Alpha-glucosidase inhibitory peptide sequence classification method based on deep residual network

Abstract

The invention relates to the technical field of bioinformatics and artificial intelligence crossing, and discloses an alpha-glucosidase inhibitory peptide sequence classification method based on a deep residual network, and the method comprises the steps: carrying out peptide sequence data preprocessing (multi-modal coding, balancing and standardization) to generate standardized input; constructing an improved deep residual network model containing multi-modal feature fusion, graph convolution enhancement and an attention weighting layer; training an optimization model (AdamW, a dynamic learning rate and an adaptive loss weight) to obtain optimal parameters; and performing classification prediction on unknown peptide sequences by using the optimal model. According to the method, an improved deep residual network model containing multi-modal feature fusion, topological structure perception-based graph convolution enhancement and attention weighted fusion is constructed, and dynamic learning rate scheduling and a self-adaptive loss weight strategy are combined, so that high-precision classification of alpha-glucosidase inhibitory peptide sequences is realized; and the screening efficiency and the generalization ability of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, in particular to an alpha-glucosidase inhibitor peptide sequence classification method based on a deep residual network. BACKGROUND

[0002] Alpha-glucosidase inhibitors exhibit important application value in regulating blood glucose and preventing and treating diabetes and its complications. Peptide substances, as a class of natural compounds with biological activity, have become the preferred source for developing new alpha-glucosidase inhibitors due to their low toxicity, high specificity, and ease of synthesis. However, the structural diversity and functional complexity of peptide sequences pose challenges to the screening and discovery of alpha-glucosidase inhibitor peptides. Therefore, developing an efficient and accurate prediction method is of important scientific significance and application prospect for accelerating the discovery of new alpha-glucosidase inhibitor peptides.

[0003] In the prior art, the screening of alpha-glucosidase inhibitor peptides mainly relies on traditional biochemical experimental methods, such as in vitro enzyme inhibition activity determination and cell level experiments. With the development of computational biology and machine learning technology, some sequence feature-based prediction models have also been applied to peptide function prediction, such as using long short-term memory network (LSTM) or a combination model of convolutional neural network and long short-term memory network (CNN-LSTM) to process peptide sequence data, aiming to identify the activity from the sequence level. These methods usually encode the peptide sequence into a numerical vector, and then input it into a deep learning model for training and prediction.

[0004] Although the prior art has made some progress in the discovery of alpha-glucosidase inhibitor peptides, there are still some deficiencies: the traditional biochemical screening method relies on a large number of experimental operations and reagent consumption, resulting in a long screening period and high cost, which is difficult to meet the demand of large-scale high-throughput screening. The sequence prediction model based on deep learning, such as LSTM or CNN-LSTM, is usually difficult to fully capture the complex local topological structure relationship between amino acids in the peptide sequence when processing the peptide sequence, and due to the limitation of network structure, it has limitations in extracting multi-scale features, resulting in relatively low classification accuracy and generalization ability of the model, and in most cases, the accuracy of the model is difficult to reach more than 85%. This is because the amino acid composition and arrangement of the peptide sequence is highly complex, and its activity is not only related to the properties of individual amino acids, but also closely related to the spatial proximity and interaction between amino acids and the global features of the entire sequence, and the comprehensive processing ability of traditional sequence models for these multi-dimensional information is limited. In addition, the natural length of the peptide sequence has variability, which brings difficulty in feature alignment to the deep learning model which needs fixed input length, and the simple padding or truncation strategy often loses information or introduces noise, affecting the learning effect of the model. At the same time, with the increase of network depth, the traditional deep learning model is also prone to gradient disappearance problem, which leads to the decline of training efficiency of the model in the deep feature learning process, and even cannot converge, further limiting the performance improvement. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an alpha-glucosidase inhibitor peptide sequence classification method based on a deep residual network, which solves the problems of low efficiency and high cost of existing alpha-glucosidase inhibitor peptide screening methods, and low classification accuracy caused by insufficient feature extraction and weak generalization ability of existing computing models.

[0006] To achieve the above purpose, the present application is implemented by the following technical solutions: an alpha-glucosidase inhibitor peptide sequence classification method based on a deep residual network, comprising the following steps:

[0007] S1, a peptide sequence data set preparation and preprocessing step, including loading an original peptide sequence data set, encoding the loaded peptide sequence data set with multi-modal amino acid features, balancing the encoded peptide sequence data set, and standardizing the sequence length to generate standardized feature input for model training;

[0008] S2, an improved deep residual network model construction step based on the standardized feature input, the deep residual network model comprising a multi-modal feature fusion layer for processing the standardized feature input, a topological structure perception based graph convolution feature enhancement module, a residual block layer, an attention mechanism based feature weighting fusion layer and an output layer;

[0009] S3, training and optimizing the constructed deep residual network model, the training and optimizing process including configuring AdamW optimizer, adopting dynamic learning rate scheduling strategy, applying uncertainty-aware adaptive loss weight strategy and setting deep residual network model saving mechanism to obtain optimal model parameters;

[0010] S4, applying the deep residual network model with the obtained optimal model parameters to perform classification prediction on unknown peptide sequence dataset.

[0011] Preferably, in the step of preparing and preprocessing the peptide sequence dataset, the step of loading the original peptide sequence dataset includes:

[0012] loading the peptide sequence dataset from a storage medium in a preset format, the first column of the peptide sequence dataset containing string data of the original peptide sequence dataset, and the second column being a binary classification label corresponding to the peptide sequence dataset, the label value being set to 0 or 1, 0 indicating that the peptide sequence dataset is not an α-glucosidase inhibitor peptide, and 1 indicating that the peptide sequence dataset is an α-glucosidase inhibitor peptide.

[0013] Preferably, in the step of preparing and preprocessing the peptide sequence dataset, the step of performing multi-modal amino acid feature encoding on the loaded peptide sequence dataset includes:

[0014] performing multi-dimensional numerical feature encoding on each amino acid in the loaded peptide sequence dataset;

[0015] the multi-dimensional numerical feature encoding including a combination of the following three features:

[0016] first, 20-dimensional one-hot encoding, the 20-dimensional one-hot encoding generating a 20-dimensional binary vector for each standard amino acid type, the binary vector having only one position as 1 and the remaining positions as 0, for representing the category information of the amino acid;

[0017] second, physicochemical property feature, the physicochemical property feature being obtained by quantifying the hydrophobicity index, molecular weight, isoelectric point, polarity and van der Waals volume of each amino acid;

[0018] third, structure information feature, the structure information feature being obtained by predicting the possibility value of the amino acid being in α-helix, β-sheet or random coil through bioinformatics tools;

[0019] The 20-dimensional one-hot encoding, the physicochemical property feature, and the structure information feature are spliced at the amino acid level to form a comprehensive high-dimensional feature vector, so that a peptide sequence dataset containing L amino acids is encoded into an L×N-dimensional feature matrix, where L is the sequence length of the peptide sequence dataset, and N is the total dimension of the comprehensive high-dimensional feature vector of each amino acid.

[0020] Preferably, the step of performing dataset balancing processing and sequence length standardization on the encoded peptide sequence dataset to generate standardized feature input for model training comprises:

[0021] The balancing processing is achieved through oversampling and / or undersampling strategies to reduce the model's preference for the majority class during training.

[0022] The input dimension of the balanced peptide sequence dataset is standardized, and the feature matrix of the peptide sequence dataset is standardized to a preset fixed sequence length L fixed .

[0023] If the original length L of the peptide sequence dataset exceeds the preset fixed length L fixed , the peptide sequence dataset is truncated to retain the features of the first L fixed amino acids.

[0024] If the original length L of the peptide sequence dataset is less than the preset fixed length L fixed , zero vectors or special markers are filled at the end of the peptide sequence dataset until the length of the peptide sequence dataset reaches the fixed length L fixed .

[0025] Preferably, in step S2, the step of constructing an improved deep residual network model comprises the following steps:

[0026] The multi-modal feature fusion layer is constructed, which is used for preliminary processing and feature fusion of the standardized feature input, converting from high-dimensional features to a format suitable for subsequent graph convolution processing.

[0027] After the output features of the multi-modal feature fusion layer, the topology structure perception-based graph convolution feature enhancement module is constructed, which is used to capture and enhance the local topological structure relationship between amino acids in the peptide sequence dataset.

[0028] After the output features of the topology structure perception-based graph convolution feature enhancement module, the residual block layer is constructed, which is used for deep feature learning through stacked residual units to extract higher-level abstract features.

[0029] After the residual block layer outputs the features, the attention mechanism-based feature weighting fusion layer is constructed, which is used to weight and optimize the learned feature maps, highlighting the feature regions important to the classification task.

[0030] After the attention mechanism-based feature weighting fusion layer outputs the features, the output layer is constructed, which is used to map the features processed layer by layer to the final classification prediction result.

[0031] Preferably, the step of constructing the multi-modal feature fusion layer comprises:

[0032] A one-dimensional convolution layer is used to perform convolution operation on the standardized feature input, which compresses and fuses the feature matrix by setting a specific convolution kernel size and output channel number, thereby generating a two-dimensional feature tensor with simplified dimensions, and adding a channel dimension to the two-dimensional feature tensor as the input of the subsequent topology structure perception-based graph convolution feature enhancement module.

[0033] The step of constructing the topology structure perception-based graph convolution feature enhancement module comprises:

[0034] The feature tensor output by the multi-modal feature fusion layer is regarded as a graph structure containing multiple nodes.

[0035] A topology adjacency matrix of the peptide sequence dataset is constructed, which describes the direct adjacent relationship of amino acids in the sequence.

[0036] On this basis, a self-connection adjacency matrix and a degree matrix are calculated.

[0037] Subsequently, a graph convolution network layer is applied to enhance the features, which extracts the topology structure features of the peptide sequence dataset by aggregating and transforming the adjacent node information.

[0038] Preferably, in the step of constructing the improved deep residual network model, the step of constructing the residual block layer comprises:

[0039] An improved ResNet34 architecture is used, which is composed of multiple groups of basic residual blocks based on the BasicBlock structure.

[0040] The residual block layer receives the features output by the topology structure perception-based graph convolution feature enhancement module.

[0041] Each BasicBlock internally contains a cross-layer connection structure, which allows input features to directly skip one or more convolutional layers and add them to the features processed by one or more convolutional layers, to achieve deep feature learning and alleviate the gradient vanishing problem;

[0042] In the step of constructing the improved deep residual network model, the step of constructing the feature weighting fusion layer based on the attention mechanism specifically comprises:

[0043] An self-attention module is inserted after the output feature map of the residual block layer;

[0044] The self-attention module converts the input feature map into a query, key and value matrix through linear transformation;

[0045] An attention weight matrix is calculated, which is obtained by the dot product of the query and key matrices and normalized by the Softmax function;

[0046] The weighted feature output is obtained by the product of the attention weight matrix and the value matrix, so as to highlight the important feature regions.

[0047] Preferably, the step of constructing the output layer specifically comprises:

[0048] After the output feature of the feature weighting fusion layer based on the attention mechanism, a global average pooling layer is connected, which pools the spatial dimension features into a single feature vector;

[0049] Then a fully connected layer is connected, which maps the single feature vector to the dimension of target classification;

[0050] A Softmax or Sigmoid activation function is applied to output the prediction probability of the peptide sequence data set belonging to alpha-glucosidase inhibitory peptide.

[0051] Preferably, the step of training and optimizing the constructed deep residual network model specifically comprises:

[0052] An AdamW optimizer is configured, and an initial learning rate is set for the optimizer;

[0053] A dynamic learning rate scheduling strategy is adopted, which includes a warm-up phase and a cosine annealing phase;

[0054] In the warm-up phase, the learning rate will be linearly increased from a small value to the configured initial learning rate within the preset warm-up phase epoch number at the beginning of training;

[0055] After the warm-up phase, the learning rate will gradually decrease according to the cosine function curve until the training ends.

[0056] An uncertainty-aware adaptive loss weighting strategy is applied, specifically as follows:

[0057] A loss weight w is dynamically assigned to each training sample i. i ;

[0058] The total loss of the model L total Defined as the average of the weighted losses:

[0059]

[0060] In the formula, L total is the total loss of the model; is the number of samples in the current training batch; w i L represents the loss weight for the i-th training sample; i Let be the training loss for the i-th training sample; i is the index of the sample in the training batch.

[0061] The uncertainty U i The model predicts the probability p of sample i. i The calculation yielded that the U i Using the entropy H(p) of the predicted probability i Quantification:

[0062] U i =H(p) i )=-(p i logp i +(1-p i log(1-p) i ));

[0063] In the formula, U i H(p) represents the uncertainty of the i-th training sample. i Let p be the predicted probability of the i-th training sample. i entropy; p i Let be the probability that the model predicts the i-th training sample as positive; log is the natural logarithm function.

[0064] The loss weight w i According to the aforementioned uncertainty U i The setting is as follows:

[0065] w i =1+λU i ;

[0066] In the formula, λ is a preset hyperparameter; U iUncertainty of the i-th training sample; w i Loss weight of the i-th training sample;

[0067] A deep residual network model saving mechanism is set, and the model saving mechanism specifically comprises: when the accuracy of the training set is higher than a preset accuracy threshold and the accuracy of the verification set is also higher than the preset accuracy threshold, the deep residual network model parameters obtained by the current training are saved to obtain optimal model parameters.

[0068] Preferably, the deep residual network model using the obtained optimal model parameters is applied to the classification prediction step of the unknown peptide sequence data set, which comprises:

[0069] The unknown peptide sequence to be predicted is preprocessed according to the steps of loading the original peptide sequence data set, performing multi-modal amino acid feature coding on the loaded peptide sequence, performing data set balancing processing and sequence length standardization on the coded peptide sequence data set, to generate a standardized feature input;

[0070] The standardized feature input is fed into the deep residual network model with the optimal model parameters;

[0071] The deep residual network model outputs the prediction probability of the unknown peptide sequence belonging to the alpha-glucosidase inhibitor peptide;

[0072] According to the prediction probability, it is judged whether the unknown peptide sequence is an alpha-glucosidase inhibitor peptide.

[0073] The application provides an alpha-glucosidase inhibitor peptide sequence classification method based on a deep residual network.

[0074] Has the following beneficial effects:

[0075] 1、The present application adopts 20-dimensional one-hot coding, physicochemical property features and structure information features to perform multi-modal numerical feature coding on each amino acid, and splices them to form a high-dimensional feature vector, which can comprehensively capture the category information, physical and chemical properties and potential structure information of the amino acids in the peptide sequence, and such comprehensive feature expression provides rich and multi-dimensional input for subsequent model learning, so that the model can understand the characteristics of the peptide sequence from different angles, thereby improving the accuracy of alpha-glucosidase inhibitor peptide recognition.

[0076] 2. The application constructs a topological structure perception-based graph convolution feature enhancement module, which regards the peptide sequence as a graph structure, captures and aggregates the direct adjacent relationship between amino acids by constructing a topological adjacency matrix between amino acids and applying a graph convolution network layer, so that the model can effectively extract the context information and local spatial arrangement features of amino acids in the peptide sequence, thereby improving the feature learning ability of the model in processing complex peptide sequence structures.

[0077] 3. The application adopts a dynamic learning rate scheduling strategy, including a warm-up phase and a cosine annealing phase, and configures an AdamW optimizer during training. The warm-up phase smoothly increases the learning rate from a low value, avoiding instability of the model at the beginning of training; the cosine annealing phase smoothly decreases the learning rate, which helps the model fine-tune parameters in the later training and avoid falling into local optimum. This learning rate scheduling strategy combined with the AdamW optimizer makes the model training process more stable and accelerates the convergence of model parameters, thereby reducing the training time and improving the performance of the final model.

[0078] 4. The application applies an uncertainty perception-based adaptive loss weight strategy, which quantifies the uncertainty of the model by calculating the entropy of the prediction probability of each sample, and dynamically adjusts the loss weight of each sample according to the uncertainty, assigning a larger loss weight to samples with higher uncertainty, so that the model pays more attention to difficult-to-classify samples during training, prompting the model to learn more discriminative features, thereby improving the classification accuracy and overall performance of the model when dealing with edge samples. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 The figure is a schematic diagram of the system structure of the embodiment of the application;

[0080] Figure 2 The figure is a schematic diagram of the method flow of the embodiment of the application;

[0081] Figure 3 The figure is a graph of the change of the accuracy of each Epoch training and verification of the embodiment of the application;

[0082] Figure 4 The figure is a molecular docking graph of the peptide sequence 1: RQFPFPRPPH of the embodiment of the application;

[0083] Figure 5 The figure is a molecular docking graph of the peptide sequence 2: LAIPVNKPGRF of the embodiment of the application;

[0084] Figure 6 The figure is a molecular docking graph of the peptide sequence 3: KFDSF of the embodiment of the application;

[0085] Figure 7Molecular docking diagram of peptide sequence 4 of the embodiment of the present application: GGESKPQFNIF;

[0086] Figure 8 Molecular docking diagram of peptide sequence 5 of the embodiment of the present application: VFCPQ. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0088] Please refer to the drawings Figure 1 , Figure 1 is a schematic diagram of the system structure according to an embodiment of the present application. The present application provides an α-glucosidase inhibitor peptide sequence classification method based on a deep residual network, which is implemented on a set of computing systems. The computing system includes one or more processors, such as a central processing unit (CPU), a graphics processing unit (GPU), or a tensor processing unit (TPU); a system memory (RAM); and a persistent memory (such as a hard drive, a solid state drive). The processor is used to execute instructions stored in the persistent memory to implement the following functional modules:

[0089] The peptide sequence dataset preparation and preprocessing module is used to load the original peptide sequence dataset from the storage medium in the preset format, encode the loaded peptide sequence dataset with multi-modal amino acid features, and perform dataset balancing processing and sequence length standardization on the encoded peptide sequence dataset to generate standardized feature input for model training.

[0090] The deep residual network model construction module is used to construct an improved deep residual network model based on the standardized feature input. The deep residual network model includes a multi-modal feature fusion layer for processing the standardized feature input, a graph convolution feature enhancement module based on topology structure perception, a residual block layer, a feature weighting fusion layer based on an attention mechanism, and an output layer.

[0091] The model training and optimization module is used to train and optimize the constructed deep residual network model. The training and optimization process includes configuring the AdamW optimizer, adopting a dynamic learning rate scheduling strategy, applying an uncertainty perception-based adaptive loss weight strategy, and setting a deep residual network model saving mechanism to obtain optimal model parameters.

[0092] The classification prediction module is used to apply the deep residual network model with the obtained optimal model parameters to perform classification prediction on unknown peptide sequence datasets.

[0093] Please refer to the attached Figure 2 , Figure 2 is a method flowchart according to an embodiment of the present application. The present application provides an alpha-glucosidase inhibitor peptide sequence classification method based on a deep residual network, which can include the following steps:

[0094] S1, a peptide sequence dataset preparation and preprocessing step is performed.

[0095] This step first loads the peptide sequence dataset from the storage medium in the preset format. The first column of the dataset contains the string data of the original peptide sequence, and the second column is the binary classification label corresponding to the peptide sequence. The label value is set to 0 or 1, where 0 represents a non-alpha-glucosidase inhibitor peptide, and 1 represents an alpha-glucosidase inhibitor peptide.

[0096] Next, for each loaded peptide sequence, each amino acid in its constituent amino acid sequence is subjected to multi-dimensional numerical feature encoding. Multi-dimensional numerical feature encoding includes a combination of 20-dimensional one-hot encoding, physicochemical property features (obtained by quantifying the hydrophobicity index, molecular weight, isoelectric point, polarity, and van der Waals volume of each amino acid), and structure information features (obtained by predicting the possibility value of the amino acid being in an alpha-helix, beta-sheet, or random coil using bioinformatics tools). The 20-dimensional one-hot encoding, physicochemical property features, and structure information features are concatenated at the amino acid level to form a comprehensive high-dimensional feature vector.

[0097] Therefore, a peptide sequence containing L amino acids is encoded as an LxN-dimensional feature matrix, where L is the sequence length of the peptide sequence, and N is the total dimension of the comprehensive high-dimensional feature vector of each amino acid. Subsequently, the step performs dataset balancing processing on the encoded peptide sequence dataset through oversampling and / or undersampling strategies.

[0098] Finally, the input dimension of all balanced peptide sequence datasets is subjected to sequence length standardization, and the feature matrix of the peptide sequence dataset is standardized to a preset fixed sequence length. If the original length L of the peptide sequence exceeds the preset fixed length L fixed , the peptide sequence is truncated to retain the features of the first L fixed amino acids;

[0099] If the original length L is less than the preset fixed length L fixed , a zero vector or a special marker is filled at the end of the peptide sequence until the length of the peptide sequence reaches the fixed length L fixed to generate standardized feature input for model training.

[0100] S2, an improved deep residual network model is constructed based on the standardized feature input step.

[0101] The deep residual network model comprises:

[0102] A multi-modal feature fusion layer, the fusion layer adopts a one-dimensional convolution layer to perform convolution operation on the standardized feature input, the one-dimensional convolution layer compresses and fuses the feature matrix by setting a specific convolution kernel size and an output channel number, thereby generating a two-dimensional feature tensor with a simplified dimension, and adding a channel dimension to the two-dimensional feature tensor to serve as the input of a subsequent graph convolution feature enhancement module based on topology structure perception.

[0103] The graph convolution feature enhancement module based on topology structure perception regards the feature tensor output by the multi-modal feature fusion layer as a graph structure containing multiple nodes, constructs a topology adjacency matrix of the peptide sequence, and the topology adjacency matrix describes the direct adjacent relationship of amino acids in the sequence. On this basis, a self-connection adjacency matrix and a degree matrix are calculated, wherein the self-connection adjacency matrix is obtained by adding a unit matrix to the topology adjacency matrix, and the diagonal elements of the degree matrix are the degrees of each node.

[0104] Subsequently, the module applies a graph convolution network layer to enhance the features, the graph convolution network layer extracts the topology structure features of the peptide sequence by aggregating and transforming the adjacent node information, and the operation process thereof generates the output feature matrix of the current layer by performing specific mathematical operations on the input feature matrix of the previous layer, the self-connection adjacency matrix, the degree matrix and the trainable weight matrix, and processing through a nonlinear activation function.

[0105] A residual block layer, the residual block layer adopts an improved ResNet34 architecture and is composed of multiple groups of basic residual blocks based on the BasicBlock structure, and receives the features output by the graph convolution feature enhancement module based on topology structure perception. Each BasicBlock contains a cross-layer connection structure, which allows the input features to directly skip one or more convolution layers and add the features processed by the convolution layers, so as to realize deep feature learning and relieve the gradient vanishing problem.

[0106] A feature weighting fusion layer based on attention mechanism, the fusion layer inserts a self-attention module after the output feature map of the residual block layer. The self-attention module converts the input feature map into a query, key and value matrix through linear transformation. The query matrix, the key matrix and the value matrix are obtained by multiplying the input feature map with the respective trainable weight matrix.

[0107] The attention weight matrix is calculated by the dot product of the query matrix and the transpose of the key matrix, scaled by dividing by the square root of the key vector dimension, and normalized by the Softmax function. Finally, the weighted feature output is obtained by multiplying the attention weight matrix and the value matrix, so as to highlight the important feature regions.

[0108] The output layer is connected to a global average pooling layer after the attention mechanism-based feature weighting fusion layer outputs the features. The global average pooling layer pools the spatial dimension features into a single feature vector. Then a fully connected layer is connected, which maps the single feature vector to the dimension of the target classification, and applies the Softmax or Sigmoid activation function to output the prediction probability of the peptide sequence dataset belonging to the alpha-glucosidase inhibitor peptide.

[0109] S3, training and optimizing the constructed deep residual network model.

[0110] The training and optimization process includes configuring the AdamW optimizer and setting the initial learning rate for the optimizer. A dynamic learning rate scheduling strategy is adopted, which includes a warm-up phase and a cosine annealing phase.

[0111] In the warm-up phase, the learning rate will be linearly increased from a small value to the configured initial learning rate within the pre-set warm-up phase epoch number at the beginning of training.

[0112] In the cosine annealing phase, the learning rate will gradually decrease according to the cosine function curve until the end of training. The learning rate calculation method in the cosine annealing phase is: the learning rate of the current epoch is calculated as the average of the minimum learning rate and the maximum learning rate, which is adjusted by the cosine function according to the proportion of the number of epochs completed and the total number of epochs in the cosine annealing phase.

[0113] Step 3 applies an adaptive loss weight strategy based on uncertainty perception to dynamically assign a loss weight to each training sample. The total loss of the model is defined as the average of the weighted loss. The total loss is obtained by multiplying the loss weight of each training sample in the current training batch by its training loss, summing them up, and then dividing by the number of samples in the current training batch.

[0114] Uncertainty is calculated by the prediction probability of the model for the sample, and uncertainty is quantified by the entropy of the prediction probability. The entropy of the prediction probability is calculated by multiplying the logarithm of the prediction probability and its complement by itself, adding them up, and then taking the negative value.

[0115] The loss weight is set according to the uncertainty, and the calculation formula is: the preset hyperparameter is multiplied by the uncertainty and then 1 is added to obtain the loss weight of the training sample. The step sets a deep residual network model saving mechanism, and the model saving mechanism specifically includes:

[0116] When the accuracy of the training set is higher than the preset accuracy threshold and the accuracy of the verification set is also higher than the preset accuracy threshold, the deep residual network model parameters obtained by the current training are saved to obtain the optimal model parameters.

[0117] S4, the deep residual network model with the obtained optimal model parameters is used to perform a classification prediction step on the unknown peptide sequence data set. The unknown peptide sequence to be predicted is preprocessed according to the steps of loading the original peptide sequence data set, encoding the loaded peptide sequence with multi-modal amino acid features, and performing data set balancing processing and sequence length standardization on the encoded peptide sequence data set in the S1 step, to generate standardized feature input. The standardized feature input is fed into the deep residual network model with the optimal model parameters. The deep residual network model outputs the prediction probability of the unknown peptide sequence belonging to the alpha-glucosidase inhibitory peptide. According to the prediction probability, it is judged whether the unknown peptide sequence is an alpha-glucosidase inhibitory peptide.

[0118] S1, a preparation and preprocessing step of a peptide sequence data set is performed. This step aims to convert the original peptide sequence data into standardized numerical feature input that can be recognized and processed by the deep residual network model.

[0119] This step includes: first, loading the original peptide sequence data set from the storage medium in the preset format. The storage medium includes but is not limited to a local file system, a network storage or a database. Each row of the data set contains a peptide sequence and its corresponding classification label. Specifically, the first column of the data set is the string data of the original peptide sequence, and the string is composed of standard amino acid letters; the second column of the data set is the binary classification label corresponding to the peptide sequence, and the label value is set to 0 or 1, where 0 represents that the peptide sequence is not an alpha-glucosidase inhibitory peptide, and 1 represents that the peptide sequence is an alpha-glucosidase inhibitory peptide. During the loading process, the system checks the validity of the peptide sequence data and the correctness of the label value.

[0120] Then, for each loaded peptide sequence, each amino acid in the amino acid sequence is encoded with multi-dimensional numerical features. This encoding process constructs a comprehensive feature representation for each amino acid by combining three types of features:

[0121] The first feature is 20-dimensional one-hot encoding. The 20-dimensional one-hot encoding generates a 20-dimensional binary vector for each of the 20 standard amino acid types. In the vector, only one position has a value of 1 and the other 19 positions have a value of 0, and the position with a value of 1 uniquely identifies the specific amino acid type.

[0122] The second feature is physico-chemical property feature. The physico-chemical property feature is obtained by quantifying a plurality of key physico-chemical properties of each amino acid, including hydrophobicity index (e.g., quantified according to Kyte-Doolittle or Hopp-Woods), molecular weight, isoelectric point, polarity, and van der Waals volume. The physico-chemical property values are obtained from an amino acid feature database and normalized, such as Min-Max normalization or Z-score normalization.

[0123] The third feature is structure information feature. The structure information feature is obtained by a bioinformatics tool to predict the likelihood value of each amino acid being in different secondary structures. The secondary structures include a-helix, b-sheet, or random coil. The prediction tool (e.g., PSIPRED or Jpred) outputs a probability score for each amino acid in the peptide sequence to belong to each of the three secondary structure types.

[0124] The 20-dimensional one-hot encoding, the physico-chemical property feature, and the structure information feature are concatenated at the amino acid level to form a comprehensive high-dimensional feature vector. If the one-hot encoding dimension is 20, the physico-chemical property feature dimension is 5, and the structure information feature dimension is 3, then the total dimension N of the comprehensive feature vector for each amino acid is 20 + 5 + 3 = 28.

[0125] Therefore, a peptide sequence containing L amino acids is encoded into an L x N-dimensional feature matrix, where L is the sequence length of the peptide sequence; and N is the total dimension of the comprehensive high-dimensional feature vector for each amino acid.

[0126] Subsequently, the encoded peptide sequence dataset is subjected to dataset balancing processing to address the imbalance in the number of positive and negative samples in the dataset. The balancing processing is achieved by over-sampling strategy and / or under-sampling strategy.

[0127] The over-sampling strategy duplicates or synthesizes new samples for the minority class samples, such as generating new synthetic samples by SMOTE (Synthetic Minority Over-sampling Technique) algorithm. The under-sampling strategy randomly or heuristically removes part of the majority class samples to reduce their number, such as random under-sampling. By selecting a suitable balancing strategy, the number of positive and negative samples tends to be balanced, or reaches a pre-set ratio.

[0128] Finally, the input dimension of all the peptide sequence datasets after the balance processing is standardized by sequence length. The feature matrix of the peptide sequence dataset is standardized to a preset fixed sequence length L fixed . fixed The selection is based on the statistical analysis of the dataset. If the original length L of the peptide sequence exceeds the preset fixed length L fixed , the peptide sequence is truncated, and the features of the first L fixed amino acids of the peptide sequence are retained. If the original length L of the peptide sequence is less than the preset fixed length L fixed , a zero vector or a special marker is filled at the end of the peptide sequence until the length of the peptide sequence reaches the fixed length L fixed .

[0129] The fill value is usually a zero vector to ensure that it is not confused with the actual amino acid feature value. After processing, the standardized feature input for training the deep residual network model is finally generated, with a dimension of (sample number, L fixed , N).

[0130] S2, constructing an improved deep residual network model step. Based on the standardized feature input, an improved deep residual network model is constructed, which aims to effectively extract local, global and topological features of the peptide sequence and perform deep learning. The deep residual network model is composed of the following functional modules in turn:

[0131] Multi-modal feature fusion layer. The multi-modal feature fusion layer is the initial input processing module of the deep residual network model. It receives the standardized feature input from step S1, and the dimension of the standardized feature input is (batch size, L fixed , N). Among them, the batch size represents the number of samples processed at a time, L fixed is the fixed sequence length, and N is the total dimension of the comprehensive high-dimensional feature vector of each amino acid.

[0132] The fusion layer uses a dimension convolution layer to perform convolution operation on the standardized feature input. The dimension convolution layer is configured with a specific convolution kernel size and output channel number. The convolution layer performs sliding window compression and fusion on the multi-dimensional peptide sequence feature vector in the sequence length dimension, and converts the input feature from dimension (L fixed , N) to dimension (L' fixed , C out ). Among them, L' fixed represents the sequence length after convolution operation, and C out represents the output channel number after convolution operation.

[0133] After the convolution operation, the fusion layer applies a Batch Normalization layer to stabilize the training process. Then, the fusion layer applies a nonlinear activation function, such as ReLU (Rectified Linear Unit), to increase the nonlinear representation ability of the model.

[0134] To adapt to the input format of the subsequent topology-aware graph convolution feature enhancement module, the fusion layer adds a channel dimension to the two-dimensional feature tensor, so that its dimension becomes (batch size, C out , L′ fixed )(batch size, C out , L′ fixed ) or (batch size, L′ fixed , C out ).

[0135] Topology-aware graph convolution feature enhancement module. The module is constructed after the output features of the multi-modal feature fusion layer, aiming to capture and enhance the local topological structure relationship between amino acids in the peptide sequence, i.e., the adjacency information of amino acids. The module regards the feature tensor output by the multi-modal feature fusion layer as a graph structure containing multiple nodes, where the feature vector of each amino acid is regarded as a node in the graph. The module constructs a topological adjacency matrix A of the peptide sequence.

[0136] The topological adjacency matrix A is an L′ fixed × L′ fixed matrix, and its element A ij describes the direct adjacent relationship of amino acid j and amino acid i in the sequence. Specifically, if amino acid j is the direct predecessor or successor of amino acid i, A ij = 1, otherwise A ij = 0. On this basis, the module calculates the self-connection adjacency matrix and the degree matrix The self-connection adjacency matrix is obtained by adding the identity matrix I to the topological adjacency matrix A, i.e., The role of the identity matrix I is to add a self-loop to each node. The degree matrix is a diagonal matrix, and its diagonal element is the degree of node i, i.e., the number of edges connected to node i (including the self-loop). Subsequently, the module applies a graph convolution network (GCN) layer to enhance the features, which extracts the topological structure features of the peptide sequence by aggregating and transforming the information of adjacent nodes. Its core operation process can be represented as:

[0137]

[0138] In the formula, H (l)input feature matrix of the l-th layer, whose rows represent nodes and columns represent node features; H (l+1) output feature matrix of the l+1-th layer; self-connection adjacency matrix; degree matrix; W (l) trainable weight matrix of the l-th layer for linear transformation of node features; σ denotes a nonlinear activation function, such as ReLU.

[0139] Through this graph convolution operation, the feature of each amino acid node will fuse the features of itself and its neighbor nodes, thus effectively encoding the local topological structure information of the peptide sequence. The module can contain one or more GCN layers, each of which can be followed by batch normalization and an activation function.

[0140] A residual block layer. The residual block layer is constructed after the output features of the topological structure-aware graph convolution feature enhancement module. This layer is used for deep feature learning through stacked residual units to extract higher-level abstract features while alleviating the gradient vanishing problem that may occur in deep network training. The residual block layer adopts an improved ResNet34 architecture, which is composed of multiple groups of basic residual blocks based on the BasicBlock structure. The residual block layer receives the features output by the topological structure-aware graph convolution feature enhancement module.

[0141] Each BasicBlock contains a cross-layer connection structure (SkipConnection) inside. The cross-layer connection structure allows the input feature x to directly skip one or more convolution layers and add the output feature F(x) after layer or multiple layers of processing, i.e., y = F(x) + x. This addition operation allows information to be directly transmitted in the network, which helps to maintain gradient flow when the network depth increases, thus achieving deep feature learning and effectively alleviating the gradient vanishing problem. Each residual block usually also contains convolution layers, batch normalization, and activation functions (such as ReLU) inside.

[0142] An attention mechanism-based feature weighting fusion layer. The fusion layer is constructed after the output feature map of the residual block layer. This layer is used to weight and optimize the learned feature map, highlighting the feature regions important to the classification task, i.e., the amino acids or regions in the peptide sequence that contribute more to inhibition, thus improving the model's ability to focus on key information.

[0143] The fusion layer inserts a self-attention module (Self-AttentionModule). The self-attention module converts the input feature map X into a query matrix Q, a key matrix K, and a value matrix V through linear transformation. The linear transformation can be represented as:

[0144] Q = XW Q

[0145] K = XW K ;

[0146] V = XW V

[0147] where X denotes the input feature map, W Q ,W K ,W V are trainable weight matrices, which can be implemented as a fully connected layer or a 1x1 convolutional layer. Then, the module computes an attention weight matrix A weights . The attention weight matrix is obtained by the dot product of the query Q and key K matrices, and is normalized by a Softmax function. The specific calculation formula is:

[0148]

[0149] where Q denotes the query matrix; K T denotes the transpose of the key matrix; d k denotes the dimension of the key vector, d k is used to scale the dot product result; and the Softmax function is used to normalize the result into a probability distribution.

[0150] Finally, the module obtains the weighted feature output Y weights by the product of the attention weight matrix A weighted and the value matrix V:

[0151] Y weighted = A weights V;

[0152] where Y weighted is the weighted feature output; A weights is the attention weight matrix; and V is the value matrix.

[0153] This process enables the model to calculate the importance of each feature element to all other feature elements, and accordingly to weight and aggregate the value matrix, thereby automatically identifying and strengthening the feature regions that contribute most to the classification task.

[0154] An output layer. The output layer is constructed after the feature weighting and fusion layer based on the attention mechanism outputs the features. The layer is used to map the features processed layer by layer to the final classification prediction result. The output layer first connects a global average pooling layer (Global Average Pooling, GAP).

[0155] The global average pooling layer performs an average pooling operation on the two-dimensional feature map outputted by the self-attention module along the sequence length dimension, thereby pooling the spatial dimension features into a single feature vector. Subsequently, a fully connected layer (Fully Connected Layer) is connected, which receives the single feature vector from the global average pooling layer and maps it to the dimension of the target classification. For a binary classification task, the output dimension of the fully connected layer is 1 or 2.

[0156] Finally, the output layer applies a Softmax or Sigmoid activation function to output the prediction probability of the peptide sequence dataset belonging to the alpha-glucosidase inhibitory peptide. If the output dimension of the fully connected layer is 1, a Sigmoid activation function is applied to output a value between 0 and 1, which directly represents the prediction probability of the positive class. If the output dimension of the fully connected layer is 2, a Softmax activation function is applied to output two probability values, which respectively represent the prediction probabilities of the negative class and the positive class, and the sum of the two is 1.

[0157] S3, training and optimization of the constructed deep residual network model; this step aims to iteratively train the constructed deep residual network model and adjust the model parameters through a series of optimization strategies to obtain optimal model parameters, so that the model achieves the required performance on the peptide sequence classification task.

[0158] This step first configures the AdamW optimizer. AdamW optimizer is a variant of Adam optimizer, which decouples weight decay from gradient update to improve the generalization ability of the model. The initial learning rate of the AdamW optimizer is set to, for example, 1 × 10 -3 or 5 × 10 -4 The optimizer updates the trainable parameters in the deep residual network model according to the gradient of the loss function.

[0159] Secondly, the step adopts a dynamic learning rate scheduling strategy, which adjusts the learning rate at different stages of training to improve training efficiency and model performance. The dynamic learning rate scheduling strategy includes a warm-up phase and a cosine annealing phase:

[0160] The warm-up phase is executed within the first pre-set number of warm-up phase epochs (e.g., 5 to 10 epochs) of training. In this phase, the learning rate is linearly increased from a small value (e.g., 1 / 100 of the initial learning rate) to the configured initial learning rate. This phase helps the model to slowly explore the parameter space in the early stage of training.

[0161] The cosine annealing phase starts after the warm-up phase ends. In this phase, the learning rate will gradually decrease according to the cosine function curve until the end of training. The cosine annealing strategy makes the learning rate decrease in a smooth manner, which helps the model to make fine parameter adjustments in the later stage of training. Its learning ratet The calculation formula of the tth epoch is:

[0162]

[0163] where lr t is the learning rate of the current tth epoch; lr min is the minimum value of the learning rate, which is usually set to a small number close to 0; lr max is the maximum value of the learning rate, which is usually set to the initial learning rate; T cur is the number of epochs that have been completed so far, calculated from the beginning of the cosine annealing phase; T total is the total number of epochs in the cosine annealing phase; cos is the cosine function; π is the circular constant.

[0164] Again, the step applies an adaptive loss weight strategy based on uncertainty perception. This strategy dynamically adjusts the contribution of each sample in the total loss calculation according to the prediction uncertainty of the deep residual network model, so as to encourage the model to pay more attention to difficult-to-classify samples. The total loss L total of the model is defined as the average of the weighted loss, and the calculation formula is:

[0165]

[0166] where L total is the total loss of the model; n i is the number of samples in the current training batch; w i is the loss weight of the ith training sample; L i is the training loss of the ith training sample; i is the index of the sample in the training batch.

[0167] The uncertainty U i is calculated through the prediction probability p i of the sample i by the deep residual network model. In a binary classification task, U i quantifies the entropy H(p i ) of the prediction probability, and the calculation formula is: U i = H(p i ) = -(p i logp i + (1-p i ) log(1-p i ));

[0168] where U i is the uncertainty of the ith training sample; H(p i ) is the entropy of the prediction probability p iis the probability that the model predicts the i-th training sample as positive class; log is the natural logarithm function.

[0169] loss weight w i according to the uncertainty U i is set as follows:

[0170] w i = 1 + λU i ;

[0171] where λ is a pre-set hyper-parameter; U i is the uncertainty of the i-th training sample; w i is the loss weight of the i-th training sample.

[0172] When λ > 0, samples with higher uncertainty (i.e., U i values are larger) will obtain larger loss weights w i , thus occupying a more dominant position in the total loss calculation, prompting the model to pay more attention to these difficult-to-classify or predict ambiguous samples during training.

[0173] Finally, the step sets up a deep residual network model saving mechanism. The model saving mechanism specifically includes: at the end of each training epoch, the performance indicators of the deep residual network model on the training set and the validation set are evaluated, for example, the accuracy. When the accuracy of the training set is higher than the pre-set accuracy threshold (for example, 0.90) and the accuracy of the validation set is also higher than the pre-set accuracy threshold (for example, 0.85), the deep residual network model parameters (including weights, biases, etc.) obtained by the current training are saved to the persistent memory. The saving strategy can ensure that the optimal model parameters are obtained and retained.

[0174] S4, using the deep residual network model with the optimal model parameters obtained to perform a classification prediction step on the unknown peptide sequence dataset. This step uses the deep residual network model with the optimal model parameters obtained after training and optimization to classify new, unknown peptide sequences to determine whether they are α-glucosidase inhibitory peptides.

[0175] This step first pre-processes the unknown peptide sequence to be predicted. The pre-processing strictly follows the process of loading the original peptide sequence dataset in S1, encoding the loaded peptide sequence dataset with multi-modal amino acid features, and standardizing the sequence length. Specifically, each amino acid in the unknown peptide sequence is encoded with a 20-dimensional one-hot encoding, physicochemical property feature encoding, and structure information feature encoding, and these features are spliced to form a high-dimensional feature vector.

[0176] Subsequently, the encoded unknown peptide sequence is standardized in sequence length by truncating or padding zero vectors to unify its length to a pre-set fixed length Lfixed The processing here aims to ensure that the feature input format and dimension of the unknown peptide sequence is consistent with the standardized feature input used in the training of the deep residual network model.

[0177] Next, the standardized feature input is fed into the deep residual network model with optimal model parameters. The deep residual network model has loaded the weight and bias parameters with the best performance obtained in the training and optimization phase of S3 step.

[0178] Subsequently, the deep residual network model performs forward propagation calculation and outputs the predicted probability of the unknown peptide sequence belonging to the α-glucosidase inhibitory peptide. The predicted probability is a value between 0 and 1, representing the likelihood of the model judging the peptide sequence as an α-glucosidase inhibitory peptide.

[0179] Finally, according to the predicted probability, the classification result of the unknown peptide sequence is determined. The determination is made by applying a preset classification threshold (e.g., 0.5). If the predicted probability is higher than the classification threshold, the unknown peptide sequence is determined to be an α-glucosidase inhibitory peptide; otherwise, it is determined to be a non-α-glucosidase inhibitory peptide. The final classification prediction result of the unknown peptide sequence is given.

[0180] Test example:

[0181] The present application provides a deep residual network-based α-glucosidase inhibitory peptide sequence classification method. The technical solutions of the present application and the technical effects achieved thereby are further described below through specific test examples.

[0182] Please refer to the attached Figure 3 , Figure 3 is the change graph of the training and validation accuracy of each Epoch according to an embodiment of the present application. During the training process, the training accuracy of the deep residual network model (shown by the blue curve) continuously rises from the initial level of about 72% to nearly 100% at the 15th training cycle (Epoch). At the same time, the validation accuracy of the deep residual network model (shown by the orange curve) shows an upward trend after the 9th training cycle, reaching nearly 98% at the 15th training cycle, close to the training accuracy curve.

[0183] The curve trend shows that the deep residual network model of the present application has effective learning ability for α-glucosidase inhibitory peptide sequence features and good generalization ability, and can accurately identify unknown peptide sequences. The method of the present application realizes high accuracy and stability of α-glucosidase inhibitory peptide sequence classification.

[0184] To further verify the screening ability of the deep residual network model in practical application, a series of experiments were carried out. The existing soybean protein hydrolysis peptide database in the laboratory was used as the screening source. First, the peptide sequences in the soybean protein hydrolysis peptide database were input into the deep residual network model to perform preliminary screening. The model classified and predicted the peptide sequences, and the peptide sequences determined as positive (i.e., the classification label was 1) were used as potential α-glucosidase inhibitory peptides.

[0185] Subsequently, the peptide sequences determined as positive by the model were further scored by the PeptideRanker activity prediction tool for secondary activity, and the peptide sequences with a score higher than 0.6 were screened out. Through the above screening process, five high-confidence α-glucosidase inhibitory peptides were finally obtained, and the specific sequences were as follows: peptide sequence 1 (SEQ ID NO: 1): RQFPFPRPPH; peptide sequence 2 (SEQ ID NO: 2): LAIPVNKPGRF; peptide sequence 3 (SEQ ID NO: 3): KFDSF; peptide sequence 4 (SEQ ID NO: 4): GGESKPQFNIF; and peptide sequence 5 (SEQ ID NO: 5): VFCPQ. Among them, the activity scores of peptide sequence 1 and peptide sequence 2 were both higher than 0.8.

[0186] In order to verify the inhibitory activity of the screened peptide sequences at the structural level, a molecular docking experiment was performed. The α-glucosidase (PDBID: 3WY1) was used as the receptor protein, and the ligand and water molecules in the receptor protein were removed for molecular docking. The AutoDockVina1.2.0 software was used for molecular docking.

[0187] The molecular docking results showed that the binding energy of peptide sequence 1 and α-glucosidase was -9.36 kcal / mol; the binding energy of peptide sequence 2 and α-glucosidase was -7.969 kcal / mol; the binding energy of peptide sequence 3 and α-glucosidase was -7.779 kcal / mol; the binding energy of peptide sequence 4 and α-glucosidase was -7.718 kcal / mol; and the binding energy of peptide sequence 5 and α-glucosidase was -7.545 kcal / mol.

[0188] Please refer to the attached Figure 4 to the attached Figure 8 The results showed that all the screened peptide sequences were successfully embedded in the active site of α-glucosidase and formed stable hydrogen bond networks and hydrophobic interactions. The binding energy data were all significantly lower than -7 kcal / mol, which is usually used as an evaluation standard for potent inhibitors. The molecular docking results confirmed that the peptide sequences screened by the classification model constructed in the present application had significant binding ability with α-glucosidase.

[0189] Further, the two peptide sequences with the highest scores (peptide sequence 1 and peptide sequence 2) were further screened by PeptideRanker and subjected to in vitro α-glucosidase inhibitory activity assay. The α-glucosidase inhibitory activity was determined by p-nitrophenyl-α-D-glucopyranoside (PNPG) method. The specific experimental steps are as follows:

[0190] In a 96-well plate, 20 μL of 10 mM phosphate buffer (PBS, 0.01 mol / L, pH = 7), 20 μL of 0.3 U / mL α-glucosidase, and 20 μL of sample solution were sequentially added and mixed well. In the sample blank group, an equal amount of PBS was used to replace the α-glucosidase. In the control group, an equal amount of PBS was used to replace the sample solution. In the blank group, only PBS was added. All the well plates were incubated at 37°C for 15 minutes.

[0191] Subsequently, the well plates were taken out, 20 μL of 2.5 mmol / L PNPG was added, mixed well, and incubated at 37°C for 30 minutes.

[0192] Finally, 80 μL of 0.2 mol / L sodium carbonate (Na2CO3) was added. Since PNPG is hydrolyzed by α-glucosidase to produce glucose and p-nitrophenol (PNP), PNP has an absorbance maximum at 405 nm, and the absorbance is measured by an enzyme marker. The α-glucosidase inhibition rate was calculated by the following formula:

[0193]

[0194] wherein the activity inhibition is the percentage of inhibition of the sample solution on α-glucosidase; A 样品 is the absorbance of the reaction system at 405 nm after adding the sample solution, α-glucosidase, and PNPG; A 样品空白 is the absorbance of the reaction system at 405 nm after adding the sample solution and PNPG (without α-glucosidase); A 空白对照 is the absorbance of the reaction system at 405 nm after adding PBS, α-glucosidase, and PNPG (without the sample solution); A 空白 is the absorbance of the reaction system at 405 nm after adding only PBS (without the sample solution, α-glucosidase, and PNPG);

[0195] The in vitro activity determination results showed that the inhibitory activities of peptide sequence 1 and peptide sequence 2 were about 37% and 49%, respectively. The results showed that both of the two candidate peptide sequences exhibited significant α-glucosidase inhibitory activity (both higher than 30%).

[0196] In conclusion, the improved deep residual network classification model realizes the performance of training accuracy higher than 95% and validation accuracy higher than 95% through the dynamic learning rate scheduling strategy and the dual threshold model saving mechanism combining the accuracy rates of the training set and the validation set.

[0197] The model effectively solves the feature processing difficulty caused by the variable length of the peptide sequence and the gradient vanishing problem in the deep network training. Based on the model, soybean protein hydrolysis peptides are screened, and molecular docking and in-vitro activity verification are combined, and it is confirmed that the screened peptide sequences (SEQ ID NO: 1 to SEQ ID NO: 5) all have alpha-glucosidase inhibitory activity, wherein the binding energy is lower than-7.5 kcal / mol, and the in-vitro inhibition rate of the active peptide is higher than 30%.

[0198] The application overcomes the limitations of long screening period and high cost of traditional biochemical screening, and provides an efficient, artificial intelligence driven alpha-glucosidase inhibitory peptide discovery scheme, which has application prospects in the field of hypoglycemic drug development or functional food.

[0199] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for classifying alpha-glucosidase inhibiting peptide sequences based on a deep residual network, characterized in that, The method comprises the following steps: S1, a peptide sequence data set preparation and preprocessing step is performed, which includes loading an original peptide sequence data set, multi-modal amino acid feature coding of the loaded peptide sequence data set, data set balancing processing of the coded peptide sequence data set, sequence length standardization, to generate standardized feature input for model training; S2, an improved deep residual network model construction step based on the standardized feature input, the deep residual network model comprising a multi-modal feature fusion layer for processing the standardized feature input, a graph convolution feature enhancement module based on topology structure perception, a residual block layer, a feature weighting fusion layer based on an attention mechanism, and an output layer; S3, a training and optimization step for the constructed deep residual network model, the training and optimization process comprising configuring an AdamW optimizer, adopting a dynamic learning rate scheduling strategy, applying an uncertainty perception based adaptive loss weight strategy, and setting a deep residual network model saving mechanism to obtain optimal model parameters; S4, a classification prediction step for an unknown peptide sequence data set using the deep residual network model with the obtained optimal model parameters.

2. The deep residual network-based classification method of a-glucosidase inhibitory peptide sequences according to claim 1, characterized in that, In the peptide sequence data set preparation and preprocessing step, the step of loading the original peptide sequence data set comprises: loading the peptide sequence data set from a storage medium in a predetermined format, the first column of the peptide sequence data set containing string data of the original peptide sequence data set, and the second column being a binary classification label corresponding to the peptide sequence data set, the label value being set to 0 or 1, 0 indicating that the peptide sequence data set is not an alpha-glucosidase inhibitor peptide, and 1 indicating that the peptide sequence data set is an alpha-glucosidase inhibitor peptide.

3. The deep residual network-based classification method of a-glucosidase inhibitor peptide sequence according to claim 2, characterized in that, In the peptide sequence data set preparation and preprocessing step, the step of multi-modal amino acid feature coding of the loaded peptide sequence data set comprises: performing multi-dimensional numerical feature coding on each amino acid in the constituent amino acid sequence of each loaded peptide sequence data set; The multi-dimensional numerical feature coding comprises a combination of the following three features: First, 20-dimensional one-hot coding, which generates a 20-dimensional binary vector for each standard amino acid type, with only one position being 1 and the remaining positions being 0, indicating the type information of the amino acid; Second, physicochemical property features obtained by quantifying the hydrophobicity index, molecular weight, isoelectric point, polarity, and van der Waals volume of each amino acid; Third, structure information features obtained by predicting the possibility value of the amino acid being in an alpha-helix, beta-sheet, or random coil using bioinformatics tools; The 20-dimensional one-hot coding, the physicochemical property features, and the structure information features are concatenated at the amino acid level to form a comprehensive high-dimensional feature vector, so that a peptide sequence data set containing L amino acids is coded as an LxN-dimensional feature matrix, where L is the sequence length of the peptide sequence data set, and N is the total dimension of the comprehensive high-dimensional feature vector of each amino acid.

4. The deep residual network-based classification method of a-glucosidase inhibitor peptide sequence according to claim 3, characterized in that, The step of performing dataset balancing processing and sequence length normalization on the encoded peptide sequence dataset to generate standardized feature input for model training comprises: The balancing processing is achieved through oversampling and / or undersampling strategies to reduce the model's preference for the majority class during training; normalizing input dimensions of all the peptide sequence datasets after the balance processing, normalizing a feature matrix of the peptide sequence dataset to a preset fixed sequence length L fixed ; If the original length L of the peptide sequence dataset exceeds the preset fixed length L fixed , then the peptide sequence dataset is truncated to retain the characteristics of the first L fixed amino acids. If the original length L of the peptide sequence dataset is less than the preset fixed length L fixed , then fill the zero vector or special mark at the end of the peptide sequence dataset until the length of the peptide sequence dataset reaches the fixed length L fixed .

5. The deep residual network-based classification method of α-glucosidase inhibitory peptide sequences according to claim 1, characterized in that, In step S2, the step of constructing the improved deep residual network model comprises the following steps: The multi-modal feature fusion layer is constructed for preliminary processing and feature fusion of the standardized feature input, converting high-dimensional features into a format suitable for subsequent graph convolution processing; After the multi-modal feature fusion layer outputs the features, the topology structure perception-based graph convolution feature enhancement module is constructed to capture and enhance the local topological structure relationship between amino acids in the peptide sequence dataset; After the topology structure perception-based graph convolution feature enhancement module outputs the features, the residual block layer is constructed for deep feature learning through stacked residual units to extract higher-level abstract features; After the residual block layer outputs the features, the attention mechanism-based feature weighting fusion layer is constructed to weight and optimize the learned feature maps, highlighting the feature regions important to the classification task; After the attention mechanism-based feature weighting fusion layer outputs the features, the output layer is constructed to map the features processed layer by layer to the final classification prediction result.

6. The deep residual network-based classification method of a-glucosidase inhibitor peptide sequence according to claim 5, wherein, The step of constructing the multi-modal feature fusion layer comprises: A one-dimensional convolution layer is used to perform convolution operation on the standardized feature input, which compresses and fuses the feature matrix by setting a specific convolution kernel size and output channel number, thereby generating a two-dimensional feature tensor with simplified dimensions, and adding a channel dimension to the two-dimensional feature tensor as input for the subsequent topology structure perception-based graph convolution feature enhancement module; The step of constructing the topology structure perception-based graph convolution feature enhancement module comprises: The feature tensor output by the multi-modal feature fusion layer is regarded as a graph structure containing multiple nodes; The topology adjacency matrix of the peptide sequence dataset is constructed, which describes the direct adjacent relationship of amino acids in the sequence; On this basis, the self-connection adjacency matrix and the degree matrix are calculated; Subsequently, the graph convolution network layer is applied to enhance the features, which extracts the topological structure features of the peptide sequence dataset by aggregating and transforming the adjacent node information.

7. The deep residual network-based classification method of a-glucosidase inhibitory peptide sequences according to claim 5, wherein, In the step of constructing the improved deep residual network model, the step of constructing the residual block layer specifically comprises: An improved ResNet34 architecture is used, which is composed of multiple groups of basic residual blocks based on the BasicBlock structure; The residual block layer receives the features output by the topology structure perception-based graph convolution feature enhancement module; Each BasicBlock contains a cross-layer connection structure inside, which allows input features to directly skip one or more convolutional layers and add them to the processed features of one or more convolutional layers to achieve deep feature learning and alleviate the problem of gradient vanishing; In the step of constructing the improved deep residual network model, the step of constructing the feature weighting fusion layer based on the attention mechanism specifically includes: An self-attention module is inserted after the output feature map of the residual block layer; The self-attention module converts the input feature map into query, key and value matrices through linear transformation; An attention weight matrix is calculated, which is obtained by the dot product of the query and key matrices and normalized by the Softmax function; The weighted feature output is obtained by the product of the attention weight matrix and the value matrix, so as to highlight the important feature regions.

8. The deep residual network-based classification method of a-glucosidase inhibitory peptide sequences according to claim 5, wherein, The step of constructing the output layer specifically includes: After the output feature of the feature weighting fusion layer based on the attention mechanism, a global average pooling layer is connected, which pools the spatial dimension features into a single feature vector; Then a fully connected layer is connected, which maps the single feature vector to the dimension of the target classification; The Softmax or Sigmoid activation function is applied, and the prediction probability of the peptide sequence data set belonging to the alpha-glucosidase inhibitory peptide is output.

9. The deep residual network-based classification method of a-glucosidase inhibitory peptide sequences according to claim 1, wherein, The step of training and optimizing the constructed deep residual network model specifically includes: Configure the AdamW optimizer and set the initial learning rate for the optimizer; A dynamic learning rate scheduling strategy is adopted, which includes a warm-up phase and a cosine annealing phase; In the warm-up phase, the learning rate will increase linearly from a small value to the configured initial learning rate within the pre-set warm-up phase epoch number of training; In the cosine annealing phase, after the warm-up phase ends, the learning rate will gradually decrease according to the cosine function curve until the end of training; An adaptive loss weight strategy based on uncertainty perception is applied, which specifically includes: dynamically assign a loss weight w to each training sample i i ; the total loss L of the model total defined as the average of the weighted losses: wherein L total is the total loss of the model; n is the number of samples in the current training batch; w i is the loss weight of the i-th training sample; L i is the training loss of the i-th training sample; i is the index of the sample in the training batch; the uncertainty U i the predicted probability p i for sample i by the model i is calculated as the U i quantified by the entropy H(p U i = H(p i ) = -(p i logp i +(1-p i )log(1-p i )) wherein U i is the uncertainty of the ith training sample; H(p i ) is the entropy of the predicted probability p i of the ith training sample; p i is the probability that the model predicts the ith training sample as positive; and log is the natural logarithm function. the loss weight w i according to the uncertainty U i is set, and the calculation formula is: w i = 1 + λU i ; where λ is a pre-set hyper-parameter; U i is the uncertainty of the i-th training sample; w i is the loss weight of the i-th training sample; Set up a deep residual network model saving mechanism, which specifically includes: when the accuracy of the training set is higher than the pre-set accuracy threshold and the accuracy of the validation set is also higher than the pre-set accuracy threshold, save the deep residual network model parameters obtained by the current training to obtain the optimal model parameters.

10. The deep residual network-based classification method of a-glucosidase inhibitor peptide sequence according to claim 9, characterized in that, The step of applying the deep residual network model with the optimal model parameters to classify and predict unknown peptide sequence data sets includes: Preprocess the unknown peptide sequence to be predicted according to the steps of loading the original peptide sequence data set, encoding the loaded peptide sequence with multi-modal amino acid features, balancing the data set and standardizing the sequence length of the encoded peptide sequence data set to generate standardized feature input; Feed the standardized feature input into the deep residual network model with the optimal model parameters; The deep residual network model outputs the prediction probability of the unknown peptide sequence belonging to the alpha-glucosidase inhibitory peptide; According to the predicted probability, it is judged whether the unknown peptide sequence is an α-glucosidase inhibitory peptide.

Images