A Method for Predicting the Function of Bioactive Peptides Based on Multi-View Multimodal Characterization Learning

Abstract

This invention discloses a method for predicting the function of bioactive peptides based on multi-view, multimodal representation learning. The method includes: extracting amino acid sequence information of peptides using multi-scale dilated convolutional CNN and bidirectional LSTM; extracting structural and functional features of peptide molecules using an ESM-2 model; processing molecular fingerprint information using convolutional CNN and Mamba structures; extracting topological information of the peptide molecular graph using traditional convolutional CNN, and processing node features using graph convolutional neural networks. All these multi-view features are ultimately fused into an aggregated feature representation, which is then passed through a fully connected layer and a sigmoid function is applied for multi-label classification. By concatenating and fusing the extracted features, a comprehensive peptide molecule feature representation is formed to predict various bioactive properties of the peptide. This multi-view, multimodal feature integration method not only enhances the model's predictive ability but also improves its flexibility and accuracy when processing complex biological data.

Description

Technical Field

[0001] This invention relates to the field of bioactive peptide function prediction technology, and specifically to a method for predicting bioactive peptide function based on multi-view, multimodal characterization learning. Background Technology

[0002] The importance of bioactive peptides in the medical field is increasingly prominent. Their diversity and specific biological functions make them regulators of various physiological processes, such as immune regulation, wound healing promotion, anti-inflammation, blood pressure regulation, and potential therapeutic approaches for cancer, diabetes, and antimicrobial treatments. These small molecule peptides, composed of short-chain amino acids, typically exhibit high selectivity and low side effects, making them strong candidates for developing novel therapeutic drugs. In today's context of increasing antibiotic resistance, bioactive peptides with significant therapeutic potential, such as AMPs, can effectively combat various microorganisms, including bacteria, fungi, and viruses, making them particularly important for the development of novel anti-infective drugs. ACPs, by targeting cancer cells and reducing side effects and drug resistance, have become an important research direction in cancer treatment. AIPs, by regulating immune responses and reducing the production of inflammatory mediators, demonstrate their therapeutic potential. ADPs, by regulating blood sugar through multiple mechanisms, are considered promising candidates for diabetes treatment and health food development. Finally, AHPs, as peptides obtainable through food protein breakdown or biotechnology, are considered ideal for antihypertensive drugs and functional foods due to their blood pressure-lowering efficacy. With the deepening of research on bioactive peptides, it is expected that they will show broader application potential in the treatment of a variety of diseases in the future, especially when traditional drugs are ineffective or accompanied by significant side effects.

[0003] With the rapid development of artificial intelligence technology, deep learning-based methods have begun to replace traditional machine learning techniques, especially in the field of predicting bioactive peptides. However, most existing models primarily utilize peptide amino acid sequence information, while neglecting the structural and functional information of peptide molecules. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems by providing a method for predicting the function of bioactive peptides based on multi-view, multimodal representation learning. The method demonstrates the effectiveness of the protein big language model in peptide prediction tasks through visualized clustering results, and effectively explains how the model predicts the function of bioactive peptides by capturing their complex features through visualized decision tree model decision process.

[0005] The technical solution of the present invention is as follows:

[0006] A method for predicting the function of bioactive peptides based on multi-view, multimodal representation learning is proposed. The method constructs an MVMR-BPF model, and the model feature extraction includes the following steps:

[0007] Constructing the graphical structure of polypeptide molecules;

[0008] Peptide amino acid sequence feature extraction: Multi-scale dilated convolutional CNN and bidirectional LSTM are used to extract peptide amino acid sequence information;

[0009] Peptide molecular graph feature extraction: Graph convolutional neural networks and convolutional CNNs are used to process molecular graph information from different perspectives, namely node features and topological information.

[0010] Residue feature extraction: Residue features were extracted from the sequence using the ESM-2 model;

[0011] Molecular fingerprint feature extraction: CNN is used in combination with TLU and FRN layers to capture local features, and the Mamba model is used to capture long-distance dependencies to generate feature vectors;

[0012] Fusion prediction: Combining peptide amino acid sequence, molecular fingerprint information, and peptide molecular map information, different modal features are extracted using multiple views, and the information is fused to complete the prediction of multifunctional peptides.

[0013] Furthermore, the extraction of the polypeptide amino acid sequence features specifically includes the following steps:

[0014] The polypeptide amino acid sequence vector, represented by integer encoding, is sent to the sequence feature extraction module.

[0015] The integer feature vector is transformed into a dense, continuous feature vector of dimension 100 through an embedding layer. eigenvectors After parallel dilated convolutional layers; for each dilated convolutional layer , For the first The feature vector of the layered polypeptide amino acid sequence is extracted through dilated convolution learning:

[0016] ,

[0017] in, It is the first The kernel weights of the layer It is a bias term. It is the expansion rate of this layer. It is the size of the convolution kernel. Represents the ReLU activation function for each layer. expansion rate They are 2, 4, and 8 respectively;

[0018] Output for each dilated convolutional layer Max pooling is applied to reduce sequence dimensionality while preserving features:

[0019] ,

[0020] in, This is the result of the pooling operation. It is the size of the pooling window. It is the step size;

[0021] In each convolutional layer Then, use Regularization is used to reduce overfitting, expressed as:

[0022] ,

[0023] in, It is the first The sum of squares of the weights of the convolutional kernels. It is the regularization coefficient, which controls the strength of regularization;

[0024] The vector after feature fusion is The feature vector result obtained after applying the Dropout layer is: :

[0025] ,

[0026] in, Indicates element-wise multiplication. It is a random vector whose elements are independent and follow a probability. A value of 1 indicates retention, with a probability of 1- A value of 0 indicates that the item is discarded.

[0027] Features processed by dilated convolutional layers and Dropout The input is fed into a bidirectional LSTM layer, which can capture the forward and backward context dependencies of the sequence.

[0028] ,

[0029] in, It is the output of a bidirectional LSTM layer, combining information from the beginning and end of the sequence;

[0030] Through flattening operation Convert to a one-dimensional vector It is then compressed into a 128-dimensional feature vector through a fully connected layer. :

[0031] ,

[0032] in, These are the weights of the fully connected layer. It is a bias term, and is applied simultaneously. Regularization, the ReLU function is used to introduce nonlinearity.

[0033] Furthermore, the residue feature extraction specifically includes the following steps:

[0034] For the input polypeptide amino acid sequence ,in Represents the first in the sequence The calculation process for the self-attention layer, consisting of 1 amino acid, is as follows:

[0035] ,

[0036] ,

[0037] ,

[0038] in, , , It is a weight matrix. It is the dimension of the key vector, used to scale the dot product result to avoid excessively large values;

[0039] Training the ESM-2 model on large-scale protein sequence data;

[0040] By performing average pooling on the features output from a specific layer of the ESM-2 model, a global feature vector with dimension 1280 is obtained. , As a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, it contains structural and functional information of the polypeptide molecule;

[0041] In the MVMR-BPF model, further feature extraction and dimensionality reduction are performed through two fully connected layers, while dropout layers are used to enhance the model's generalization ability, ultimately resulting in a compressed feature vector with 64 dimensions. We further learned about the structural and functional characteristics of polypeptide amino acid sequences.

[0042] Furthermore, the molecular fingerprint feature extraction specifically includes the following steps:

[0043] Multiple molecular fingerprints can be generated simultaneously from amino acid sequences using the RDKit library;

[0044] Two different fingerprint features were combined using a point-by-point summation method. and Combined into a comprehensive feature vector :

[0045] ,

[0046] Fusion molecular fingerprint feature vector Further feature extraction is performed through a deep convolutional network, where each convolutional layer... The following layers are configured with FRN layers and TLU activation layers; the FRN layers are located within the convolutional layers. Output Perform the operation above:

[0047] ,

[0048] in, and They are The mean and variance, It is a very small constant to avoid division by zero, while It is a learnable parameter;

[0049] Following the FRN layer, the TLU activation layer is enhanced. The representation of the nonlinear characteristics of a vector yields a vector. :

[0050] ,

[0051] in, It is the activation threshold, a learnable parameter used to control the non-linear threshold of the activation layer;

[0052] Extracted features We delve deeper into the Mamba model; the Mamba model comprises multiple custom MambaBlocks, and the core operations within each MambaBlock can be described by a discretized version of the following continuous system:

[0053] ,

[0054] in, , and These are the state transition matrix, input matrix, and output matrix after discretization adjustment. The specific discretization steps are as follows:

[0055] ,

[0056] in, The time scale parameter is used; MambaBlock uses this method to capture long-range dependencies in amino acid sequences.

[0057] Finally, through feature extraction and context analysis layers, a 64-dimensional feature vector was extracted. .

[0058] Furthermore, the molecular graph topology information processing in the polypeptide molecular graph feature extraction includes the following steps:

[0059] The Rdkit library was used to construct a molecular graph from the polypeptide amino acid sequence; network topology statistics were calculated from the molecular graph.

[0060] Network metrics organization The feature matrix is ​​obtained by taking the network topology characteristics of an amino acid in each row, thus yielding the topology feature vector. ;

[0061] Configure a model containing a three-layer two-dimensional convolutional neural network, starting from topological feature vectors. Extracting and utilizing network topology information;

[0062] In the model, for the first Topological feature vectors of convolutional layers Perform convolution operation to obtain :

[0063] ,

[0064] in, and They are the first Convolution kernels and bias terms of layers, Represents a convolution operation; ReLU is the activation function.

[0065] Get the first The convolutional layer returns The vector is then processed through a batch normalization layer to obtain... ,

[0066] ,

[0067] in, and These are the mean and variance of the batch data. and These are learnable scaling and offset parameters. It is a small constant to prevent division by zero;

[0068] After batch normalization, global average pooling is used to reduce the feature dimensionality, resulting in... vector:

[0069] ,

[0070] in, and These are the height and width of the feature map, respectively; and to prevent learned... Vector overfitting was addressed by adding a Dropout layer to the module with a dropout rate of 0.3.

[0071] ,

[0072] It is a random binary vector. Indicates element-wise multiplication;

[0073] Finally, network topology features are learned and extracted from molecular graph topology information. And it was used to further predict the properties of polypeptide molecules.

[0074] Furthermore, the molecular graph node feature processing in the polypeptide molecular graph feature extraction includes the following steps:

[0075] Molecular diagrams represented graphically are labeled as... ,in, It is a set of nodes, representing atoms in an amino acid sequence, and It is a set of edges representing chemical bonds between atoms. Each node contains a multi-dimensional feature vector that comprehensively reflects the atomic properties. The maximum number of nodes is set to 100. Each node is described by a feature vector containing codes for different chemical properties, ultimately resulting in a node feature matrix of the molecular graph. and adjacency matrix ;

[0076] Graph convolutional neural networks are used to extract molecular graph node features. A network architecture containing two layers of graph convolution is set up, and graph convolution operations are performed on the node feature matrix:

[0077] ,

[0078] in, It is the first The node feature matrix of the layer It is an adjacency matrix with added self-loops. and identity matrix , yes The degree matrix, whose elements , It is the first The learnable weight matrix for each layer; to prevent overfitting, Dropout is applied after each layer's graph convolution:

[0079] ,

[0080] Next, the feature vector output by the graph convolutional network is processed through a global average pooling layer. From the last floor Dimensionality reduction yields the global feature representation:

[0081] ,

[0082] in, It is the total number of nodes in the graph. It is a node In the feature vectors of the last layer, It represents the structural and functional properties of amino acid sequences and is suitable for further prediction of polypeptide molecular properties.

[0083] Furthermore, the fusion prediction specifically includes the following steps:

[0084] After extracting the corresponding feature representations from each view, the Concatenate layer is used to merge the five different input features along the feature dimension, including molecular sequence features. Molecular fingerprint characteristics Network topology information Graphical representation features Molecular structure and functional characteristics extracted from ESM-2 The feature fusion vector is obtained. ;

[0085] A Dropout layer is introduced with a dropout rate of 0.3; a Dense layer with sigmoid activation function is used to predict target biological activity, and L2 regularization is added.

[0086] Furthermore, the MVMR-BPF model is trained using the Adam optimizer, where the learning rate is controlled by the lr parameter. The training process employs binary cross-entropy as the loss function, and the training progress is monitored using the accuracy metric.

[0087] .

[0088] Furthermore, the network topology statistical indicators include the average neighbor degree, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality, and proximity centrality of nodes.

[0089] Furthermore, the feature vector description of the chemical property encoding includes atom type, connectivity, implicit valence, and whether it is an aromatic property.

[0090] Compared with existing technologies, the advantages of this invention are:

[0091] 1. A method for predicting the function of bioactive peptides based on multi-view multimodal characterization learning. The proposed method provides a new perspective and interpretable model for the efficient identification and functional prediction of bioactive peptides. It shows significant performance improvement in all aspects of the bioactive peptide identification task, especially in terms of coverage and accuracy.

[0092] 2. A method for predicting the function of bioactive peptides based on multi-view, multi-modal representation learning. This method deeply mines the node features of peptide molecules, constructs a corresponding molecular graph, and fully considers key information such as the topological structure and atomic properties of peptide molecules. By comprehensively extracting the feature representation of the molecular graph of peptide molecules from two unique perspectives, it can more comprehensively reflect the physical structural features of peptide molecules and the physicochemical features of their atomic nodes. The use of graph convolutional neural networks effectively extracts the node information in the molecular graph, enhancing the understanding and representation ability of molecular structural complexity.

[0093] 3. A method for predicting the function of bioactive peptides based on multi-view multimodal representation learning. This method utilizes the ESM-2 protein language model to perform in-depth structural and functional analysis of peptide sequences. It not only improves the accuracy of sequence analysis but also reveals the potential role of the large protein language model in biological system analysis, overcoming the limitations of traditional deep learning structure prediction for proteins.

[0094] 4. A method for predicting the function of bioactive peptides based on multi-view, multi-modal representation learning is proposed. A multi-view, multi-modal peptide representation learning architecture is designed, and different view models are designed to extract different features of peptide molecules from multiple dimensions, targeting various modalities such as chemical properties, physical properties, and bioactivity. In the molecular fingerprint feature extraction view, CNN and TLU / FRN layers are used to extract local features, and the Mamba module is used to capture long-range dependencies, ultimately generating feature vectors. Through heterogeneous feature fusion technology, the interaction of various features is comprehensively analyzed, significantly improving the prediction accuracy of peptide molecule behavior. Attached Figure Description

[0095] Figure 1 This is a general framework diagram of the MVMR-BPF model of the present invention.

[0096] Figure 2 This is a structural diagram of the polypeptide sequence feature extraction of the present invention.

[0097] Figure 3 This is a schematic diagram of the multi-molecular fingerprint fusion feature extraction structure of the present invention.

[0098] Figure 4 This is a schematic diagram of the molecular graph network topology information processing module of the present invention.

[0099] Figure 5 This is a structural diagram of the molecular graph node feature processing of the present invention. Detailed Implementation

[0100] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0102] Please see Figure 1-5 A method for predicting the function of bioactive peptides based on multi-view, multimodal representation learning is proposed, which constructs an MVMR-BPF model, such as... Figure 1 As shown, the model feature extraction includes the following steps:

[0103] Constructing the graphical structure of polypeptide molecules;

[0104] Peptide amino acid sequence feature extraction: Multi-scale dilated convolutional CNN and bidirectional LSTM are used to extract peptide amino acid sequence information;

[0105] Peptide molecular graph feature extraction: Graph convolutional neural networks and convolutional CNNs are used to process molecular graph information from different perspectives, namely node features and topological information.

[0106] Residue feature extraction: Residue features were extracted from the sequence using the ESM-2 model;

[0107] Molecular fingerprint feature extraction: CNN is used in combination with TLU and FRN layers to capture local features, and the Mamba model is used to capture long-distance dependencies to generate feature vectors;

[0108] Fusion prediction: Combining peptide amino acid sequence, molecular fingerprint information, and peptide molecular map information, different modal features are extracted using multiple views, and the information is fused to complete the prediction of multifunctional peptides.

[0109] like Figure 2 As shown, the extraction of polypeptide amino acid sequence features specifically includes the following steps:

[0110] The polypeptide amino acid sequence vector, represented by integer encoding, is sent to the sequence feature extraction module.

[0111] The integer feature vector is transformed into a dense, continuous feature vector of dimension 100 through an embedding layer. eigenvectors After parallel dilated convolutional layers; for each dilated convolutional layer , For the first The feature vector of the layered polypeptide amino acid sequence is extracted through dilated convolution learning:

[0112] ,

[0113] in, It is the first The kernel weights of the layer It is a bias term. It is the expansion rate of this layer. It is the size of the convolution kernel. Represents the ReLU activation function for each layer. expansion rate They are 2, 4, and 8 respectively;

[0114] Output for each dilated convolutional layer Max pooling is applied to reduce sequence dimensionality while preserving features:

[0115] ,

[0116] in, This is the result of the pooling operation. It is the size of the pooling window. It is the stride; parallel dilated convolutional networks allow the capture of sequence features within different ranges at different dilation layers, from local to global, enhancing the model's ability to understand sequence features.

[0117] In each convolutional layer Then, use Regularization is used to reduce overfitting, expressed as:

[0118] ,

[0119] in, It is the first The sum of squares of the weights of the convolutional kernels. It is the regularization coefficient, which controls the strength of regularization; after dilated convolution feature fusion, a Dropout layer is introduced to further improve the model's generalization ability.

[0120] The vector after feature fusion is The feature vector result obtained after applying the Dropout layer is: :

[0121] ,

[0122] in, Indicates element-wise multiplication. It is a random vector whose elements are independent and follow a probability. A value of 1 indicates retention, with a probability of 1- A value of 0 indicates that the item is discarded.

[0123] Features processed by dilated convolutional layers and Dropout The input is fed into a bidirectional LSTM layer, which can capture the forward and backward context dependencies of the sequence.

[0124] ,

[0125] in, It is the output of a bidirectional LSTM layer, combining information from the beginning and end of the sequence;

[0126] Through flattening operation Convert to a one-dimensional vector It is then compressed into a 128-dimensional feature vector through a fully connected layer. :

[0127] ,

[0128] in, These are the weights of the fully connected layer. It is a bias term, and is applied simultaneously. Regularization, specifically the ReLU function, is used to introduce nonlinearity and improve the expressive power of the model.

[0129] The residue feature extraction process includes the following steps: The protein language model ESM-2 is used, which treats protein sequences as a natural language, where each amino acid is considered a character, to perform in-depth structural and functional feature extraction on polypeptide amino acid sequences. The core of the model is an autoregressive neural network, specifically a Transformer based on an attention mechanism. The Transformer structure includes an encoder and a decoder, both relying on self-attention layers to process long sequence data.

[0130] For the input polypeptide amino acid sequence ,in Represents the first in the sequence The calculation process for the self-attention layer, consisting of 1 amino acid, is as follows:

[0131] ,

[0132] ,

[0133] ,

[0134] in, , , It is a weight matrix. is the dimension of the bond vector, used to scale the dot product result to avoid excessively large values; this self-attention mechanism enables the Transformer to capture dependencies between different positions in the sequence, thereby learning the global characteristics of the protein sequence and obtaining the polypeptide molecular structure and functional feature vector Y.

[0135] The ESM-2 model was trained on large-scale protein sequence data. The main objective was to capture the complex relationships between sequence and structure / function through learning from a large number of protein sequences. The training process employed a masking training strategy, which involved randomly masking parts of the sequence and having the model predict these masked amino acids. This forced the model to learn to capture long-range dependencies and contextual information within the sequence.

[0136] By performing average pooling on the features output from a specific layer of the ESM-2 model, a global feature vector with dimension 1280 is obtained. , As a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, it contains structural and functional information of the polypeptide molecule;

[0137] In the MVMR-BPF model, further feature extraction and dimensionality reduction are performed through two fully connected layers, while dropout layers are used to enhance the model's generalization ability, ultimately resulting in a compressed feature vector with 64 dimensions. We further learned about the structural and functional characteristics of polypeptide amino acid sequences.

[0138] like Figure 3 As shown, molecular fingerprint feature extraction specifically includes the following steps:

[0139] Multiple molecular fingerprints can be generated simultaneously from amino acid sequences using the RDKit library; ECFP generates features by considering the topological connectivity of atoms, while Daylight fingerprints focus on the overall structural pattern of the molecule.

[0140] Two different fingerprint features were combined using a point-by-point summation method. and Combined into a comprehensive feature vector :

[0141] ,

[0142] Fusion molecular fingerprint feature vector Further feature extraction is performed through a deep convolutional network, where each convolutional layer... The following layers are configured with FRN layers and TLU activation layers; the FRN layers are located within the convolutional layers. Output Perform the operation above:

[0143] ,

[0144] in, and They are The mean and variance, It is a very small constant to avoid division by zero, while It is a learnable parameter;

[0145] Following the FRN layer, the TLU activation layer is enhanced. The representation of the nonlinear characteristics of a vector yields a vector. :

[0146] ,

[0147] in, The activation threshold is a learnable parameter used to control the nonlinear threshold of the activation layer. This combination of layer structures helps the model to train stably without depending on the batch size, and improves the ability to capture nonlinear features by setting the threshold, thereby enhancing the model's performance.

[0148] To enhance the model's contextual understanding capabilities, the extracted features The process then delves into the Mamba model for in-depth processing. The Mamba model comprises multiple custom MambaBlocks, each of which effectively processes and infers contextual information about the amino acid sequence by combining 1D convolution, self-attention mechanisms, and specialized selective scanning techniques. The core operations within each MambaBlock can be described by a discretized version of the following continuous system:

[0149] ,

[0150] in, , and These are the state transition matrix, input matrix, and output matrix after discretization adjustment. The specific discretization steps are as follows:

[0151] ,

[0152] in, The time scale parameter is used; MambaBlock uses this method to capture long-range dependencies in amino acid sequences; through this structured approach, MambaBlock enables the model to not only effectively capture long-range dependencies in amino acid sequences, but also to deeply analyze the biological functions of the sequences while maintaining computational efficiency.

[0153] Finally, through feature extraction and context analysis layers, a 64-dimensional feature vector was extracted. .

[0154] like Figure 4 As shown, the molecular graph topology information processing in polypeptide molecular graph feature extraction includes the following steps:

[0155] The Rdkit library was used to construct a molecular graph from the polypeptide amino acid sequence; network topology statistics were calculated from the molecular graph.

[0156] Network metrics organization The feature matrix is ​​obtained by taking the network topology characteristics of an amino acid in each row, thus yielding the topology feature vector. ;

[0157] Configure a model containing a three-layer two-dimensional convolutional neural network, starting from topological feature vectors. Extracting and utilizing network topology information;

[0158] In the model, for the first Topological feature vectors of convolutional layers Perform convolution operation to obtain :

[0159] ,

[0160] in, and They are the first Convolution kernels and bias terms of layers, Represents a convolution operation; ReLU is the activation function.

[0161] Get the first The convolutional layer returns The vector is then processed through a batch normalization layer to obtain... ,

[0162] ,

[0163] in, and These are the mean and variance of the batch data. and These are learnable scaling and offset parameters. It is a small constant to prevent division by zero;

[0164] After batch normalization, global average pooling is used to reduce the feature dimensionality, resulting in... vector:

[0165] ,

[0166] in, and These are the height and width of the feature map, respectively; and to prevent learned... Vector overfitting was addressed by adding a Dropout layer to the module with a dropout rate of 0.3.

[0167] ,

[0168] It is a random binary vector. Indicates element-wise multiplication;

[0169] Finally, network topology features are learned and extracted from molecular graph topology information. This was then used to further predict the properties of polypeptide molecules, improving the model's accuracy in predicting different polypeptide amino acid sequences.

[0170] like Figure 5 As shown, the molecular graph node feature processing in polypeptide molecular graph feature extraction includes the following steps:

[0171] Molecular diagrams represented graphically are labeled as... ,in, It is a set of nodes, representing atoms in an amino acid sequence, and It is a set of edges representing chemical bonds between atoms. Each node contains a multi-dimensional feature vector that comprehensively reflects the atomic properties. The maximum number of nodes is set to 100; this not only conforms to the length of most polypeptide sequences but also facilitates network processing. Each node is described by a feature vector containing codes for different chemical properties, ultimately yielding a node feature matrix of the molecular graph. and adjacency matrix ;

[0172] Graph convolutional neural networks are used to extract molecular graph node features. A network architecture containing two layers of graph convolution is set up, and graph convolution operations are performed on the node feature matrix:

[0173] ,

[0174] in, It is the first The node feature matrix of the layer It is an adjacency matrix with added self-loops. and identity matrix , yes The degree matrix, whose elements , It is the first The learnable weight matrix for each layer; to prevent overfitting, Dropout is applied after each layer's graph convolution:

[0175] ,

[0176] Next, the feature vector output by the graph convolutional network is processed through a global average pooling layer. From the last floor Dimensionality reduction yields the global feature representation:

[0177] ,

[0178] in, It is the total number of nodes in the graph. It is a node In the feature vectors of the last layer, It represents the structural and functional properties of amino acid sequences and is suitable for further prediction of polypeptide molecular properties.

[0179] Fusion prediction specifically includes the following steps:

[0180] After extracting the corresponding feature representations from each view, the Concatenate layer is used to merge the five different input features along the feature dimension, including molecular sequence features. Molecular fingerprint characteristics Network topology information Graphical representation features Molecular structure and functional characteristics extracted from ESM-2 The feature fusion vector is obtained. This strategy ensures that information captured from various biomolecules can be fully utilized, enhancing the model's predictive capabilities.

[0181] A Dropout layer is introduced with a dropout rate of 0.3; a Dense layer with sigmoid activation function is used to predict target biological activity, and L2 regularization is added.

[0182] The MVMR-BPF model is trained using the Adam optimizer, where the learning rate is controlled by the lr parameter. The training process uses binary cross-entropy as the loss function, and the training progress is monitored using the accuracy metric.

[0183] .

[0184] Network topology statistics include the average neighbor degree of a node, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality, and proximity centrality.

[0185] The feature vector description of chemical property encoding includes atom type, connectivity, implicit valence, and whether it is an aromatic property.

[0186] Model Evaluation

[0187] Evaluation functions and metrics:

[0188] The proposed MVMR-BPF model also uses five-fold cross-validation to demonstrate its effectiveness on the benchmark dataset. The benchmark dataset is randomly divided into five parts; one part is used for evaluation at a time, while the remaining four parts are used as training data for the model. Furthermore, to better evaluate the model's performance in predicting the function of bioactive peptides, several common evaluation metrics for multi-label classification are employed, including: precision, coverage, accuracy, absolute correctness, and absolute error. These metrics are defined as follows:

[0189] ,

[0190] ,

[0191] ,

[0192] ,

[0193] ,

[0194] ,

[0195] Where N represents the total number of all multifunctional bioactive peptides obtained after processing the benchmark dataset, and M represents the functional types of the low-redundancy dataset. This represents the union in set theory. Represented as the intersection in set theory, This indicates that the number of elements L is counted. Let be the subset of true labels for the i-th sample. It is represented as a subset of the predicted labels for the i-th sample.

[0196] Dataset: The benchmark therapeutic peptide dataset contains 6300 bioactive peptide sequences, including five bioactive peptides with different therapeutic effects: AMP, ACP, ADP, AHP, and AIP. Among them, 198 peptide sequences possess two therapeutic properties. Let... To represent a polypeptide sequence, define a function. To describe the therapeutic effects of peptide sequences. Here, Represents polypeptide sequence It has specific therapeutic effects, and This indicates that the polypeptide sequence does not have the therapeutic effect. Each polypeptide sequence... All are encoded using a one-hot vector Let be the vector, where the dimension of the vector depends on the possible categories of the sequence.

[0197] Peptide amino acid sequence feature representation: The benchmark dataset used stores peptide amino acid sequences in text format files. In these data files, amino acid sequences are encoded using single-letter symbols, with each amino acid represented by a corresponding English letter. Considering the potential limitations of directly inputting amino acid sequences represented by English characters into the computational model, the amino acid sequences are converted into a numerical form that is easily processed by machines using integer encoding.

[0198] Specifically, 20 standard amino acids in the sequence are assigned an integer code from 1 to 20, thereby transforming the original amino acid sequence of each polypeptide into a feature vector of numerical expression. For example, amino acid Alanine (A) is encoded as 1, Cysteine ​​(C) as 2, and so on, until Tyrosine (Y) is encoded as 20.

[0199] Considering the sequence length variations of multifunctional bioactive peptides, the peptide sequences in the dataset range from 5 to 517 amino acids in length. To ensure consistency in feature vector dimensions and preserve the integrity of the original amino acid sequence information, the feature vector dimensions obtained from the peptide sequences were standardized to 517. Based on this, a zero-padding method was used to standardize the length of all peptide sequences to 517 amino acids, thereby ensuring the uniformity of data input and the effectiveness of model training.

[0200] Molecular fingerprinting: Since a single chemical fingerprint cannot capture all the key structural or property characteristics of a peptide molecule, this study employed a stacked approach combining ECFP (Extended-Connectivity Fingerprints) and Daylight molecular fingerprints to describe the physicochemical properties of peptides. By combining different types of molecular fingerprints, multiple features of the compound are captured, thereby improving the performance of the MVMR-BPF model in predicting the function of bioactive peptides. By combining these two fingerprinting methods, we can more comprehensively describe peptide molecules and improve the accuracy and robustness of the model in predicting complex biological activities.

[0201] Molecular graph topology information: Using the original amino acid sequence of the polypeptide as the dataset, and utilizing the Rdkit library, atoms are used as graph nodes and molecular bonds are used as undirected edges to convert it into a molecular graph. ,in It is the set of nodes in the diagram, representing the atoms in the molecule, and It is a set of edges, representing the chemical bonds between atoms. Then, the topological information of the molecular graph is extracted using Rdkit, including seven types of features: degree, degree centrality, betweenness centrality, eccentricity, clustering coefficient, proximity centrality, and average neighbor degree.

[0202] Performance comparison:

[0203] To verify the effectiveness of the MVMR-BPF model in predicting the function of bioactive peptides, a series of comparative experiments were designed to compare MVMR-BPF with other state-of-the-art multi-label classification models. The experiments covered bioactive peptide multi-label classifiers such as MLBP, MPMABP, MMDB, and SCN-MLTPP, as well as pre-training and general multi-label classification models EML, RAKEL, RBRL, and MLDF.

[0204] Effectiveness of Multi-View Features: Experimental data show that although models such as MLBP and MPMABP effectively extract local and global features of peptides through improved multi-scale CNNs, BiGRU, and BiLSTM, they only use amino acid sequences as input, ignoring the multi-view features of peptides, resulting in limited feature extraction capabilities. In contrast, MMDB, SCN-MLTPP, and our MVMR-BPF models not only utilize amino acid sequence information but also obtain multi-view features through amino acid sequence transformation, significantly improving performance metrics. MVMR-BPF, in particular, demonstrates superior experimental results supported by rich peptide molecular-level features.

[0205] Application of protein language models: Compared to SCN-MLTPP, although the latter incorporates seven different peptide molecular feature representations and achieves some performance improvement, its feature extraction process relies too heavily on specialized experimental tools, increasing experimental complexity. In contrast, the MVMR-BPF model extracts functional and structural features of amino acid sequences using only the RDKit library and the ESM-2 protein language model, simplifying the experimental process and demonstrating higher performance and stability in identifying the therapeutic functions of bioactive peptides.

[0206] Strategies to Enhance Feature Extraction Capabilities: Compared to models like SCN-MLTPP and Pre-Train, MVMR-BPF demonstrates superior feature extraction performance. While SCN-MLTPP obtains feature representations using specialized tools, it doesn't employ deep learning models to further learn these features, instead relying solely on capsule networks for feature fusion, which limits its performance. Conversely, while the Pre-Train model utilizes natural language processing techniques to handle peptide sequences and achieves some innovative results, it neglects the complex molecular-level information of peptides, impacting its prediction accuracy. MVMR-BPF, on the other hand, uses the ESM-2 model instead of RoBert for transfer learning and utilizes different feature extraction modules in parallel in subsequent tasks, significantly improving prediction performance.

[0207] Specifically, we employed the same five-fold cross-validation method as MLBP, MPMABP, SCN-MLTPP, and Pre-Train, and conducted a detailed performance comparison. According to the data in the table, the MVMR-BPF model demonstrated superior performance in predicting the properties of bioactive peptides, achieving the highest accuracy, coverage, precision, absolute correctness, and the lowest absolute error rate.

[0208] These results reveal that traditional single-label prediction methods—which primarily rely on machine learning and statistical methods to handle a single data dimension—generally perform poorly when dealing with peptide molecules with complex properties. Compared to traditional methods, multi-label deep learning approaches can significantly improve performance. The MVMR-BPF model, by introducing multi-view features to enhance feature extraction capabilities and utilizing a large language model for initial feature extraction, significantly improves performance in bioactive peptide prediction tasks. MVMR-BPF can comprehensively capture the features of peptide molecules, providing a more efficient and comprehensive solution, representing a breakthrough in predicting the therapeutic functions of bioactive peptides.

[0209]

[0210] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning, characterized in that, The MVMR-BPF model is constructed, and the feature extraction process includes the following steps: Constructing the graphical structure of polypeptide molecules; Peptide amino acid sequence feature extraction: Multi-scale dilated convolutional CNN and bidirectional LSTM are used to extract peptide amino acid sequence information; Peptide molecular graph feature extraction: Graph convolutional neural networks and convolutional CNNs are used to process molecular graph information from different perspectives, namely node features and topological information. Residue feature extraction: Residue features were extracted from the sequence using the ESM-2 model; Molecular fingerprint feature extraction: CNN is used in combination with TLU and FRN layers to capture local features, and the Mamba model is used to capture long-distance dependencies to generate feature vectors; Fusion prediction: Combining peptide amino acid sequence, molecular fingerprint information, and peptide molecular map information, different modal features are extracted using multiple views, and the information is fused to complete the prediction of multifunctional peptides.

2. The method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1, characterized in that, The extraction of polypeptide amino acid sequence features specifically includes the following steps: The polypeptide amino acid sequence vector, represented by integer encoding, is sent to the sequence feature extraction module. The integer feature vector is transformed into a dense, continuous feature vector of dimension 100 through an embedding layer. eigenvectors After parallel dilated convolutional layers; for each dilated convolutional layer , For the first The feature vector of the layered polypeptide amino acid sequence is extracted through dilated convolution learning: ， in, It is the first The convolutional kernel weights of the layer, It is a bias term. It is the expansion rate of this layer. It is the size of the convolution kernel. Represents the ReLU activation function for each layer. expansion rate They are 2, 4, and 8 respectively; Output for each dilated convolutional layer Max pooling is applied to reduce sequence dimensionality while preserving features: ， in, This is the result of the pooling operation. It is the size of the pooling window. It is the step size; In each convolutional layer Then, use Regularization is used to reduce overfitting, expressed as: ， in, It is the first The sum of squares of the weights of the convolutional kernels. It is the regularization coefficient, which controls the strength of regularization; The vector after feature fusion is The feature vector result obtained after applying the Dropout layer is: : ， in, Indicates element-wise multiplication. It is a random vector whose elements are independent and follow a probability. A value of 1 indicates retention, with a probability of 1- A value of 0 indicates that the item is discarded. Features processed by dilated convolutional layers and Dropout The input is fed into a bidirectional LSTM layer, which can capture the forward and backward context dependencies of the sequence. ， in, It is the output of a bidirectional LSTM layer, combining information from the beginning and end of the sequence; Through flattening operation Convert to a one-dimensional vector It is then compressed into a 128-dimensional feature vector through a fully connected layer. : ， in, These are the weights of the fully connected layer. It is a bias term, and is applied simultaneously. Regularization, the ReLU function is used to introduce nonlinearity.

3. The method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1, characterized in that, The residue feature extraction specifically includes the following steps: For the input polypeptide amino acid sequence ,in Represents the first in the sequence The calculation process for the self-attention layer, consisting of 1 amino acid, is as follows: ， ， ， in, , , It is a weight matrix. It is the dimension of the key vector, used to scale the dot product result to avoid excessively large values; Training the ESM-2 model on large-scale protein sequence data; By performing average pooling on the features output from a specific layer of the ESM-2 model, a global feature vector with dimension 1280 is obtained. , As a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, it contains structural and functional information of the polypeptide molecule; In the MVMR-BPF model, further feature extraction and dimensionality reduction are performed through two fully connected layers, while dropout layers are used to enhance the model's generalization ability, ultimately resulting in a compressed feature vector with 64 dimensions. We further learned about the structural and functional characteristics of polypeptide amino acid sequences.

4. The method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1, characterized in that, The molecular fingerprint feature extraction specifically includes the following steps: Multiple molecular fingerprints can be generated simultaneously from amino acid sequences using the RDKit library; Two different fingerprint features were combined using a point-by-point summation method. and Combined into a comprehensive feature vector : ， Fusion molecular fingerprint feature vector Further feature extraction is performed through a deep convolutional network, where each convolutional layer... The following layers are configured with FRN layers and TLU activation layers; the FRN layers are located within the convolutional layers. Output Perform the operation above: ， in, and They are The mean and variance It is a very small constant to avoid division by zero, while It is a learnable parameter; Following the FRN layer, the TLU activation layer is enhanced. The representation of the nonlinear characteristics of a vector yields a vector. : ， in, It is the activation threshold, a learnable parameter used to control the non-linear threshold of the activation layer; Extracted features We delve deeper into the Mamba model; the Mamba model comprises multiple custom MambaBlocks, and the core operations within each MambaBlock can be described by a discretized version of the following continuous system: ， in, , and These are the state transition matrix, input matrix, and output matrix after discretization adjustment. The specific discretization steps are as follows: ， in, The time scale parameter is used; MambaBlock uses this method to capture long-range dependencies in amino acid sequences. Finally, through feature extraction and context analysis layers, a 64-dimensional feature vector was extracted. .

5. The method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1, characterized in that, The molecular graph topology information processing in the polypeptide molecular graph feature extraction includes the following steps: The Rdkit library was used to construct a molecular graph from the polypeptide amino acid sequence; network topology statistics were calculated from the molecular graph. Network metrics organization The feature matrix is ​​obtained by taking the network topology characteristics of an amino acid in each row, thus yielding the topology feature vector. ; Configure a model containing a three-layer two-dimensional convolutional neural network, starting from topological feature vectors. Extracting and utilizing network topology information; In the model, for the first Topological feature vectors of convolutional layers Perform convolution operation to obtain : ， in, and They are the first Convolution kernels and bias terms of layers, Represents a convolution operation; ReLU is the activation function. Get the first The convolutional layer returns The vector is then processed through a batch normalization layer to obtain... , ， in, and These are the mean and variance of the batch data. and These are learnable scaling and offset parameters. It is a small constant to prevent division by zero; After batch normalization, global average pooling is used to reduce the feature dimensionality, resulting in... vector: ， in, and These are the height and width of the feature map, respectively; and to prevent learned... Vector overfitting was addressed by adding a Dropout layer to the module with a dropout rate of 0.

3. ， It is a random binary vector. Indicates element-wise multiplication; Finally, network topology features are learned and extracted from molecular graph topology information. And it was used to further predict the properties of polypeptide molecules.

6. A method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1 or 5, characterized in that, The molecular graph node feature processing in the polypeptide molecular graph feature extraction includes the following steps: Molecular diagrams represented graphically are labeled as... ,in, It is a set of nodes, representing atoms in an amino acid sequence, and It is a set of edges representing chemical bonds between atoms. Each node contains a multi-dimensional feature vector that comprehensively reflects the atomic properties. The maximum number of nodes is set to 100. Each node is described by a feature vector containing codes for different chemical properties, ultimately resulting in a node feature matrix of the molecular graph. and adjacency matrix ; Graph convolutional neural networks are used to extract molecular graph node features. A network architecture containing two layers of graph convolution is set up, and graph convolution operations are performed on the node feature matrix: ， in, It is the first The node feature matrix of the layer, It is an adjacency matrix with added self-loops. and identity matrix , yes The degree matrix, whose elements , It is the first The learnable weight matrix for each layer; to prevent overfitting, Dropout is applied after each layer's graph convolution: ， Next, the feature vector output by the graph convolutional network is processed through a global average pooling layer. From the last floor Dimensionality reduction yields the global feature representation: ， in, It is the total number of nodes in the graph. It is a node In the feature vectors of the last layer, It represents the structural and functional properties of amino acid sequences and is suitable for further prediction of polypeptide molecular properties.

7. The method for predicting the function of bioactive peptides based on multi-view multimodal characterization learning according to claim 1, characterized in that, The fusion prediction specifically includes the following steps: After extracting the corresponding feature representations from each view, the Concatenate layer is used to merge the five different input features along the feature dimension, including molecular sequence features. Molecular fingerprint characteristics Network topology information Graphical representation features Molecular structure and functional characteristics extracted from ESM-2 The feature fusion vector is obtained. ; A Dropout layer is introduced with a dropout rate of 0.3; a Dense layer with sigmoid activation function is used to predict target biological activity, and L2 regularization is added.

8. The method for predicting the function of bioactive peptides based on multi-view, multimodal characterization learning according to claim 1, characterized in that, The MVMR-BPF model is trained using the Adam optimizer, where the learning rate is controlled by the lr parameter. The training process uses binary cross-entropy as the loss function, and the training progress is monitored using the accuracy metric. 。 9. The method for predicting the function of bioactive peptides based on multi-view multimodal characterization learning according to claim 5, characterized in that, The network topology statistics include the average neighbor degree, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality, and proximity centrality of nodes.

10. The method for predicting the function of bioactive peptides based on multi-view multimodal characterization learning according to claim 6, characterized in that, The feature vector description of the chemical property encoding includes atom type, connectivity, implicit valence, and whether it is an aromatic property.

Images