In step (1), the Evolutionary Scale Modeling-2 protein language model is used to extract features from the peptides in the peptide biological activity dataset, including: The protein sequence data of the peptides in the Excel file were read using the Pandas library. The format of the protein sequence data of each peptide was converted into a format acceptable to the Evolutionary Scale Modeling-2 protein language model. The Evolutionary Scale Modeling-2 protein language model was then loaded to calculate the feature representation of the protein sequence data of each peptide, generating 320 feature vectors for each peptide.

5. The method for regression prediction of the biological activity of a polypeptide according to claim 1, characterized in that: In step (2), the long short-term memory network model includes three LSTM layers and three dense layers connected in sequence; The number of neurons in the three LSTM layers are 128, 64, and 32 respectively, and the LSTM layer returns the hidden state values ​​of all time steps; Among the three dense layers, the number of neurons in the last dense layer is 1, and the number of neurons in the first two dense layers is optimized by the bionic meta-heuristic algorithm.

6. The method for regression prediction of the biological activity of a polypeptide according to claim 1, characterized in that: The bionic meta-heuristic algorithm is a whale optimization algorithm, a sparrow search algorithm or a Harris eagle algorithm.

7. The method for regression prediction of the biological activity of a polypeptide according to claim 1, characterized in that: In step (3), the evaluation index for evaluating the bioactivity regression prediction model is at least one of absolute error, root mean square error, goodness of fit and explained variance.

8. A regression prediction device for the biological activity of a polypeptide, characterized in that: comprising an activity prediction module, wherein the activity prediction module comprises one or more biological activity regression prediction models constructed by the method according to any one of claims 1 to 7; Substitute the peptide sequences to be tested into the selected biological activity regression prediction model in batches to obtain the corresponding biological activity data values ​​of the peptides.

Images