Peptide oxidation resistance prediction method and system based on pre-training model and descriptor

Abstract

The invention relates to the technical field of bioactive peptides, and particularly discloses a peptide oxidation resistance prediction method and system based on a pre-training model and descriptors, and the method comprises the following steps: obtaining an amino acid sequence of a to-be-predicted polypeptide and a physicochemical characteristic description index of each amino acid; enabling the coded amino acid sequence to pass through a BERT embedding layer to form high-dimensional vector representation, inputting the high-dimensional vector representation into a pre-trained deep neural network model to obtain features of a peptide sequence, and inputting the features of the peptide sequence into a pre-trained artificial neural network to obtain deep representation of the peptide sequence; according to the physicochemical characteristic description index of each amino acid, generating a characteristic vector of each amino acid, and constructing the characteristic of the polypeptide; and inputting the depth representation of the peptide sequence and the characteristics of the polypeptide into a predictor to obtain a prediction result of the oxidation resistance of the peptide. According to the method, the fine-tuned model characterization is combined with the traditional peptide sequence descriptor, so that the prediction accuracy of the antioxidant peptide and the stability of the model are successfully improved.

Description

Technical Field

[0001] The present invention relates to the technical field of bioactive peptides, and in particular to a method and system for predicting the antioxidant properties of peptides based on a pre-trained model and descriptors. Background Art

[0002] Antioxidant peptides generally need to be obtained by testing different peptides produced by hydrolyzing proteins with various proteases. This way of identifying antioxidant peptides is extremely slow and cumbersome, belonging to a trial-and-error method. With the continuous development of science and technology, scientists have paid more and more attention to developing methods for predicting antioxidant peptides based on computers, thereby reducing the workload in the laboratory. In recent years, peptide antioxidant predictors based on deep learning have been developed, and relatively excellent models include AnOxPePred and AnOxPP.

[0003] AnOxPePred is based on a convolutional neural network (CNN) model, which shows better prediction performance compared to the traditional k-nearest neighbor (k-NN) method and can more accurately predict the antioxidant properties of peptides. AnOxPP is based on a bidirectional long short-term memory (BiLSTM) neural network, combined with traditional amino acid descriptors, further improving the prediction accuracy and being superior to AnOxPePred. Although the application of neural networks has significantly accelerated the research process in predicting the antioxidant properties of peptides, there is still room for improvement in the prediction accuracy of these methods. How to more accurately extract the characteristics of polypeptides and effectively combine more biological information is an important challenge for further improving the prediction accuracy and model reliability. Summary of the Invention

[0004] The present invention aims to solve the problems of unsatisfactory prediction accuracy of peptide antioxidant properties and low model reliability. To this end, the present invention provides a method and system for predicting the antioxidant properties of peptides based on a pre-trained model and descriptors, which deeply integrates a large-scale pre-trained language model and traditional descriptors. By fine-tuning on a high-quality peptide antioxidant property dataset, it can more precisely capture the complex relationship between peptide characteristics and antioxidant properties. Combining the fine-tuned model representation with traditional peptide sequence descriptors forms a multi-faceted and multi-dimensional feature input, and a more efficient prediction model is obtained, successfully improving the prediction accuracy of antioxidant peptides and the stability of the model.

[0005] The present invention provides a method for predicting the antioxidant properties of peptides based on a pre-trained model and descriptors, and the technical solution adopted is as follows: including: Obtain the amino acid sequence of the polypeptide to be predicted and the physicochemical property description index of each amino acid; The encoded amino acid sequence is passed through the BERT embedding layer to form a high-dimensional vector representation, and then input into a pre-trained deep neural network model to obtain the features of the peptide sequence. Subsequently, it is input into a pre-trained artificial neural network to obtain the deep representation of the peptide sequence; According to the physicochemical property description indexes of each amino acid, the feature vector of each amino acid is generated to construct the characteristic features of the polypeptide; The deep representation of the peptide sequence and the characteristic features of the polypeptide are input into the predictor to obtain the prediction result of peptide antioxidant activity.

[0006] Furthermore, the deep neural network model adopts the Transformer structure with a multi-head self-attention mechanism; The artificial neural network includes three Linear layers, and a ReLU layer is set after each Linear layer.

[0007] Furthermore, the deep neural network model includes twelve consecutive Transformer layers, one fully connected layer, and one output layer. An 8-head self-attention mechanism is introduced in the Transformer layer.

[0008] Furthermore, the pre-training process of the deep neural network model and the artificial neural network is as follows: The deep neural network model is pre-trained using the pre-training dataset to obtain the pre-trained deep neural network model; The pre-trained deep neural network model is frozen, and an artificial neural network is connected behind it, and pre-training is carried out again to obtain the pre-trained artificial neural network.

[0009] Furthermore, the physicochemical property description indexes of amino acids include hydrophobicity, polarity, charge, size, amino acid composition, and GRID.

[0010] Furthermore, the hydrophobicity includes gravy, hopp-woods, janin, and kytedoolittle, The polarity includes polarity and grantham, The charge includes charge_phys, ISAECI, Z-Scales, and z-3, The size includes ABHPRK, argos, and bulkiness, The amino acid composition includes MSW, pepcats, and PPCALI, GRID includes t_scale.

[0011] Furthermore, the construction process of the characteristic features of the polypeptide is as follows: Normalize the physicochemical property description indexes of amino acids to obtain the normalized values of the indexes, and splice all the normalized values to obtain the amino acid feature vector; Arrange the amino acid feature vectors in order to obtain an intermediate feature vector. The dimension of the intermediate feature vector is [sequence length, feature dimension], where the sequence length is the number of amino acids in the peptide sequence, and the feature dimension is the number of indexes of each amino acid; Perform standardization or normalization processing on the intermediate feature vector to obtain the characteristic features of the polypeptide.

[0012] Furthermore, the predictor includes three layers of Linear layers and one layer of sigmoid layer, and a ReLU layer is set after each Linear layer.

[0013] Furthermore, the number of predictors is odd; When the number of predictors is multiple, a voting mechanism is adopted to determine the peptide antioxidant prediction result.

[0014] The present invention also provides a peptide antioxidant prediction system based on a pre-trained model and a descriptor. The technical solution adopted is as follows: including: a data acquisition module, a first feature calculation module, a second feature calculation module, and an antioxidant prediction module, The data acquisition module is used to acquire the amino acid sequence of the polypeptide to be predicted and the physicochemical property description indexes of each amino acid; The first feature calculation module is used to pass the encoded amino acid sequence through the BERT embedding layer to form a high-dimensional vector representation, input it into a pre-trained deep neural network model to obtain the features of the peptide sequence, and input it into a pre-trained artificial neural network to obtain the deep representation of the peptide sequence; The second feature calculation module is used to generate each amino acid feature vector according to the physicochemical property description indexes of each amino acid and construct the characteristic features of the polypeptide; The antioxidant prediction module is used to input the deep representation of the peptide sequence and the characteristic features of the polypeptide into the predictor to obtain the peptide antioxidant prediction result.

[0015] One or more of the above technical solutions in the embodiments of the present invention have at least one of the following technical effects: The present invention innovatively achieves a deep integration of the large model framework and traditional descriptors. The present invention adopts a unique feature fusion strategy: the deep representation of the peptide sequence is carefully refined through a pre-trained deep neural network model and an artificial neural network, combined with antioxidant peptide data. This process enables the deep representation of the peptide sequence to deeply mine complex and key feature information of antioxidant peptides in terms of molecular structure, chemical activity, functional group interaction, etc., possessing powerful representation ability and deep semantic understanding ability; while the characteristic features of the polypeptide are obtained based on traditional descriptors, covering basic composition information of the peptide chain reflected by amino acid composition, spatial conformation features contained in secondary structure information, and classical chemical and physical properties such as hydrophobicity. These traditional descriptors can provide stable and reliable structural information from the chemical essence level.

[0016] The present invention organically integrates the deep representation of the peptide sequence with the characteristic features of the polypeptide, giving full play to the powerful representation ability of the deep learning model and the complementary advantages of classical chemical descriptors, and further significantly improving the prediction performance. In the model structure, after ingeniously fusing the fine-tuned deep learning model with traditional descriptor features, deep feature learning is carried out through a deep neural network, so as to accurately capture more abundant and diverse peptide sequence information. The quantitative features extracted based on traditional descriptors provide additional structural information for the model, effectively helping to improve the accuracy and robustness of the prediction results. In order to achieve better performance, the present invention innovatively adopts a multi-layer neural network structure, enabling the model to efficiently explore and determine the optimal relationship pattern in the complex feature space, thus demonstrating excellent performance and high reliability in the prediction task of peptide antioxidant activity.

[0017] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the practice of the present invention. Brief Description of the Drawings

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 is a flowchart of the method provided by the present invention.

[0020] Figure 2 is a structural diagram of the pre-trained model of the method provided by the present invention.

[0021] Figure 3 is a structural diagram of the predictor provided by the present invention.

[0022] Figure 4 This is the accuracy result graph of the method provided by the present invention and the existing model.

[0023] Figure 5 This is the structural block diagram of the system provided by the present invention.

[0024] Reference numerals: 100, data acquisition module; 200, first feature calculation module; 300, second feature calculation module; 400, antioxidant property prediction module. Detailed implementation manners

[0025] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without making creative efforts based on the embodiments in the present invention belong to the scope protected by the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0026] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without conflict, those skilled in the art can combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0027] The following will be combined with Figures 1 to 5 to further elaborate on the present invention, and describe a method and system for predicting the antioxidant property of peptides based on a pre-trained model and descriptors of the present invention: In this embodiment, as Figure 1 shown, a method for predicting the antioxidant property of peptides based on a pre-trained model and descriptors is provided, including the following steps: Step 1: Obtain the amino acid sequence of the polypeptide to be predicted and the physicochemical property description indexes of each amino acid.

[0028] Determine the polypeptide to be predicted, obtain its amino acid sequence, and the physicochemical property description indexes of each amino acid.

[0029] Step 2: Encode the amino acid sequence, pass the encoded amino acid sequence through the BERT embedding layer to form a high-dimensional vector representation, input the high-dimensional vector representation into the pre-trained deep neural network model to obtain the features of the peptide sequence, and input the features of the peptide sequence into the pre-trained artificial neural network to obtain a deep representation of the peptide sequence.

[0030] The deep neural network model and artificial neural network used in this embodiment are pre-trained. Since there is no ready-made data set, this embodiment first collects and organizes data, which is divided into two parts: one part is the collection and organization of data for pre-training, and the other part is the collection and organization of antioxidant peptide and non-antioxidant peptide data. The details are as follows: (1) Data collection and organization for pre-training. The specific steps are as follows: S101, based on the excellent performance of protein databases in data comprehensiveness, as well as its significant advantages in data accuracy, data update timeliness, etc., chose the Uniprot database as the main database for retrieval. The Uniprot database is a comprehensive protein resource library that integrates protein sequences and functional information from many sources, covering a large amount of protein data identified from different species.

[0031] S102, downloading all data from the Uniprot database, including information such as sequence name, sequence length, sequence source, etc.

[0032] S103, according to the consensus reached in the peptide-related field, it is believed that sequences with a length of 1 to 50 can be considered peptides. Therefore, the sequence length is set to be sorted from small to large, and the search is performed to retain peptide data with a length of 1 to 50.

[0033] S104, the peptide sequences retained in the third step are deduplicated, and data with missing key information (such as incomplete or missing amino acid information in the sequence) are removed.

[0034] S105, after the above processing, 18913252 high-quality peptide sequence data are finally retained. The peptide sequences are extracted separately and saved as pre-training data sets for subsequent models.

[0035] (2) Collection and organization of antioxidant peptide and non-antioxidant peptide data. The specific steps are as follows: S201, downloaded peptide sequences from the DFBP database and BIOPEP-UWM database, used data cleaning technology to accurately remove duplicate data entries, and used rigorous experimental result consistency screening standards to screen out sample data with differences. After this refined processing, a data set of 1060 antioxidant peptides with free radical scavenging activity was successfully constructed and defined as a positive sample data set.

[0036] S202. For non - antioxidant peptides, through in - depth mining and sorting of a large amount of literature, 1060 peptide sequences that highly match the positive sample dataset in terms of sequence length dimension and quantity distribution trend are carefully selected as the negative sample dataset.

[0037] S203. 72 highly active antioxidant peptides that have no intersection with the above - mentioned positive and negative sample datasets are collected from the literature as the validation set of the model. The purpose is to strictly test and evaluate the accuracy, reliability, and generalization ability of the model constructed in the subsequent steps of this embodiment to ensure the effectiveness and scientific nature of the model in actual application scenarios.

[0038] To ensure that the pre - training data can be correctly recognized by the model, in this embodiment, the amino acid sequence is first encoded to convert the amino acid sequence into a numerical form that can be understood by a computer. For this purpose, integer encodings between 4 and 24 are assigned to 20 common amino acids. For example, the encoding of amino acid A is 4. In addition, it is also stipulated that 1 represents the start symbol of the sequence, 2 represents the end symbol of the sequence, 3 represents the padding symbol, and 0 represents non - standard amino acids. The role of the padding symbol is to pad the sequence to a unified length, ensuring that each input sequence has a consistent dimension, thereby making the model training process more stable and consistent. The use of the padding symbol can effectively handle variable - length sequences and ensure that the network is not affected when processing input sequences of different lengths. To improve the robustness and generalization ability of the model, a random masking strategy is also introduced. Specifically, the random masking will randomly mask a part of the information in the input sequence during the training process, prompting the model to learn more abundant and potential feature representations during training, thereby preventing overfitting and enabling it to better adapt to the missing or incomplete data that may exist in the real world.

[0039] The encoded amino acid sequence is passed through the BERT embedding layer. The BERT embedding layer (BERT Embedding) converts the encoded amino acid sequence into a high - dimensional vector representation, thereby capturing the semantic relationships and context information between amino acids in the sequence. The BERT embedding layer maps the numerical encoding of each amino acid into a high - dimensional space, enabling the model to represent each amino acid and its relationship with other amino acids in vector form, providing rich semantic features for the subsequent deep - learning module.

[0040] As Figure 2As shown, in this embodiment, a deep neural network model is constructed, adopting a Transformer structure with a multi-head self-attention mechanism introduced. Specifically, the deep neural network model includes twelve consecutive Transformer layers, a fully connected layer, and an output layer. An 8-head self-attention mechanism is introduced in the Transformer layer. A high-dimensional vector is used to represent the input to the pre-trained deep neural network model to obtain the features of the peptide sequence.

[0041] The 8-head self-attention mechanism can capture information from different subspaces by calculating multiple attention heads in parallel. This enables the model to flexibly handle complex context relationships and more effectively capture long-range dependencies in the input sequence. By stacking 12 Transformer layers, deep feature representations can be constructed, thereby enhancing the model's ability to capture complex patterns and long-range dependencies. Each Transformer layer can not only finely extract local features but also globally model the complex interactions between amino acids in the sequence, greatly enhancing the model's expressive power and learning ability.

[0042] The sequence representation after being processed by the Transformer layer will be passed to the fully connected layer (Fully Connected Layer), further enhancing the model's expressive ability through non-linear transformation to ensure that it can accurately output the final result according to the task requirements. The design of the output layer aims to provide appropriate prediction or classification results according to the specific goals of the task. During the training process, the output of the model will be compared with the original input sequence, and the difference between the predicted value and the true value will be calculated through the loss function (Loss Function) to guide the optimization of the model. The feedback of the loss function will be passed to the previous layers to gradually adjust the model's parameters. Through the backpropagation algorithm, the model will gradually optimize the weights in each iteration, approaching the optimal solution, thereby improving the overall performance.

[0043] After the construction of the deep neural network model is completed, in this embodiment, more than 18 million peptide data collected previously are used for pre-training. Through pre-training, it is ensured that the model can have sufficient learning ability and generalization performance in multiple tasks. The pre-trained deep neural network model is obtained by pre-training the deep neural network model using the pre-training dataset.

[0044] To optimize the training process, several key hyperparameters are set. Among them, the batch size is set to 32 to ensure that in each training epoch, the model can effectively process an appropriate amount of samples, thus achieving a good balance between computational efficiency and training stability. A smaller batch size helps reduce the gradient variance during each update, thereby enhancing the generalization ability of the model, while the value of 32 can ensure a sufficient amount of training samples to support complex model learning. In terms of the learning rate, a relatively conservative setting is adopted, and the learning rate is set to 5e-06. This learning rate selection can ensure the stable convergence of the model in the initial stage of training, while avoiding the phenomena of gradient explosion or oscillation caused by too large step sizes. The lower learning rate also helps to refine the optimization process, prompting the model to make more accurate weight adjustments in the later stage of training to gradually approach the global optimal solution. This setting helps to maintain high stability during the long-term training process and maximize the final performance of the model. During training, the number of epochs is set to 10.

[0045] This embodiment also constructs a neural network, as Figure 3 shown, which includes three Linear layers, three ReLU layers, and one sigmoid layer. The ReLU layers are arranged after the Linear layers. The first six layers of this neural network can be used as an artificial neural network to calculate and output the deep representation of the peptide sequence; all seven layers can be used as a predictor to calculate and output the peptide antioxidant prediction results.

[0046] The neural network adopts three Linear + ReLU layers to enhance the model's learning ability for complex features and improve its non-linear expression ability. After each layer of linear transformation, using the ReLU activation function can introduce non-linearity, enabling the model to capture complex patterns and potential relationships in the data. Specifically, the Linear layer performs a linear transformation on the input through a weight matrix, and the ReLU activation function, through non-linear mapping, enables the network to effectively learn richer and more abstract features. Through this design, the expressiveness of the model can be improved, helping it better adapt to the data features with complex non-linear relationships in the antioxidant peptide prediction task.

[0047] In this embodiment, the pre-trained deep neural network model is frozen, and an artificial neural network is connected behind it, and pre-training is carried out again to obtain the pre-trained artificial neural network. The artificial neural network includes three Linear layers, and a ReLU layer is arranged after each Linear layer.

[0048] The fine-tuning process of the artificial neural network. Specifically, freezing the parameters of the pre-trained deep neural network model can preserve the superiority of the pre-trained deep neural network model and enable the entire model to better adapt to specific task requirements. It can effectively reduce the computational amount, prevent overfitting, and enable the model to focus on learning the features related to antioxidant peptide prediction. This strategy enables the fine-tuning process to converge quickly and improves the prediction accuracy of the model in specific tasks.

[0049] In this embodiment, a deep neural network model and an artificial neural network (the fine-tuned model) obtained by pre-training and fine-tuning in combination with antioxidant peptide data are used to calculate the deep representation of the peptide sequence. Through the full integration and in-depth utilization of antioxidant peptide data, the fine-tuned model can accurately capture the key features of antioxidant peptides in multiple aspects such as molecular structure, chemical activity, and functional group interaction from a more comprehensive and in-depth perspective, and then convert these complex and diverse feature information into a deep representation of the peptide sequence with high representativeness and distinctiveness. The deep representation of the peptide sequence is generally saved in pkl format for use in the subsequent fusion steps.

[0050] Step 3: Generate the feature vector of each amino acid according to the physicochemical property description index of each amino acid, and construct the characteristic features of the polypeptide.

[0051] Different from the deep representation of the peptide sequence, the characteristic features of the polypeptide are constructed based on traditional amino acid descriptors. Traditional amino acid descriptors cover classic characterization information such as the physicochemical properties of amino acids, including hydrophobicity, hydrophilicity, charge characteristics, etc. By systematically analyzing and integrating these traditional descriptors corresponding to each amino acid in the peptide sequence, key information in another dimension reflecting the characteristics of the peptide can be extracted in a way based on the classical biochemistry theoretical framework, and then the information contained in these traditional amino acid descriptors can be converted into the characteristic features of the polypeptide with unique significance and value.

[0052] In this embodiment, the physicochemical property description indexes of amino acids include hydrophobicity, polarity, chargeability, size and volume, amino acid composition, and GRID. Specifically: (1) Hydrophobicity: gravy: The hydrophobicity score (Hydropathy Value) of an amino acid, used to describe the hydrophobic characteristics of the protein surface.

[0053] hopp-woods: A scale of amino acid hydrophobicity proposed by Hopp and Woods, which measures the tendency of amino acid residues towards a hydrophobic environment in an aqueous phase.

[0054] janin: The hydrophobicity score of Janin, which is used to evaluate the hydrophobicity of protein amino acids.

[0055] kytedoolittle: The Kyte-Doolittle hydrophobicity index, which represents the hydration and hydrophobicity of amino acids and is usually used to predict the structure of proteins.

[0056] (2)Polarity: polarity: Represents the polarity of amino acids, usually referring to the hydrophilicity degree of amino acid side chains in the aqueous phase. Amino acids with higher polarity are more soluble in water.

[0057] grantham: The Grantham polarity scale, which quantifies the polarity of amino acids and is usually used to predict the impact of amino acid substitutions on protein structure.

[0058] (3)Charge: charge_phys: Refers to the charge characteristics of amino acids under specific pH conditions, usually related to the acid-base properties of amino acids.

[0059] ISAECI: A scoring method based on the charge characteristics and structure of amino acids, which is used to predict the behavior of amino acids in proteins.

[0060] Z-Scales: z-3: Represents a scoring method based on the Z-scale, usually related to the three-dimensional structure stability of proteins, the interactions between amino acids, etc. z-5: Represents another Z-scale scoring method, similar to z-3 but may consider different parameters and interactions.

[0061] (4)Size and Bulkiness: ABHPRK: This index measures the volume or size of amino acids, usually described by calculating the space occupied by amino acid residues.

[0062] argos: The ARGOS scoring method, which is used to evaluate the volume characteristics of amino acids in proteins and the complexity of the three-dimensional space structure.

[0063] bulkiness: The volume index of amino acids, which is used to measure the space occupancy of amino acid side chains.

[0064] (5)Amino Acid Composition: MSW: An index of the diversity or complexity of amino acid composition, which is used to analyze the amino acid composition of proteins.

[0065] pepcats: Represents the compositional distribution of amino acid categories, usually referring to the proportions of different categories of amino acids in a protein.

[0066] PPCALI: Represents the category score of amino acid composition components, used to predict the amino acid type distribution in a protein.

[0067] (6)GRID (General Receptor and Ligand Interaction Descriptor): t_scale: Usually related to the three-dimensional spatial arrangement of a protein, t_scale is used to describe the physical or chemical properties at different positions in space.

[0068] The specific construction process of the characteristic features of the polypeptide is as follows: For each amino acid, detailed calculations are carried out based on the collected indicators, and the physicochemical property description indicators of the amino acid are normalized to obtain the normalized values of the indicators; ensure that each indicator is on a unified scale. Calculations can be performed using python software and the modlAMP software package.

[0069] All the normalized values are concatenated to obtain the amino acid feature vector.

[0070] The amino acid feature vectors of each amino acid in the peptide sequence are arranged in order to construct a complete tensor, obtaining the intermediate feature vector. The dimension of the intermediate feature vector is [sequence length, feature dimension], where the sequence length is the number of amino acids in the peptide sequence, and the feature dimension is the number of indicators for each amino acid.

[0071] To ensure the consistency of different features on the scale and avoid some features having too much or too little influence on the model. The intermediate feature vector is standardized or normalized to obtain the characteristic features of the polypeptide. Standardization or normalization is crucial for subsequent steps and can improve the training effect and prediction accuracy of the model. The characteristic features of the polypeptide are also saved in pkl format for use in fusion in subsequent steps.

[0072] Step 4: Input the deep representation of the peptide sequence and the characteristic features of the polypeptide into the predictor to obtain the peptide antioxidant property prediction result. The predictor includes three layers of Linear layers and one layer of sigmoid layer, and a ReLU layer is set after each Linear layer. The deep representation of the peptide sequence and the characteristic features of the polypeptide are concatenated and then input into the predictor. If the predictor outputs 1, it means that the peptide antioxidant property prediction result is that the polypeptide to be predicted belongs to the antioxidant peptide; if the predictor outputs 0, it means that the peptide antioxidant property prediction result is that the polypeptide to be predicted does not belong to the antioxidant peptide.

[0073] The number of predictors is odd and can be 1 or more. When there is 1 predictor, the peptide antioxidant prediction result can be directly obtained. When the number of predictors is multiple, for example, 5, a voting mechanism is adopted to determine the peptide antioxidant prediction result.

[0074] The predictors used in this embodiment are also pre-trained. The data collected and sorted in step 2, through steps 2 and 3 respectively, the depth representation of the peptide sequence and the characteristic features of the polypeptide calculated, and then the concatenated data obtained by concatenating the two form a concatenated dataset for the predictors to use in pre-training. It should be noted that in order to ensure the smooth progress of pre-training, a comprehensive data quality verification and evaluation work is also carried out on the depth representation of the peptide sequence and the characteristic features of the polypeptide, carefully checking whether there are data missing phenomena and abnormal values and other adverse conditions in these two types of feature vectors. If data is missing, that data entry is deleted; if an anomaly is found, the mean / median filling method is used for processing (if a certain feature value is missing, the missing part is filled with the mean or median of that feature) to ensure the integrity and accuracy of the data. Good concatenated data provides a solid and comprehensive data basis for constructing an accurate and efficient predictor in the follow-up, so as to ensure that the predictor can fully absorb and integrate key information from different feature dimensions.

[0075] During the training process, this embodiment adopts a five-fold cross-validation method, evenly dividing the concatenated dataset into five parts, where four parts are used as the training set for the parameter learning and optimization of the model, and the remaining one part is used as the validation set to evaluate the performance of the model on unseen data. Five-fold cross-validation can not only help detect the performance of the model on different data subsets, but also effectively reduce the bias that may be brought about by uneven data division, and improve the robustness and reliability of the model. After 10 epochs of training, five predictors are finally retained: 1 optimal model and 4 relatively good models.

[0076] To evaluate the performance of the predictor on peptide data that has not appeared in the training set and ensure the accuracy and generalization ability of the model, two different schemes were adopted. Scheme 1: Use all the data, train the final model under the same training strategy, and name this model PAnoxP. Scheme 2: Integrate five predictors and make decisions based on a voting mechanism. In this method, the results of five different predictors in cross-validation are integrated together, and the final prediction result is determined through a voting mechanism. Specifically, votes are cast according to the prediction outputs of each predictor. If at least three predictors predict the result as 1, the final prediction result is regarded as 1; otherwise, it is regarded as 0. Let PAnoxP-V represent the result of the model voting mechanism. This integration method aims to improve the overall performance by combining the prediction results of multiple models, thereby reducing the errors that may occur in a single model and enhancing the accuracy and robustness of the prediction. During this process, accuracy is selected as the core evaluation index, and its calculation method is the ratio of the number of correctly predicted samples to the total number of samples. After obtaining this index, this embodiment compares and analyzes it with existing predictors of the same type.

[0077] The results are as Figure 4 shown. The accuracy of PAnoxP is 0.8750, which is significantly better than the existing excellent predictor AnOxPP. This result indicates that this method has significant advantages in predicting peptide antioxidant properties and performs well in terms of accuracy and robustness, further verifying the feasibility and effectiveness of this method. The accuracy of PAnoxP-V is 0.8194, which is comparable to the performance of the existing excellent predictor AnOxPP. This result indicates that this method matches the current excellent predictors in terms of accuracy, further proving the feasibility and effectiveness of this method. Based on the above analysis, the conclusion is drawn that PAnoxP based on all the data shows significant advantages, indicating that combining the fine-tuned model with traditional descriptors can significantly improve the prediction performance and enhance the generalization ability of the model. This feature fusion technology provides a comprehensive and efficient solution for the field of peptide antioxidant prediction and demonstrates its great potential in bioinformatics and drug design.

[0078] In this embodiment, as Figure 5 shown, a peptide antioxidant prediction system based on a pre-trained model and descriptors is also provided. The technical solution adopted is as follows: It includes: a data acquisition module 100, a first feature calculation module 200, a second feature calculation module 300, and an antioxidant prediction module 400.

[0079] The data acquisition module 100 is used to obtain the amino acid sequence of the polypeptide to be predicted and the physicochemical property description indexes of each amino acid; The first feature calculation module 200 is configured to form a high-dimensional vector representation by passing the encoded amino acid sequence through a BERT embedding layer, input it into a pre-trained deep neural network model to obtain the features of the peptide sequence, and input it into a pre-trained artificial neural network to obtain the deep representation of the peptide sequence; The second feature calculation module 300 is configured to generate a feature vector for each amino acid according to the physicochemical property description index of each amino acid, and construct the characteristic features of the polypeptide; The antioxidant property prediction module 400 is configured to input the deep representation of the peptide sequence and the characteristic features of the polypeptide into a predictor to obtain the peptide antioxidant property prediction result.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors, characterized in that: include: Obtaining the amino acid sequence of the polypeptide to be predicted and the physical and chemical property description index of each amino acid; The encoded amino acid sequence is passed through the BERT embedding layer to form a high-dimensional vector representation, which is then input into a pre-trained deep neural network model to obtain the features of the peptide sequence, and then input into a pre-trained artificial neural network to obtain a deep representation of the peptide sequence; Based on the physical and chemical property description index of each amino acid, a feature vector of each amino acid is generated to construct the characteristic features of the peptide; The deep characterization of peptide sequences and the characteristic features of peptides were input into the predictor to obtain the peptide antioxidant activity prediction results.

2. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors as claimed in claim 1, characterized in that: The deep neural network model adopts the Transformer structure that introduces a multi-head self-attention mechanism; The artificial neural network consists of three Linear layers, and a ReLU layer is set after each Linear layer.

3. A peptide antioxidant activity prediction method based on a pre-trained model and a descriptor as claimed in claim 2, characterized in that: The deep neural network model includes twelve sequentially connected Transformer layers, a fully connected layer and an output layer, and an 8-head self-attention mechanism is introduced in the Transformer layer.

4. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors according to any one of claims 1 to 3, characterized in that: The process of pre-training deep neural network models and artificial neural networks is: Pre-training the deep neural network model using the pre-training data set to obtain a pre-trained deep neural network model; Freeze the pre-trained deep neural network model, and then connect the artificial neural network to it, and perform pre-training again to obtain the pre-trained artificial neural network.

5. A peptide antioxidant activity prediction method based on a pre-trained model and a descriptor as claimed in claim 1, characterized in that: The physical and chemical properties of amino acids can be described by indicators including hydrophobicity, polarity, charge, size, volume, amino acid composition and GRID.

6. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors as claimed in claim 5, characterized in that: Hydrophobicity includes gravy, hopp-woods, janin and kytedoolittle, Polarity includes polarity and grantham. Charge properties include charge_phys, ISAECI, Z-Scales and z-3, Size includes ABHPRK, argos and bulkiness. The amino acid composition includes MSW, pepcats and PPCALI, GRID includes t_scale.

7. A peptide antioxidant activity prediction method based on a pre-trained model and a descriptor as claimed in claim 1, 5 or 6, characterized in that: The process of constructing the characteristic features of a peptide is as follows: Normalize the physical and chemical property description indexes of amino acids to obtain normalized values ​​of the indexes, and concatenate all normalized values ​​to obtain amino acid feature vectors; Arrange the amino acid feature vectors in order to obtain an intermediate feature vector whose dimension is [sequence length, feature dimension], where the sequence length is the number of amino acids in the peptide sequence and the feature dimension is the number of indicators for each amino acid; The intermediate feature vectors are standardized or normalized to obtain the characteristic features of the peptide.

8. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors as claimed in claim 1, characterized in that: The predictor consists of three linear layers and one sigmoid layer, and each linear layer is followed by a ReLU layer.

9. A peptide antioxidant activity prediction method based on a pre-trained model and descriptors as claimed in claim 1, characterized in that: The number of predictors is an odd number; When there are multiple predictors, a voting mechanism is used to determine the peptide antioxidant activity prediction results.

10. A peptide antioxidant prediction system based on pre-trained models and descriptors, characterized in that: The method for predicting the antioxidant activity of a peptide based on a pre-trained model and a descriptor according to any one of claims 1 to 9 comprises: a data acquisition module, a first feature calculation module, a second feature calculation module and an antioxidant prediction module. A data acquisition module is used to obtain the amino acid sequence of the polypeptide to be predicted and the physical and chemical property description index of each amino acid; The first feature calculation module is used to pass the encoded amino acid sequence through the BERT embedding layer to form a high-dimensional vector representation, input it into the pre-trained deep neural network model to obtain the features of the peptide sequence, and input it into the pre-trained artificial neural network to obtain the deep representation of the peptide sequence; The second feature calculation module is used to generate a feature vector for each amino acid based on the physical and chemical property description index of each amino acid, and construct the characteristic features of the polypeptide; The antioxidant activity prediction module is used to input the deep characterization of peptide sequences and the characteristic features of peptides into the predictor to obtain the peptide antioxidant activity prediction results.

Images