High-throughput screening method and device for antioxidant polypeptides

Abstract

The invention discloses a high-throughput screening method and device for antioxidant polypeptides, and the method comprises the following steps: (1) collecting polypeptides to construct a polypeptide data set which comprises antioxidant peptides, non-antioxidant peptides and random peptides; (2) encoding polypeptides in the polypeptide data set, extracting features, forming vector representation, and dividing into a training set and a test set; (3) training the learning model by adopting the training set and the test set to obtain a prediction model of the antioxidant polypeptide; and (4) substituting to-be-screened polypeptide sequences into the prediction model in batches, and outputting the antioxidant probability of the polypeptide by the prediction model. According to the method and the device, the oxidation resistance of the polypeptide can be quickly predicted in a high-throughput manner, and meanwhile, relatively high accuracy and a model Morse correlation coefficient are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of bioactive peptides, and in particular to a high-throughput screening method and device for antioxidant peptides based on confidence learning and protein language models. Background Art

[0002] Oxidation is an important reaction in biological processes that can produce free radicals. Free radicals are atomic groups containing an unpaired electron, which are usually unstable and highly reactive. A small amount of active free radicals in the body have certain functions, such as immunity and signal transduction processes. However, too many active free radicals will have destructive behavior, leading to damage to normal cells and tissues in the human body, thereby causing a variety of diseases such as heart disease, Parkinson's disease and tumors. Free radicals can also be produced in food. When the oils, additives, pigments, etc. in food are exposed to ultraviolet rays, heavy metals or air with free radicals, oxidation reactions will occur to produce free radicals, which will lead to bad odors, changes in texture and compounds that are harmful to health. Antioxidants are a class of substances that can help capture and neutralize free radicals, thereby inhibiting the damage of free radicals to the human body. Therefore, adding antioxidants to food is an important method to prevent or delay food oxidation, improve food stability and extend storage life. As consumers have higher requirements for food safety and health, the use of peptides derived from food proteins as antioxidants has attracted more and more attention from researchers.

[0003] Confident learning is an emerging, principled framework that can identify label errors, characterize label noise, and apply it to noisy learning. Confident learning can find incorrectly labeled samples in a dataset and directly estimate the joint distribution of noisy labels and true labels, which is theoretically reasonable. At the same time, confident learning can use any model, unlike many noisy learning methods that are strongly coupled with the model and training process. Currently, confident learning has been successfully used to find label errors in popular computer vision and natural language processing (NLP) benchmark datasets such as MINIST.

[0004] 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.

[0005] Machine learning and deep learning are branches of artificial intelligence. By using statistical and computer science methods, computer systems can extract patterns, rules and knowledge from large amounts of data, and use this knowledge to perform tasks such as prediction, classification and identification, so that computer systems can learn from data and make decisions or predictions based on the learned knowledge.

[0006] Moreover, the existing technology rarely uses the method of combining confidence learning and ESM-2 to optimize benchmark data sets and predict antioxidant peptides, and the antioxidant peptide classification models in existing patents and literature perform poorly. The present invention can make up for these shortcomings, quickly and high-throughput obtain antioxidant peptide information, and ensure a high accuracy rate and model Matthews correlation coefficient, providing a reference for the study of food-derived bioactive peptides. Summary of the invention

[0007] The present invention provides a method and device for high-throughput screening of antioxidant polypeptides, which can quickly and high-throughput predict the antioxidant properties of polypeptides while ensuring a high accuracy rate and model Matthews correlation coefficient.

[0008] The technical solution of the present invention is as follows:

[0009] A high-throughput screening method for antioxidant polypeptides comprises the following steps:

[0010] (1) collecting peptides to construct a peptide data set, wherein the peptide data set includes antioxidant peptides, non-antioxidant peptides and random peptides;

[0011] (2) Encode the peptides in the peptide dataset and extract features to form a vector representation, which is divided into a training set and a test set;

[0012] (3) Using the training set and the test set to train the learning model, a prediction model for antioxidant peptides was obtained;

[0013] (4) Substituting the peptide sequences to be screened into the prediction model in batches, the prediction model outputs the antioxidant probability of the peptides.

[0014] In step (1), the collected peptides come from the data sets of the antioxidant peptide classification models UniDL4BioPep and AnOxPePred published in the literature.

[0015] The peptide data set contains 1404 peptides, including 687 antioxidant peptides, 217 non-antioxidant peptides and 500 random peptides. Among these peptides, antioxidant activity is defined as 1, and no antioxidant activity is defined as 0. These peptides have been tested by existing peptide classification models and have high quality.

[0016] In step (2), the polypeptide may be encoded using AAC (amino acid composition), DPC (dipeptide composition), CTDC (physical and chemical property composition), CTDT (physical and chemical property conversion), CTDD (physical and chemical property distribution), GAAC (grouped amino acid composition) or using the Evolutionary Scale Modeling-2 (ESM-2) protein language model.

[0017] The physicochemical properties include: hydrophobicity, van der Waals forces, polarity, polarizability, charge properties, surface tension, secondary structure, and solvent accessibility.

[0018] AAC generates 20 feature vectors, DPC generates 400 feature vectors, CTDC generates 38 feature vectors, CTDT generates 194 feature vectors, CTDD generates 38 feature vectors, GAAC generates 4 feature vectors, and the ESM-2 protein language model is used for encoding to generate 320 feature vectors.

[0019] Preferably, in step (2), the Evolutionary Scale Modeling-2 (ESM-2) protein language model is used for encoding.

[0020] 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 the perspective of word vector features, position encoding features, residue-residue interaction features and window features. It is more comprehensive than traditional peptide embedding methods such as sequence embedding and physical and chemical property embedding.

[0021] Preferably, step (2) comprises:

[0022] (2-1) The Evolutionary Scale Modeling-2 protein language model is used to encode the peptides in the peptide dataset and extract features to form vector representations;

[0023] (2-2) The peptide dataset after feature extraction is cleaned using the Cleanlab library and divided into a training set and a test set after cleaning.

[0024] The peptide dataset after feature extraction is cleaned using the Cleanlab library, including:

[0025] The peptide data after feature extraction is loaded into the LogisticRegression model for training, and the cross_val_split function in the sklearn library is used for five-fold cross validation to output the predicted probability of antioxidant activity for each peptide;

[0026] The peptide features and predicted probabilities were input into the Cleanlab library, and the find_label_issues function was used to find label errors in the data set. The data with label errors were deleted to obtain the cleaned data set.

[0027] The present invention combines confidence learning Cleanlab with the latest protein language model Evolutionary Scale Modeling 2 to perform label cleaning on the collected benchmark data set, effectively avoiding the degradation of data set quality caused by experimental operation errors in different laboratories in the previous development of polypeptide antioxidant classification models, making the invented model more convincing.

[0028] In step (3), the learning models include: K-Nearest Neighbor (KNN) model, Random Forest (RF) model, Recurrent Neural Network (RNN) model, Convolutional Neural Network (CNN) model, Multi-Layer Perceptron (MLP) model, LightGBM (LightGradient Boosting Machine, LGBM) model, Logistic Regression (LR) model, and Support Vector Machine (SVM) model.

[0029] Preferably, in step (2), the Evolutionary Scale Modeling-2 protein language model is used to encode the polypeptide; in step (3), the learning model is a support vector machine model.

[0030] When this technical solution is adopted, the balance accuracy of the model is as high as 94.8%, and the Matthews correlation coefficient is as high as 0.892.

[0031] In step (4), when the prediction model outputs that the antioxidant probability of the polypeptide is greater than 80%, the polypeptide is considered to be a potential antioxidant polypeptide.

[0032] The present invention also provides a high-throughput screening device for antioxidant polypeptides, comprising a data analysis module, wherein the data analysis module comprises a prediction model for antioxidant polypeptides;

[0033] The peptide sequences to be screened are batch-substituted into the prediction model, and the prediction model outputs the antioxidant probability of the peptide.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] The present invention can rapidly, efficiently and accurately screen food-borne bioactive peptides with antioxidant properties, while reducing the screening cycle and cost, thus providing assistance for the future development of food-borne bioactive peptides. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a high-throughput antioxidant peptide screening method according to an embodiment of the present invention;

[0037] Figure 2 Flowcharts developed and applied for the model of embodiments of the present invention;

[0038] Figure 3 Compare the performance of ESM-2 embedding and 6 other traditional embedding models on the SVM model;

[0039] Figure 4 Comparison of Matthews correlation coefficient values ​​for the models of six machine learning methods, two deep learning methods combined with ESM-2, and six other peptide embedding methods;

[0040] Figure 5 Comparison of equilibrium accuracy values ​​of the models of 6 machine learning methods, 2 deep learning methods combined with ESM-2, and other 6 peptide embedding methods. DETAILED DESCRIPTION

[0041] 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.

[0042] Example:

[0043] like Figure 1 and Figure 2 As shown, the high-throughput screening method for antioxidant polypeptides based on confidence learning and protein language model of the present invention comprises the following steps:

[0044] S1. Collect antioxidant peptide benchmark datasets from the best publicly available antioxidant peptide classification models;

[0045] The antioxidant peptides screened here are all from the data sets of the existing antioxidant peptide classification models. This is to make a fair comparison of the model performance and prevent the intentional design of the data set from resulting in good performance but poor generalization ability. The benchmark data set of antioxidant peptides in S1 is completely from the best antioxidant peptide classification models UniDL4BioPep and AnOxPePred published in the literature, with a total of 1404 peptides, including 687 antioxidant peptides, 217 non-antioxidant peptides and 500 peptides randomly retrieved from the Uniprot database. Among these peptides, antioxidant activity is defined as 1, and non-antioxidant activity is defined as 0.

[0046] S2, use multiple peptide embedding methods to extract peptide features and form a unique vector representation for each peptide;

[0047] The specific peptide embedding methods include: AAC amino acid composition, dipeptide composition DPC, CTDC composition-conversion-distribution composition, CTDT composition-conversion-distribution conversion, CTDD composition conversion-distribution distribution, GAAC grouping amino acid composition and Evolutionary Scale Modeling-2 (ESM-2) protein language model, so as to express the characteristics of peptides. Their embedding methods include the following steps:

[0048] S2-1. Load the FASTA format sequence of the data set into txt format, input the iLearnPlus-Omega program, calculate AAC, DPC, CTDC, CTDT, CTDD and GAAC respectively, extract the feature representation of each sequence from the feature results and save them, among which AAC generates 20 feature vectors, DPC generates 400 feature vectors, CTDC generates 38 feature vectors, CTDT generates 194 feature vectors, CTDD generates 38 feature vectors, and GAAC generates 4 feature vectors.

[0049] S2-2. Use the Pandas library to read the protein sequence data in the Excel file, convert each protein sequence data into a format acceptable to the Evolutionary Scale Modeling-2 protein language model, then load the Evolutionary Scale Modeling-2 protein language model, calculate the feature representation, and generate 320 feature vectors for each peptide.

[0050] S3. The ESM-2 feature extraction dataset was cleaned using the Cleanlab program based on the confidence learning principle, and the vector representations of related peptides in all feature extraction methods were cleared according to the noise in the recognition. After cleaning, it was divided into training set and test set in a ratio of 8:2. The method of using Evolutionary Scale Modeling-2 is as follows:

[0051] S3-1. 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.

[0052] S3-2. Use the pre-trained version esm2_t6_8M_UR50D() in ESM-2, which has 6 representation layers. Select the last representation layer. Use Pandas to read the antioxidant peptide dataset, convert it into a format acceptable to ESM-2, perform peptide representation on the antioxidant peptide dataset, and generate 320 feature vectors for each peptide.

[0053] How to use Cleanlab:

[0054] S3-3. Use Pandas to read the feature representation dataset table after ESM-2 representation, load the LogisticRegression model for training, use the cross_val_split function in the sklearn library for five-fold cross validation, and output the predicted probability of antioxidant activity for each peptide;

[0055] S3-4. Input the output peptide features and predicted probabilities into the Cleanlab library, use the find_label_issues function to find label errors in the data set, and delete all data with "True" in the filter label to obtain the cleaned data set. This function is a confidence learning method that directly estimates the joint distribution between given labels and unknown labels, identifies noise data, and enhances the quality of the data set.

[0056] S4. Use a variety of peptide representations in combination with multiple machine learning models and deep learning models (including: K-Nearest Neighbor (KNN) model, Random Forest (RF) model, Recurrent Neural Network (RNN) model, Convolutional Neural Network (CNN) model, Multi-Layer Perceptron (MLP) model, LightGBM (Light Gradient Boosting Machine, LGBM) model, Logistic Regression (LR) model, Support Vector Machine (SVM) model) to predict the post-cleaning antioxidant peptide data set and obtain multiple prediction models;

[0057] S5. Use a variety of indicators to evaluate the performance of the above prediction models. Take the Matthews correlation coefficient as the selection criterion and select the best models, which are Evolutionary Scale Modeling-2 (ESM-2) and SVM models.

[0058] refer to Figure 3-Figure 5 Table 1 shows the Matthews correlation coefficient and equilibrium accuracy of the peptide embedding method + model structure constructed by combining 7 different peptide embedding methods, 6 machine learning models and 2 deep learning models in the present invention. Figure 4 and Figure 5 It can be seen that the balance accuracy and Matthews correlation coefficient of the model after Cleanlab cleaning are higher than those of the model before cleaning; Figure 4 and Figure 5 It can be seen that the Matthews correlation coefficient and balanced accuracy of the ESM-2+SVM model are the highest, which are 0.892 and 0.948 respectively; the closer the balanced accuracy value and the Matthews correlation coefficient value are to 1, the stronger the binary classification ability of the corresponding model combination. Figure 3 It can also be seen that the ESM-2 peptide embedding method is also superior to other traditional peptide embedding methods.

[0059] Table 1 Comparison of the optimal peptide embedding combination and model performance with the optimal model performance in current research

[0060]

[0061] S6. Substitute the batch polypeptide sequences into the model, and the polypeptides predicted by the model output with an antioxidant probability greater than 80% are considered as potential antioxidant polypeptides;

[0062] The generalization ability of the model was tested using peptide data outside the benchmark dataset. The model tested 20 antioxidant peptides at one time, and correctly identified 7 out of 8 antioxidant peptides and 7 out of 8 non-antioxidant peptides, with an accuracy rate of 87.5%, which has high practical application value.

[0063] S7. Develop local applications based on the above optimal model.

[0064] Users can use the program directly on their local computers to upload large quantities of peptide sequences or input a single sequence to predict the antioxidant value of the peptide in real time. The application outputs the antioxidant prediction results corresponding to each peptide sequence, and if the peptide is predicted to be an antioxidant peptide, it will also output the prediction probability to help users better judge its antioxidant activity. Users do not need to have any relevant knowledge, just upload the sequence to complete the prediction.

[0065] 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 high-throughput screening method for antioxidant polypeptides, characterized in that: The following steps are involved: (1) collecting peptides to construct a peptide data set, wherein the peptide data set includes antioxidant peptides, non-antioxidant peptides and random peptides; (2) Encode the peptides in the peptide dataset and extract features to form a vector representation, which is divided into a training set and a test set; (3) Using the training set and the test set to train the learning model, a prediction model for antioxidant peptides was obtained; (4) Substituting the peptide sequences to be screened into the prediction model in batches, the prediction model outputs the antioxidant probability of the peptides.

2. The high-throughput screening method for antioxidant polypeptides according to claim 1, characterized in that: In step (2), the polypeptide is encoded in the following ways: AAC, DPC, CTDC, CTDT, CTDD, GAAC or using the Evolutionary Scale Modeling-2 protein language model for encoding.

3. The high-throughput screening method for antioxidant polypeptides according to claim 2, characterized in that: AAC generates 20 feature vectors, DPC generates 400 feature vectors, CTDC generates 38 feature vectors, CTDT generates 194 feature vectors, CTDD generates 38 feature vectors, GAAC generates 4 feature vectors, and the ESM-2 protein language model is used for encoding to generate 320 feature vectors.

4. The high-throughput screening method for antioxidant polypeptides according to claim 1, characterized in that: Step (2) comprises: (2-1) The Evolutionary Scale Modeling-2 protein language model is used to encode the peptides in the peptide dataset and extract features to form vector representations; (2-2) The peptide dataset after feature extraction is cleaned using the Cleanlab library and divided into a training set and a test set after cleaning.

5. The high-throughput screening method for antioxidant polypeptides according to claim 4, characterized in that: The peptide dataset after feature extraction is cleaned using the Cleanlab library, including: The peptide data after feature extraction is loaded into the LogisticRegression model for training, and the cross_val_split function in the sklearn library is used for five-fold cross validation to output the predicted probability of antioxidant activity for each peptide; The peptide features and predicted probabilities were input into the Cleanlab library, and the find_label_issues function was used to find label errors in the data set. The data with label errors were deleted to obtain the cleaned data set.

6. The high-throughput screening method for antioxidant polypeptides according to claim 1, characterized in that: In step (3), the learning models include: K nearest neighbor model, random forest model, recurrent neural network model, convolutional neural network model, multi-layer perception model, LightGBM model, logistic regression model, and support vector machine model.

7. The high-throughput screening method for antioxidant polypeptides according to claim 1, characterized in that: In step (2), the Evolutionary Scale Modeling-2 protein language model is used to encode the polypeptide; in step (3), the learning model is a support vector machine model.

8. The high-throughput screening method for antioxidant polypeptides according to claim 1, characterized in that: In step (4), when the prediction model outputs that the antioxidant probability of the polypeptide is greater than 80%, the polypeptide is considered to be a potential antioxidant polypeptide.

9. A high-throughput screening device for antioxidant polypeptides, comprising a data analysis module, wherein the data analysis module comprises a prediction model for antioxidant polypeptides constructed by the method according to any one of claims 1 to 8; The peptide sequences to be screened are batch-substituted into the prediction model, and the prediction model outputs the antioxidant probability of the peptide.

Images