Accurate active peptide recognition method fusing polypeptide three-dimensional characteristics and deep learning

Abstract

The invention relates to an active peptide intelligent identification method fusing polypeptide three-dimensional structure characteristics and deep learning. Atomic-scale three-dimensional structure information of polypeptide is obtained through AlphaFold-3, and multi-dimensional tensor characterization is constructed in combination with amino acid sequence characteristics; a two-stage deep learning framework is innovatively designed, in the first stage, an improved auto-encoder is adopted to carry out nonlinear dimensionality reduction on original high-dimensional features, in the second stage, a convolution auto-encoder is adopted to extract local spatial features, and a noise filtering module is introduced to enhance model robustness. According to the method, the antioxidant peptide is taken as a research object, the training accuracy of 99.34% and the independent verification accuracy of 98.29% are realized on 2916 polypeptide data sets, the method is successfully migrated to an anti-aging peptide recognition task, and the prediction performance of 98% or above is kept. According to the method, a new structure-function associated active peptide recognition normal form is established, an efficient and accurate calculation tool is provided for discovery of polypeptide drugs, and the method has active peptide recognition universality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of polypeptides, and particularly relates to a method for accurately identifying active peptides by combining fusion polypeptide three-dimensional features and deep learning. BACKGROUND

[0002] In recent years, with the rapid development of artificial intelligence technology, deep learning has shown great application potential in the field of biological medicine. As a key bioactive substance between small molecules and proteins, active polypeptides have become an important target for the development of new drugs due to their high specificity, low immunogenicity and designability. Traditional active peptide recognition mainly relies on two methods: (1) empirical models based on sequence-activity relationship, which have the limitation of only capturing shallow features such as amino acid composition; (2) experimental screening methods, which have the bottleneck of long cycle and high cost. Deep learning models can automatically extract more complex feature patterns, significantly improving the efficiency and accuracy of polypeptide function prediction. So far, deep learning strategies have been widely used in the identification of functional peptides such as antibacterial peptides, antioxidant peptides and anticancer peptides, indicating that deep learning has become an important tool for active peptide research.

[0003] Existing deep learning models are usually optimized for a single activity type, and this "one task one model" paradigm cannot share the structure-activity relationship knowledge between different active peptides, resulting in the need to model from scratch for new active peptide recognition. Moreover, current polypeptide feature representation strategies mainly rely on basic information such as amino acid sequences or physicochemical properties, which limits the accuracy and universality of deep learning model prediction, making it difficult to establish a general method suitable for active peptide screening. Recently, a study published in Nature Biomedical Engineering (2024) pointed out that the biological activity of polypeptides is closely related to their three-dimensional conformational dynamics, and introducing polypeptide three-dimensional features into deep learning models will greatly improve the recognition performance of active peptides. However, the conformational flexibility of polypeptides poses challenges to traditional molecular dynamics simulation in terms of conformational sampling efficiency and energy optimization accuracy. More importantly, existing methods lack effective three-dimensional feature representation strategies, making it difficult to integrate structural information into deep learning frameworks. AlphaFold-3 is an artificial intelligence program developed by the DeepMind team, the winner of the 2024 Nobel Prize, to predict the structure and interaction of biological molecules such as polypeptides and proteins and their complexes. The release of AlphaFold-3 provides a new approach to the above problems: using AlphaFold-3 to predict the three-dimensional structure of polypeptides and extract polypeptide three-dimensional features, a multi-scale three-dimensional feature representation strategy can be established. This method can overcome the limitations of pure sequence analysis and avoid the one-sidedness of a single conformation, providing a more comprehensive structural information basis for active peptide screening. Therefore, introducing polypeptide three-dimensional features into deep learning with the help of AlphaFold-3 will help break through the limitations of traditional sequence analysis. SUMMARY

[0004] The present application provides a new strategy for fusing polypeptide three-dimensional features with deep learning for the purpose of accurate identification of active peptides. By integrating the polypeptide structure information output by AlphaFold-3, a multi-scale feature representation system is constructed, including atomic three-dimensional spatial coordinates and amino acid sequence information, etc. On this basis, a deep learning model based on autoencoder is developed to achieve accurate identification of different active peptides. At the same time, the feature dimension of active peptide identification is spanned from sequence features to three-dimensional conformation.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] An active peptide accurate identification method fusing polypeptide three-dimensional features with deep learning, comprising the following steps:

[0007] Step 1: Collect several reported antioxidant peptides from AODB database, and randomly obtain several polypeptides from UniProt database as non-antioxidant peptides. Combine and randomly arrange the several antioxidant peptides and several non-antioxidant peptides to form a data set, and use 80%-90% of the data set as a training set for model training, and the remaining data as an independent validation set for model verification after training.

[0008] Step 2: Optimize the three-dimensional structure of the polypeptide sequence in step 1 using AlphaFold-3 program, extract atomic number, atomic type, three-dimensional coordinates (X / Y / Z), and amino acid type and amino acid sequence information, and form a multi-dimensional matrix.

[0009] Step 3: Construct a deep learning model based on autoencoder, train and verify the data in the training set, and obtain a deep learning model that can accurately identify antioxidant peptides.

[0010] Step 4: Use the deep learning model obtained in step 3 to predict the activity of polypeptides in the independent validation set, and calculate the prediction accuracy.

[0011] Obtain several reported anti-aging peptides from AagingBase database, and randomly obtain several polypeptides from UniProt database as non-anti-aging peptides. Combine the data set and divide the training set and independent validation set according to the above step 1, extract the three-dimensional features of the anti-aging peptides according to the above step 2, train the model according to the above step 3, and verify the accuracy of the strategy to identify anti-aging peptides according to the above step 4.

[0012] Preferably, in step 1, the number ratio of antioxidant peptides and non-antioxidant peptides is 4:6-6:4, and preferably, the number of antioxidant peptides and non-antioxidant peptides is consistent, both being 1458.

[0013] Preferably, in step 2, the multi-dimensional matrix of the optimized polypeptide structure includes the atomic number, atomic species, corresponding X coordinate, Y coordinate, Z coordinate, and amino acid species and amino acid sequence information of the extracted polypeptide; the matrix dimension is 7 x the number of atoms. Since the number of atoms of different polypeptides is inconsistent, the "0" padding method is used to make the dimensions of all matrices consistent.

[0014] Preferably, in step 3, the deep learning model mainly includes two algorithms based on autoencoder, the first one is a traditional autoencoder (AE) which realizes data dimension reduction by adopting a three-layer fully connected structure, contains a 256-dimensional hidden layer and one potential hidden space, and the reduced data is stored in the hidden layer; the second one is a convolutional autoencoder (CAE) which is composed of a symmetric three-layer convolutional encoder and a decoder, equipped with a parallel noise processing branch, adopts a 5x5 convolution kernel, and uses a dropout layer to prevent overfitting, AE and CAE are trained for 2000 rounds, the batch size is 8, and a dynamic learning rate adjustment strategy is adopted.

[0015] Preferably, in step 4, the features of polypeptides in the training set are learned and summarized, and the polypeptides in the independent validation set are verified, outputting "0" and "1", "0" indicating that the polypeptide is a non-antioxidant peptide, and "1" indicating that the polypeptide is an antioxidant peptide.

[0016] The deep learning model constructed for accurately identifying antioxidant peptides is applied to the identification of anti-aging peptides, and all structures and parameters of the model are consistent, only the anti-aging peptides are replaced with antioxidant peptides as the input of the model.

[0017] Preferably, the number of anti-aging peptides and non-anti-aging peptides is 311.

[0018] Compared with the prior art, the beneficial effects of the present application are:

[0019] 1、The present application first constructs a deep learning identification framework that fuses the three-dimensional spatial information of polypeptides, and develops a new strategy for high-precision active peptide identification. The present application aims to identify active peptides, constructs a data set containing active peptides and non-active peptides, based on the data set, constructs a deep learning model based on autoencoder, and trains it to identify active peptides. The present application integrates the atomic information and sequence information of the polypeptides extracted by AlphaFold-3 calculation optimization to establish a multi-dimensional feature matrix, breaks through the limitation of deep learning model relying only on sequence and one-dimensional features, and realizes the paradigm shift from "sequence analysis" to "three-dimensional structure driven".

[0020] 2, The application innovatively develops a dual-channel self-encoder fusion model architecture. A hybrid structure of a traditional autoencoder (AE) and a convolutional autoencoder (CAE) is adopted, wherein the AE is responsible for global feature compression, and the CAE processes local spatial features, and the model robustness is enhanced through a parallel noise processing branch, and finally an accuracy of 99.34% is achieved in the antioxidant peptide recognition task, which is higher than that of existing antioxidant peptide recognition models.

[0021] 3, The application first combines the three-dimensional features of polypeptides with a self-encoder deep learning model to identify antioxidant peptides, which is more efficient, accurate and time-saving than traditional methods. And it is successfully adapted to the identification of anti-aging peptides, providing a new general paradigm for active peptide identification across activity types. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 (A) is an AlphaFold-3 polypeptide three-dimensional feature extraction schematic diagram; (B) is a self-encoder flowchart; (C) is a convolutional autoencoder flowchart; (D) is a model performance verification flowchart;

[0023] Figure 2 (Antioxidant and non-antioxidant peptide amino acid frequency distribution diagram)

[0024] Figure 3 (Antioxidant and non-antioxidant peptide hydrophobicity distribution diagram)

[0025] Figure 4 (Antioxidant and non-antioxidant peptide charge distribution diagram)

[0026] Figure 5 (Error diagram during self-encoder training)

[0027] Figure 6 (Training loss and validation loss diagram of the deep learning model that can accurately identify antioxidant peptides)

[0028] Figure 7 (Training loss and validation loss trend diagram of the deep learning model that can accurately identify antioxidant peptides) DETAILED DESCRIPTION

[0029] The technical solutions in the application will be described below in conjunction with the drawings and examples. Obviously, the described examples are only a part of the examples of the application, not all examples. Based on the examples in the application, all other examples obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0030] Example 1:

[0031] This embodiment details the method for constructing the active peptide dataset and the three-dimensional feature extraction process. The specific steps are as follows:

[0032] Step 1: Obtain 1458 experimentally validated antioxidant peptide sequences from the AODB database (https: / / aodb.idruglab.cn / ) as positive data, and randomly select 1458 non-antioxidant peptide sequences from the UniProt database (https: / / www.uniprot.org / ) as negative data.

[0033] Step 2: After merging the antioxidant peptides and non-antioxidant peptides from Step 1, we obtain 2916 peptide sequences. These sequences are then divided into a training set (2624 sequences) and an independent validation set (292 sequences) at a ratio of 9:1 to ensure that the two types of samples are evenly distributed in the training set.

[0034] Step 3: Optimize the geometry of each polypeptide using the AlphaFold-3 program, and extract atomic-level features from the optimized structure. Figure 1 A), including: atom number, atom type, X / Y / Z three-dimensional coordinates, amino acid type, and amino acid sequence position. A 7×N dimensional feature matrix is ​​constructed (N is the number of atoms), and zero padding is used for matrices with fewer atoms than the maximum value to achieve dimensional uniformity.

[0035] The following demonstrates the construction and training process of a deep learning model incorporating 3D features. Using Python 3.6 and PyTorch 2.0 as the programming language and framework, a deep learning model for predicting antioxidant peptides was built. The deep learning model comprises two main architectures: a conventional encoder (AE) and a convolutional encoder (CAE), both designed to extract latent characterizations of peptides and predict peptide activity. The conventional encoder focuses on reducing data dimensionality and extracting global features (…). Figure 1 B), while the convolutional encoder replaces fully connected layers with convolutional operations to capture local patterns (B). Figure 1 C). The specific steps are as follows:

[0036] Step 4: The AE architecture consists of two main elements: the encoder and the decoder. The encoder calculation formula is as follows:

[0037]

[0038] Among them The data after encoder dimensionality reduction, For input data, and These are the weight matrix and bias vector of the AE encoder, respectively. This is the activation function.

[0039] The decoder calculation formula is as follows:

[0040]

[0041] wherein is the reconstructed data, and are the weight matrix and bias vector of the AE decoder, respectively, is an activation function.

[0042] In order to evaluate the effectiveness of AE in reducing the dimension of the original data, the mean square error (MSE) verification is used to evaluate the error of and The calculation formula is as follows

[0043]

[0044] wherein N is the number of input samples.

[0045] Step 5: The reduced dimension data obtained in step 4 is taken as the input of CAE. In order to enhance the robustness of the model and avoid excessive dependence on a small number of features after dimension reduction, data expansion is needed. The calculation formula is as follows:

[0046]

[0047] wherein is the expanded data of , which represents generating a random sparse matrix with a dimension of , N refers to the number of samples, C refers to the expanded column, and K refers to the sparse filling count, which is set to 30.

[0048] The further input of CAE is composed of an encoder and a decoder. Unlike AE, CAE introduces multi-layer convolution to extract features. The calculation formula of CAE encoder is as follows:

[0049]

[0050] wherein Z is the feature set extracted by the CAE encoder; and are the weight matrix and bias vector of the CAE encoder, respectively; is a convolution operation; BatchNorm represents the normalized output of each layer to speed up the training convergence. Dropout represents randomly shielding some neurons to prevent overfitting, which is set to 0.5.

[0051] In order to avoid model overfitting, a Gaussian noise with the same size as and the noise features are extracted by convolution, and finally combined with the encoding results. The calculation formula is as follows:

[0052]

[0053] wherein is a unit matrix, denotes the feature set extracted by convolution; is a convolution operation; and are the weight matrix and bias vector during convolution extraction, respectively; denotes and the sum.

[0054] The CAE decoder restores the global structure of the data by reverse convolution operation, and the fully connected layer focuses on the key features, and finally outputs the discrimination probability of the antioxidant peptide. The calculation formula is as follows:

[0055]

[0056] wherein, denotes the prediction output generated by the CAE decoder; FC is a fully connected layer; ConvTranspose1D denotes a reverse convolution operation. and are the weight matrix and bias vector of the CAE decoder, respectively.

[0057] In order to evaluate the prediction performance of the model, the MSE verification is used to evaluate the error of and . The calculation formula is as follows:

[0058]

[0059] wherein, N is the number of input samples, is the true label of the original data.

[0060] Step 6: The deep learning model is evaluated by 5-fold cross-validation. First, the confusion matrix is constructed according to the prediction results and the true labels, and then the accuracy (ACC), true positive rate (TPR), true negative rate (TNR), precision (Pre), efficiency (EFF) and F1 value (F1) and other indicators are calculated to comprehensively evaluate the model performance (D), and the calculation formula is as follows: Figure 1

[0061]

[0062] wherein, TN, TP, FN and FP respectively denote the number of true negatives, the number of true positives, the number of false negatives and the number of false positives.

[0063] ​Step 7: Remove the labels of the data in the independent validation set, and use the trained deep learning model to make predictions, and calculate the prediction performance of the model according to the test indicators in step 6.

[0064] Embodiment 2

[0065] This embodiment details the accuracy of the constructed active peptide precise identification method fusing polypeptide three-dimensional features and deep learning in identifying another active peptide (anti-aging peptide). From the AagingBase database, 311 reported anti-aging peptides were obtained, and the same number of polypeptides were randomly obtained from the UniProt database as non-anti-aging peptides. Steps 2-7 were executed, and the ACC, TPR, TNR, Pre, EFF, and F1 indicators of the deep learning model in anti-aging peptide identification were calculated.

[0066] Example analysis

[0067] The present application proposes an active peptide precise identification method fusing polypeptide three-dimensional features and deep learning, wherein the polypeptide three-dimensional features are extracted by AlphaFold-3 calculation, and the deep learning is based on the principle of autoencoder network, relying on python software and pytorch library to establish a deep learning neural network model. In order to prove the feasibility of the method of the present application, first, taking the identification of antioxidant peptides as the target, the polypeptide three-dimensional information matrix extracted by AlphaFold-3 is used for feature extraction and model training by using the deep learning model, and a deep learning model for precise identification of antioxidant peptides is constructed. Then, the same research procedure is used for the identification of anti-aging peptides.

[0068] In order to verify the accuracy of the method of the present application in the identification of antioxidant peptides, 1458 antioxidant peptides and 1458 non-antioxidant peptides were collected, and the main features related to antioxidant properties of antioxidant peptides and non-antioxidant peptides were analyzed and compared, including amino acid characteristics, charge and hydrophobicity. Compared with non-antioxidant peptides, antioxidant peptides are rich in Ala, Cys, His, Ile, Met, Gln, Arg, Val and Thr, while Glu, Phe, Leu, Pro and Trp are reduced ( Figure 2 ). Among them, Ala, Cys, Ile and Met are determined as key residues for peptide antioxidant activity. In addition, compared with non-antioxidant peptides, the antioxidant peptide group contains more hydrophobicity ( Figure 3 ) and negative charge ( Figure 4) sequence. More hydrophobicity and negative charge are beneficial to the antioxidant properties of peptides. Atomic-level three-dimensional features of polypeptides in the dataset were extracted using AlphaFold-3, and a two-stage deep learning framework was designed for feature extraction and classification prediction of high-dimensional biological data. In the first stage, an improved AE model was used to compress the original 7xN-dimensional sparse polypeptide data to a low-dimensional latent space through nonlinear dimensionality reduction. Its unique hierarchical encoding structure effectively captured the key feature patterns in the polypeptide sequence. Experimental results show that the reconstruction error is controlled at a low level (MSE < 0.05, Figure 5 ), which verifies the effectiveness of feature extraction. In the second stage of prediction modeling, a CAE model that integrates convolutional operations and noise filtering is designed. This stage takes the features after AE dimensionality reduction as input, extracts local sequence features through multi-scale convolution, and innovatively introduces a random noise filtering module to enhance the robustness of the model. Experiments show that the accuracy (ACC), true positive rate (TPR), true negative rate (TNR), precision (Pre), efficiency (EFF), and F1 value (F1) of the deep learning model on the training set are 99.34%, 99.23%, 99.54%, 99.54%, 99.39%, and 99.39%, respectively. Then, the training loss and validation loss of the deep learning model are stabilized at 0.006 ± 0.003 and 0.015 ± 0.008, respectively, after 800 iterations. Figure 6 With the increase of iteration times, the training loss and validation loss gradually decrease and tend to be stable. Figure 7 Moreover, the trend of training loss curve and validation loss curve is similar and tends to overlap, indicating that the deep learning model has strong generalization ability and can be effectively applied to unknown data. The ACC, TPR, TNR, Pre, EFF, and F1 values of the trained deep learning model on the independent validation set are 98.29%, 97.95%, 98.63%, 98.63%, 98.29%, and 98.28%, respectively. Importantly, the steps of training the deep learning model on the antioxidant peptides are applied to the identification of anti-aging peptides, and the ACC, TPR, TNR, Pre, EFF, and F1 values obtained are 98.39%, 98.71%, 98.08%, 98.07%, 98.08%, and 98.40%, respectively.

[0069] Table 1 Performance index results of BioPepAE in different datasets

[0070] From the above experimental results, the polypeptide three-dimensional feature based on AlphaFold-3 extraction and the active peptide recognition method of the two-stage deep learning model proposed in the application show excellent performance in the antioxidant peptide prediction task. The deep learning model maintains high accuracy on the independent validation set, showing excellent discrimination ability and generalization. It is particularly noteworthy that when the model is migrated to the anti-aging peptide recognition task, it can still maintain high-precision prediction performance, fully proving that the method has cross-active peptide type universality. Through the fusion of three-dimensional structure features and deep learning technology, the limitations of traditional sequence analysis models are broken through, and a new paradigm of active peptide intelligent recognition driven by "structure-function" correlation is established, providing an efficient and reliable computing tool for high-throughput screening and rational design of polypeptides.

[0071] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand.

Claims

1. A method for precise identification of active peptides by fusing three-dimensional features of polypeptides with deep learning, characterized in that, Comprising the following steps: Step 1: Obtain several active peptides verified by experiments as positive data from an active peptide database; randomly obtain several polypeptides as non-active peptides, i.e. negative data, from a UniProt database; Step 2: Combine and randomly arrange the positive data and negative data in step 1 to form a data set, and divide the data set into a training set and an independent validation set; Step 3: Optimize the structure of all polypeptides in the training set and the independent validation set in step 2 using the AlphaFold-3 program, and obtain an output file; Step 4: Open the output file with the highest confidence in step 3 using Notepad, and extract the atomic information and sequence information of each polypeptide, and construct a matrix representing the peptide molecule based on the atomic information and sequence information of the polypeptide; the reconstructed training set is used as the input of the deep learning model; Step 5: Use the encoder in the traditional autoencoder to reduce the dimensionality of the input data in step 4, use the decoder to reconstruct, and use the mean square error to verify; Step 6: Store the data after dimensionality reduction in step 5 in the hidden layer; Step 7: Use The matrix expands the data after dimensionality reduction in step 6; Step 8: Use the data expanded in step 7 as the input of the convolutional autoencoder, and use cross-validation to train and verify the model, obtaining the trained model; Step 9: Verify the accuracy of the trained model using the independent validation set.

2. The method for accurate identification of active peptides by integrating three-dimensional features of peptides and deep learning according to claim 1, characterized in that: In step 1, the active peptides are antioxidant peptides, the non-active peptides are non-antioxidant peptides, and the antioxidant peptides are obtained from the AODB database.

3. The method for accurate identification of active peptides by integrating three-dimensional features of peptides and deep learning according to claim 1, characterized in that: The active peptides are anti-aging peptides, the non-active peptides are non-anti-aging peptides, and the anti-aging peptides are obtained from the AagingBase database.

4. The method for accurate identification of active peptides by integrating three-dimensional features of peptides and deep learning according to claim 1, characterized in that: In step 1, the ratio of the number of positive data and negative data is 4:6-6:

4. 5.The method according to claim 1 or 4, wherein the fusion polypeptide is characterized by the following three-dimensional features: (1) a hydrophobic core, (2) a hydrophilic loop, and (3) a hydrophilic tail. In step 1, the ratio of the number of positive data and negative data is 1:

1. 6.The method of claim 1, wherein the fusion polypeptide is characterized by a three-dimensional feature. In step 2, the ratio of the training set and the independent validation set is 8:2-9:

1. 7.The method of claim 1, wherein the fusion polypeptide is characterized by a three-dimensional feature. In step 4, the atomic information of the polypeptide includes atomic number, atomic type and three-dimensional coordinates; the sequence information of the polypeptide includes the type of amino acid and the amino acid sequence. 8.The method according to claim 1, wherein, In step 4, the deep learning model includes two algorithms based on autoencoder, the first one is traditional autoencoder, which realizes data dimensionality reduction with three-layer full connection structure, contains 256-dimensional hidden layer and one latent space, and the data after dimensionality reduction is stored in the hidden layer; the second one is convolutional autoencoder, which is composed of symmetric 3-layer convolutional encoder and decoder, equipped with parallel noise processing branch, adopts 5x5 convolution kernel, and uses dropout Layer to prevent overfitting, both traditional autoencoder and convolutional autoencoder are trained for 2000 rounds, with batch size of 8, and dynamic learning rate adjustment strategy is adopted. 9.The method according to claim 1, wherein, In step 7, to enhance the robustness of the model and avoid over-reliance on a small number of features after dimensionality reduction, the matrix is expanded as follows: matrix is expanded as follows: （1） where R is the data after dimension reduction by a traditional encoder; where is the extended data of R; represents generating a random sparse matrix with dimension N indicates the number of samples, C indicates the number of extended columns, and K indicates the sparse filling count. 10.The method according to claim 1, wherein, In step 8, the convolutional autoencoder replaces the full connection layer of the traditional autoencoder with a convolutional layer, specifically: （2） wherein, is the output after convolution; is the element in the mth row and nth column of the input; is the weight in the convolution kernel; b is the bias term; is the activation function; M and N are the size of the convolution kernel.

Images