Bioactive peptide function prediction method based on multi-view multi-modal representation learning

Abstract

The invention discloses a bioactive peptide function prediction method based on multi-view multi-modal representation learning. The method comprises the following steps: extracting polypeptide amino acid sequence information by using a multi-scale expansion convolution CNN and a bidirectional LSTM; extracting structure and functional characteristics of polypeptide molecules by using an ESM-2 model; the molecular fingerprint information is processed through a convolutional CNN and a Mamba structure; topological information of the polypeptide molecular graph is extracted through a traditional convolutional CNN, and node features are processed through a graph convolutional neural network. All the multi-view features are finally fused into aggregated feature representation through feature fusion, and then multi-label classification is carried out through a full connection layer and by applying a sigmoid function. The extracted features are finally subjected to feature splicing and fusion to form a comprehensive polypeptide molecule feature representation so as to predict various biological activity features of the polypeptide. The multi-view and multi-modal feature integration method not only enhances the prediction ability of the model, but also improves the flexibility and accuracy of processing complex biological data.

Description

Technical Field

[0001] The present invention relates to the technical field of bioactive peptide function prediction, and in particular to a bioactive peptide function prediction method based on multi-view multi-modal representation learning. Background Art

[0002] Bioactive peptides are becoming increasingly important in the medical field. Their diversity and specific biological functions make them regulators of various physiological processes, such as immunomodulation, wound healing, anti-inflammation, blood pressure regulation, and potential therapeutic options for anticancer, antidiabetes and antimicrobial treatments. These small molecule peptides composed of short-chain amino acids usually have high selectivity and low side effects, and are considered to be strong candidates for the development of new therapeutic drugs. In today's increasingly serious antibiotic resistance, bioactive peptides with significant therapeutic potential, such as AMPs, can effectively fight a variety of microorganisms including bacteria, fungi, and viruses, and their potential for development as new anti-infective drugs is particularly important. ACP has become an important research direction for cancer treatment by targeting cancer cells and reducing side effects and drug resistance. AIP has shown its therapeutic potential by regulating immune responses and reducing the production of inflammatory mediators. ADP regulates blood sugar through multiple mechanisms and is considered a promising candidate for diabetes treatment and health food development. Finally, AHP, as a peptide that can be obtained through food protein decomposition or biotechnology, is considered an ideal choice for antihypertensive drugs and functional foods due to its ability to lower blood pressure. As research on bioactive peptides deepens, it is expected that they will show wider application potential in the treatment of various 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 prediction of bioactive peptides. However, most existing models mainly use peptide amino acid sequence information, while ignoring the structure and function information of peptide molecules. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the function of bioactive peptides based on multi-view multimodal representation learning in response to the above-mentioned problems. The method demonstrates the effectiveness of the protein large language model in the peptide prediction task through the visualization of clustering effects, and effectively explains how the model predicts the function of bioactive peptides by capturing the complex characteristics of bioactive peptides through the decision-making process of the visualization decision tree model.

[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 MVMR-BPF model is constructed. The model feature extraction includes the following steps:

[0007] Construct the graph structure of peptide molecules;

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

[0009] Peptide molecular graph feature extraction: Graph convolutional neural network and convolutional CNN methods are used to process molecular graph information from different perspectives of node features and topological information respectively;

[0010] Residue feature extraction: Use the ESM-2 model to extract residue features from the sequence;

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

[0012] Fusion prediction: Combining the peptide amino acid sequence, molecular fingerprint information, and peptide molecular graph information, using multiple views to extract different modal features, and fusing information to complete multifunctional peptide prediction.

[0013] Furthermore, the polypeptide amino acid sequence feature extraction specifically comprises the following steps:

[0014] Send the peptide amino acid sequence vector represented by integer encoding to the sequence feature extraction module;

[0015] The integer feature vector is converted into a dense continuous feature vector x with a dimension of 100 through the embedding layer, and the feature vector x passes through parallel dilated convolutional layers; for each dilated convolutional layer l, x (l) is the feature vector of the peptide amino acid sequence in the first layer, extracted through dilated convolution learning:

[0016]

[0017] in, is the convolution kernel weight of layer l, b (l) is the bias term, r l is the dilation rate of the layer, k is the size of the convolution kernel, max(0,z) represents the ReLU activation function, and for each layer l, the dilation rate r l They are 2, 4, and 8 respectively;

[0018] For each dilated convolutional layer output x (l) , apply the maximum pooling operation to reduce the sequence dimension and retain the features:

[0019]

[0020] Among them, p (l) [i] is the result of the pooling operation, M is the size of the pooling window, and s is the step size;

[0021] After each convolutional layer l, l2 regularization is used to reduce overfitting, expressed as:

[0022] L total =L+λ∑ l ∥W (l) ∥ 2 ,

[0023] Among them, ∥W (l) ∥ 2 is the sum of the squares of the convolution kernel weights of the lth layer, and λ is the regularization coefficient, which controls the strength of regularization;

[0024] The vector after feature fusion is f, and the feature vector result after applying the Dropout layer is f′:

[0025] f′=f⊙d,

[0026] Where ⊙ represents element-wise multiplication, d is a random vector whose elements are independent and obey probability p of 1 for retention and probability 1-p of 0 for discarding;

[0027] The feature f′ processed by the dilated convolution layer and Dropout is input to the bidirectional LSTM layer. The bidirectional LSTM can capture the forward and backward context dependencies of the sequence:

[0028] h = BiLSTM(f′),

[0029] Among them, h is the output of the bidirectional LSTM layer, combining the previous and next information of the sequence;

[0030] Convert h to a one-dimensional vector h by flattening it flat , and compressed into a 128-dimensional feature vector X1 through a fully connected layer:

[0031] X1=ReLU(W fc h flat +b fc ),

[0032] Among them, W fc is the weight of the fully connected layer, b fc is the bias term, l2 regularization is applied at the same time, and the ReLU function is used to introduce nonlinearity.

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

[0034] For the input polypeptide amino acid sequence X=(x1,x2,…,x n ), where x i Represents the i-th amino acid in the sequence. The calculation process of the self-attention layer is as follows:

[0035] q i =W Q x i ,k i =W K x i ,v i =W V x i ,

[0036]

[0037] Among them, W Q ,W K ,W V is the weight matrix, d k is the dimension of the key vector, used to scale the dot product result to avoid too large values;

[0038] Train the ESM-2 model on large-scale protein sequence data;

[0039] By performing average pooling on the features output by the specific layer of the ESM-2 model, an overall feature vector Y with a dimension of 1280 is obtained. Y is used as a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, which contains the structural and functional information of the polypeptide molecule.

[0040] In the MVMR-BPF model, two fully connected layers are used to perform further feature extraction and dimensionality reduction, and a dropout layer is used to enhance the generalization ability of the model, ultimately obtaining a 64-dimensional compressed feature vector Y1; the structural and functional characteristics of the peptide amino acid sequence are further learned.

[0041] Furthermore, the molecular fingerprint feature extraction specifically includes the following steps:

[0042] Use the RDKit library to simultaneously generate multiple molecular fingerprints from amino acid sequences;

[0043] The two different feature fingerprints f1 and f2 are combined into a comprehensive feature vector f by point-by-point addition:

[0044] f=f1+f2,

[0045] The fused molecular fingerprint feature vector f is further extracted through a deep convolutional network, where each convolutional layer l is followed by an FRN layer and a TLU activation layer; the FRN layer is at the output f of the convolutional layer l. (l) To operate on:

[0046]

[0047] Among them, μ and σ 2 They are f (l)The mean and variance of , ∈ is a small constant to avoid division by zero, and τ is a learnable parameter;

[0048] After the FRN layer, the TLU activation layer enhances y (l) The nonlinear characteristics of the vector are expressed as vector z (l) :

[0049] z (l) =max(y (l) ,α),

[0050] Among them, α is the activation threshold, which is a learnable parameter used to control the nonlinear threshold of the activation layer;

[0051] The extracted features z (l) Enter the Mamba model for in-depth processing; the Mamba model includes multiple customized MambaBlocks. The core operation inside each MambaBlock can be described by the discretized version of the following continuous system:

[0052] h′(t)=A·h(t)+B·x(t), y(t)=C·h(t),

[0053] Among them, A, B and C are the state transfer matrix, input matrix and output matrix adjusted by discretization, respectively. The specific discretization steps are:

[0054] A=exp(ΔA),B=(ΔA) -1 (exp(ΔA)-I)·ΔB,

[0055] Where Δ is the time scale parameter; MambaBlock uses this method to capture long-range dependencies in amino acid sequences;

[0056] Finally, through the feature extraction and context analysis layers, a 64-dimensional feature vector M1 is extracted.

[0057] Furthermore, the molecular graph topology information processing in the polypeptide molecular graph feature extraction includes the following steps:

[0058] The peptide amino acid sequence was constructed into a molecular graph using the Rdkit library; network topology statistical indicators were calculated through the molecular graph;

[0059] The network indicators were organized into a 100 × 7 feature matrix, where each row represents the network topological property of an amino acid, and the topological feature vector N was obtained;

[0060] Set up a model consisting of a three-layer two-dimensional convolutional neural network to extract and utilize network topology information from the topological feature vector N;

[0061] The topological feature vector N of the lth convolutional layer in the model(l) , perform convolution operation to get Y (l) :

[0062] Y (l) =ReLU(W (l) *N (l) +b (l) ),

[0063] Among them, W (l) and b (l) They are the convolution kernel and bias term of the lth layer, * indicates the convolution operation, and ReLU is the activation function;

[0064] Get Y returned by the lth convolutional layer (l) vector, and then pass through the batch normalization layer to get Z (l) ,

[0065]

[0066] in, and is the mean and variance of the batch data, Y (l) and β (l) are learnable scaling and offset parameters, ∈ is a small constant to prevent division by zero;

[0067] After batch normalization, global average pooling is used to reduce the feature dimension and obtain P (l) vector:

[0068]

[0069] Among them, N H and N w are the height and width of the feature map respectively; and in order to prevent the learned P (l) The vector is overfitted, and a Dropout layer is added to the module, with the dropout rate set to 0.3:

[0070] N1 (l) =P (l) ⊙D (l) ,

[0071] D (l) is a random binary vector, ⊙ represents element-wise multiplication;

[0072] Finally, the network topology information feature N1 is learned and extracted from the molecular graph topology information, and is used for further prediction of the properties of polypeptide molecules.

[0073] Furthermore, the molecular graph node feature processing in the polypeptide molecular graph feature extraction includes the following steps:

[0074] The molecular graph is represented in the form of a graph, marked as G(V,E), where V is a set of nodes, representing atoms in the amino acid sequence, and E is a set of edges, representing chemical bonds between atoms. Each node contains a multidimensional feature vector that comprehensively reflects the properties of the atom. The upper limit of the number of nodes is set to 100. Each node is described by a feature vector containing different chemical property codes, and finally the node feature matrix H and adjacency matrix A of the molecular graph are obtained.

[0075] Use graph convolutional neural network to extract molecular graph node features, set up a network architecture containing two layers of graph convolution, and perform graph convolution operations on the node feature matrix:

[0076]

[0077] Among them, H (l) is the node feature matrix of the lth layer, is the adjacency matrix A with self-loops added and the identity matrix I, yes The degree matrix, whose elements W (l) is the learnable weight matrix of layer l; to prevent overfitting, Dropout is applied after each layer of graph convolution:

[0078] H (l+1) = Dropout(H (l+1) ,0.5),

[0079] Then, the feature vector H output by the graph convolutional network is pooled through the global average pooling layer. (L) From the last layer L, we can reduce the dimension and get the global feature representation:

[0080]

[0081] Where N is the total number of nodes in the graph, H' i (L) It is the feature vector of node i in the last layer. G1 represents the structural and functional characteristics of the amino acid sequence and is suitable for further prediction of the properties of polypeptide molecules.

[0082] Furthermore, the fusion prediction specifically includes the following steps:

[0083] After extracting the corresponding feature representations from each view, the Concatenate layer is used to merge five different input features in the feature dimension, including molecular sequence feature X1, molecular fingerprint feature M1, network topology information N1, graph representation feature G1, and molecular structure and function feature Y1 extracted from ESM-2, to obtain the feature fusion vector C;

[0084] The Dropout layer was introduced and the dropout rate was set to 0.3. The Dense layer with sigmoid activation function was used to predict the target biological activity, and L2 regularization was added.

[0085] Furthermore, the MVMR-BPF model is trained using the Adam optimizer, where the learning rate is controlled by the lr parameter. The training process of the model uses binary cross entropy as the loss function, and the training progress is monitored by the accuracy indicator:

[0086]

[0087] Furthermore, the network topology statistical indicators include the average neighbor degree, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality and closeness centrality of the nodes.

[0088] Furthermore, the characteristic vector description of the chemical property encoding includes atom type, connectivity, implicit valence, and whether it is aromatic.

[0089] Compared with the prior art, the present invention has the following beneficial effects:

[0090] 1. A method for predicting the function of bioactive peptides based on multi-view multimodal representation learning. The proposed method provides a new perspective and interpretable model for the efficient identification and function prediction of bioactive peptides. It shows significant performance improvements in all aspects of bioactive peptide identification tasks, especially in terms of coverage and accuracy.

[0091] 2. A method for predicting the function of bioactive peptides based on multi-view multimodal representation learning. It conducts in-depth mining of the node features of peptide molecules and constructs the corresponding molecular graph, fully considering key information such as the topological structure and atomic properties of peptide molecules. By comprehensively extracting the molecular graph feature representation of peptide molecules from two unique perspectives, it can more comprehensively reflect the physical structural characteristics of peptide molecules and the physicochemical characteristics of their atomic nodes. It uses graph convolutional neural networks to effectively extract node information in the molecular graph, enhancing the understanding and representation capabilities of the complexity of molecular structures.

[0092] 3. A method for predicting the function of bioactive peptides based on multi-view multimodal representation learning, using the ESM-2 protein language model to conduct 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 protein large language models in biological system analysis, surpassing the limitations of traditional protein deep learning structure prediction;

[0093] 4. A method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning. A multi-view multi-modal peptide representation learning architecture is designed. Different view models are designed to extract different features of peptide molecules from multiple dimensions based on multiple modal characteristics such as chemical properties, physical properties and biological activity. In the molecular fingerprint feature extraction view, CNN and TLU, FRU layers are used to extract local features, and the Mamba module is used to capture long-distance dependencies, and finally generate feature vectors; through the fusion technology of heterogeneous features, the interaction of various features is comprehensively analyzed, which significantly improves the prediction accuracy of peptide molecular behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 This is the overall framework diagram of the MVMR-BPF model of the present invention.

[0095] Figure 2 Extract the structural diagram of the polypeptide sequence characteristics of the present invention.

[0096] Figure 3 It is a schematic diagram of the multi-molecule fingerprint fusion feature extraction structure of the present invention.

[0097] Figure 4 This is a schematic diagram of the molecular graph network topology information processing module of the present invention.

[0098] Figure 5 This is a structural diagram of the molecular graph node feature processing of the present invention. DETAILED DESCRIPTION

[0099] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0100] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0101] See also Figure 1-5 , a method for predicting the function of bioactive peptides based on multi-view multimodal representation learning, constructing the MVMR-BPF model, such as Figure 1 As shown, model feature extraction includes the following steps:

[0102] Construct the graph structure of peptide molecules;

[0103] Peptide amino acid sequence feature extraction: Use multi-scale dilated convolutional CNN and bidirectional LSTM to extract peptide amino acid sequence information;

[0104] Peptide molecular graph feature extraction: Graph convolutional neural network and convolutional CNN methods are used to process molecular graph information from different perspectives of node features and topological information respectively;

[0105] Residue feature extraction: Use the ESM-2 model to extract residue features from the sequence;

[0106] Molecular fingerprint feature extraction: CNN is used with TLU and FRU layers to capture local features, and the Mamba model is used to capture long-distance dependencies to generate feature vectors;

[0107] Fusion prediction: Combining the peptide amino acid sequence, molecular fingerprint information, and peptide molecular graph information, using multiple views to extract different modal features, and fusing information to complete multifunctional peptide prediction.

[0108] like Figure 2 As shown, the polypeptide amino acid sequence feature extraction specifically includes the following steps:

[0109] Send the peptide amino acid sequence vector represented by integer encoding to the sequence feature extraction module;

[0110] The integer feature vector is converted into a dense continuous feature vector x with a dimension of 100 through the embedding layer, and the feature vector x passes through parallel dilated convolutional layers; for each dilated convolutional layer l, x (l) is the feature vector of the peptide amino acid sequence in the first layer, extracted through dilated convolution learning:

[0111]

[0112] in, is the convolution kernel weight of layer l, b (l) is the bias term, r l is the dilation rate of the layer, k is the size of the convolution kernel, max(0,z) represents the ReLU activation function, and for each layer l, the dilation rate r l They are 2, 4, and 8 respectively;

[0113] For each dilated convolutional layer output x (l) , apply the maximum pooling operation to reduce the sequence dimension and retain the features:

[0114]

[0115] Among them, p (l)[i] is the result of the pooling operation, M is the size of the pooling window, and s is the step size. The parallel dilated convolutional network allows the capture of sequence features in different ranges in different dilation layers, from local to global, enhancing the model's ability to understand sequence features.

[0116] After each convolutional layer l, l2 regularization is used to reduce overfitting, expressed as:

[0117] L total =L+λ∑ l ∥W (l) ∥ 2 ,

[0118] Among them, ∥W (l) ∥ 2 is the sum of the squares of the convolution kernel weights of the lth layer, and λ is the regularization coefficient, which controls the strength of regularization. After the dilated convolution feature fusion, the Dropout layer is introduced to further improve the generalization ability of the model.

[0119] The vector after feature fusion is f, and the feature vector result after applying the Dropout layer is f′:

[0120] f′=f⊙d,

[0121] Where ⊙ represents element-wise multiplication, d is a random vector whose elements are independent and obey probability p of 1 for retention and probability 1-p of 0 for discarding;

[0122] The feature f′ processed by the dilated convolution layer and Dropout is input to the bidirectional LSTM layer. The bidirectional LSTM can capture the forward and backward context dependencies of the sequence:

[0123] h = BiLSTM(f′),

[0124] Among them, h is the output of the bidirectional LSTM layer, combining the previous and next information of the sequence;

[0125] Convert h to a one-dimensional vector h by flattening it flat , and compressed into a 128-dimensional feature vector X1 through a fully connected layer:

[0126] X1=ReLU(W fc h flat +b fc ),

[0127] Among them, W fc is the weight of the fully connected layer, b fc It is a bias term. At the same time, l2 regularization is applied. The ReLU function is used to introduce nonlinearity to improve the expressiveness of the model.

[0128] Residue feature extraction specifically includes the following steps: The protein language model ESM-2 is used, which regards the protein sequence as a natural language, in which each amino acid is regarded as a character, and the in-depth structural and functional feature extraction of the polypeptide amino acid sequence is performed. The core of the model is an autoregressive neural network, namely the Transformer based on the attention mechanism. The Transformer structure includes an encoder and a decoder, both of which rely on the self-attention layer to process long sequence data.

[0129] For the input polypeptide amino acid sequence X=(x1,x2,…,x n ), where x i Represents the i-th amino acid in the sequence. The calculation process of the self-attention layer is as follows:

[0130] q i =W Q x i ,k i =W K x i ,v i =W V x i ,

[0131]

[0132] Among them, W Q ,W K ,W V is the weight matrix, d k is the dimension of the key vector, which is used to scale the dot product result to avoid excessively large values; this self-attention mechanism enables Transformer to capture the 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.

[0133] The ESM-2 model is trained on large-scale protein sequence data; the main purpose is to capture the complex relationship between sequence, structure and function by learning a large number of protein sequences. The training process adopts a masking training strategy, that is, by randomly masking some amino acids in the sequence and letting the model predict these masked amino acids, so as to force the model to learn to capture long-range dependencies and contextual information in the sequence.

[0134] By performing average pooling on the features output by the specific layer of the ESM-2 model, an overall feature vector Y with a dimension of 1280 is obtained. Y is used as a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, which contains the structural and functional information of the polypeptide molecule.

[0135] In the MVMR-BPF model, two fully connected layers are used to perform further feature extraction and dimensionality reduction, and a dropout layer is used to enhance the generalization ability of the model, ultimately obtaining a 64-dimensional compressed feature vector Y1; the structural and functional characteristics of the peptide amino acid sequence are further learned.

[0136] like Figure 3 As shown, molecular fingerprint feature extraction specifically includes the following steps:

[0137] The RDKit library is used to simultaneously generate multiple molecular fingerprints from amino acid sequences; ECFP generates features by considering the topological connectivity of atoms, while Daylight fingerprints focus on the overall structural patterns of molecules.

[0138] The two different feature fingerprints f1 and f2 are combined into a comprehensive feature vector f by point-by-point addition:

[0139] f=f1+f2,

[0140] The fused molecular fingerprint feature vector f is further extracted through a deep convolutional network, where each convolutional layer l is followed by an FRN layer and a TLU activation layer; the FRN layer is at the output f of the convolutional layer l. (l) To operate on:

[0141]

[0142] Among them, μ and ∈ 2 They are f (l) The mean and variance of , ∈ is a small constant to avoid division by zero, and τ is a learnable parameter;

[0143] After the FRN layer, the TLU activation layer enhances y (l) The nonlinear characteristics of the vector are expressed as vector z (l) :

[0144] z (l) =max(y (l) ,α),

[0145] Among them, α is the activation threshold, which is a learnable parameter used to control the nonlinear threshold of the activation layer; such a combination of layer structures helps the model to perform stable training without relying on the batch size, and improves the ability to capture nonlinear features by setting the threshold, thereby enhancing the performance of the model.

[0146] In order to enhance the context understanding ability of the model, the extracted features z (l)Enter the Mamba model for in-depth processing; the Mamba model includes multiple customized MambaBlocks, each of which effectively processes and infers the contextual information of the amino acid sequence by combining 1D convolution, self-attention mechanism and special selective scanning technology. The core operation inside each MambaBlock can be described by the discretized version of the following continuous system:

[0147] h′(t)=A·h(t)+B·x(t), y(t)=C·h(t),

[0148] Among them, A, B and C are the state transfer matrix, input matrix and output matrix adjusted by discretization, respectively. The specific discretization steps are:

[0149] A=exp(ΔA),B=(ΔA) -1 (exp(ΔA)-I)·ΔB,

[0150] Among them, Δ is the time scale parameter; MambaBlock uses this method to capture the long-range dependencies in amino acid sequences; MambaBlock uses this structured method to enable the model to not only effectively capture the long-range dependencies in amino acid sequences, but also deeply analyze the biological functions of the sequences while maintaining computational efficiency.

[0151] Finally, through the feature extraction and context analysis layers, a 64-dimensional feature vector M1 is extracted.

[0152] like Figure 4 As shown, the molecular graph topology information processing in the peptide molecular graph feature extraction includes the following steps:

[0153] The peptide amino acid sequence was constructed into a molecular graph using the Rdkit library; network topology statistical indicators were calculated through the molecular graph;

[0154] The network indicators were organized into a 100 × 7 feature matrix, where each row represents the network topological property of an amino acid, and the topological feature vector N was obtained;

[0155] Set up a model consisting of a three-layer two-dimensional convolutional neural network to extract and utilize network topology information from the topological feature vector N;

[0156] The topological feature vector N of the lth convolutional layer in the model (l) , perform convolution operation to get Y (l) :

[0157]

[0158] Among them, W (l) and b (l)They are the convolution kernel and bias term of the lth layer, * indicates the convolution operation, and ReLU is the activation function;

[0159] Get Y returned by the lth convolutional layer (l) vector, and then pass through the batch normalization layer to get Z (l) ,

[0160]

[0161] in, and is the mean and variance of the batch data, Y (l) and β (l) are learnable scaling and offset parameters, ∈ is a small constant to prevent division by zero;

[0162] After batch normalization, global average pooling is used to reduce the feature dimension and obtain P (l) vector:

[0163]

[0164] Among them, N H and N W are the height and width of the feature map respectively; and in order to prevent the learned P (l) The vector is overfitted, and a Dropout layer is added to the module, with the dropout rate set to 0.3:

[0165] N1 (l) =P (l) ⊙D (l) ,

[0166] D (l) is a random binary vector, ⊙ represents element-wise multiplication;

[0167] Finally, the network topology information feature N1 is extracted from the molecular graph topology information and used for further prediction of the properties of polypeptide molecules, thus improving the prediction accuracy of the model for different polypeptide amino acid sequences.

[0168] like Figure 5 As shown, the molecular graph node feature processing in the polypeptide molecular graph feature extraction includes the following steps:

[0169] The molecular graph is represented in the form of a graph, marked as G(V,E), where V is a set of nodes, representing the atoms in the amino acid sequence, and E is a set of edges, representing the chemical bonds between atoms. Each node contains a multidimensional feature vector that comprehensively reflects the properties of the atoms. The upper limit of the number of nodes is set to 100; this not only meets the length of most polypeptide sequences, but also facilitates network processing. Each node is described by a feature vector containing different chemical property codes, and finally the node feature matrix H and adjacency matrix A of the molecular graph are obtained;

[0170] Use graph convolutional neural network to extract molecular graph node features, set up a network architecture containing two layers of graph convolution, and perform graph convolution operations on the node feature matrix:

[0171]

[0172] Among them, H (l) is the node feature matrix of the lth layer, is the adjacency matrix A with self-loops added and the identity matrix I, yes The degree matrix, whose elements W (l) is the learnable weight matrix of layer l; to prevent overfitting, Dropout is applied after each layer of graph convolution:

[0173] H (l+1) = Dropout(H (l+1) ,0.5),

[0174] Then, the feature vector H output by the graph convolutional network is pooled through the global average pooling layer. (L) From the last layer L, we can reduce the dimension and get the global feature representation:

[0175]

[0176] Where N is the total number of nodes in the graph, H' i (L) It is the feature vector of node i in the last layer. G1 represents the structural and functional characteristics of the amino acid sequence and is suitable for further prediction of the properties of polypeptide molecules.

[0177] Fusion prediction specifically includes the following steps:

[0178] After the corresponding feature representations are extracted from each view, the Concatenate layer is used to merge five different input features in the feature dimension, including molecular sequence feature X1, molecular fingerprint feature M1, network topology information N1, graph representation feature G1, and molecular structure and function feature Y1 extracted from ESM-2, to obtain the feature fusion vector C; this strategy ensures that the information captured from various biological molecular angles can be fully utilized, thereby enhancing the predictive ability of the model.

[0179] The Dropout layer was introduced and the dropout rate was set to 0.3. The Dense layer with sigmoid activation function was used to predict the target biological activity, and L2 regularization was added.

[0180] The MVMR-BPF model is trained using the Adam optimizer, where the learning rate is controlled by the lr parameter. The model training process uses binary cross entropy as the loss function, and the accuracy indicator is used to monitor the training progress:

[0181]

[0182] Network topology statistical indicators include the average neighbor degree, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality, and closeness centrality of nodes.

[0183] The feature vector descriptions of chemical properties include atom type, connectivity, implicit valence, and whether it is aromatic.

[0184] Model Evaluation

[0185] Evaluation functions and indicators:

[0186] The proposed MVMR-BPF model also uses a five-fold cross-validation method to prove the effectiveness of the model on the benchmark dataset. The data in the benchmark dataset is randomly divided into five parts, one of which is used for evaluation each time, and the remaining four are used as training data for the model. Furthermore, in order to better evaluate the performance of the model in predicting the function of bioactive peptides, several common evaluation indicators of multi-label classification are used, including: precision, coverage, accuracy, absolute correctness and absolute error. The definitions of these evaluation indicators are as follows:

[0187]

[0188] Among them, N represents the total number of all multifunctional bioactive peptides obtained after the benchmark data set is processed, M represents the functional type of the low-redundancy data set, ∪ represents the union in set theory, ∩ represents the intersection in set theory, ‖‖ represents the statistics of the number of elements L, L i Represented as the true label subset of the i-th sample, Denotes the subset of labels predicted for the i-th sample.

[0189] Dataset: The Benchmark Therapeutic Peptide Dataset collects a total of 6,300 bioactive peptide sequences, including five bioactive peptides with different therapeutic effects, namely AMP, ACP, ADP, AHP and AIP. Among them, 198 peptide sequences have two therapeutic properties. Let S represent a peptide sequence, and define a function f:S→{0,1} to describe the therapeutic effect of the peptide sequence. Here, f(S)=1 means that the peptide sequence S has a specific therapeutic effect, while f(S)=0 means that the peptide sequence does not have this therapeutic effect. Each peptide sequence S is represented by a one-hot encoded vector X, where the dimension of the vector depends on the possible category of the sequence.

[0190] Peptide amino acid sequence feature representation: The benchmark data set used is to store the peptide amino acid sequence in a text format file. In these data files, the amino acid sequence is encoded by a single letter symbol, and each amino acid is represented by a corresponding English letter. Considering the possible limitations of directly inputting the amino acid sequence represented by English characters into the computational model, the amino acid sequence is converted into a numerical form that is easy for machine processing by integer encoding.

[0191] Specifically, the 20 standard amino acids in the sequence are assigned an integer code from 1 to 20, thereby converting the original amino acid sequence of each polypeptide into a digitally expressed feature vector. For example, the amino acid Alanine (A) is coded as 1, Cysteine ​​(C) is 2, and so on, until Tyrosine (Y) is coded as 20.

[0192] Considering the differences in sequence lengths of multifunctional bioactive peptides, the lengths of peptide sequences in the dataset range from 5 to 517 amino acids. To ensure the consistency of feature vector dimensions and retain the integrity of the original amino acid sequence information, the feature vector dimensions obtained from the peptide sequences were unified to 517. Based on this, the zero-padding method was used to standardize the lengths of all peptide sequences to 517 amino acids, thereby ensuring the uniformity of data input and the effectiveness of model training.

[0193] Molecular fingerprint representation: Since a single chemical fingerprint cannot capture all the key structures or properties of a peptide molecule, this study used the ECFP (Extended-Connectivity Fingerprints) and Daylight molecular fingerprint stacking methods to describe the physicochemical properties of peptides. The combination of different types of molecular fingerprints was used to capture multiple features of the compound, thereby improving the performance of the MVMR-BPF model in predicting the function of bioactive peptides. By combining these two fingerprints, we can more comprehensively describe peptide molecules and improve the accuracy and robustness of the model in predicting complex biological activities.

[0194] Molecular graph topology information: The original amino acid sequence of the peptide is used as the data set, and the Rdkit library is used to convert atoms as graph nodes and molecular bonds as undirected edges of the graph into a molecular graph G(V,E), where V is the set of nodes in the graph, representing atoms in the molecule, and E is the set of edges, representing chemical bonds between atoms. Then, Rdkit is used to extract the topological information of the molecular graph, including seven types of features: node degree, degree centrality, betweenness centrality, eccentricity, clustering coefficient, closeness centrality, and average neighbor degree.

[0195] Performance comparison:

[0196] In order to verify the effectiveness of the MVMR-BPF model in the task of predicting the function of bioactive peptides, a series of comparative experiments were designed to compare MVMR-BPF with other advanced multi-label classification models. The experiments covered bioactive peptide multi-label classifiers such as MLBP, MPMABP, MMDB, SCN-MLTPP, as well as Pre-Train and general multi-label classification models EML, RAKEL, RBRL and MLDF.

[0197] 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 CNN, BiGRU and BiLSTM, they only take amino acid sequences as input and ignore the multi-view features of peptides, resulting in limited feature extraction capabilities. In contrast, MMDB, SCN-MLTPP and our MVMR-BPF models not only use amino acid sequence information, but also obtain multi-view features through amino acid sequence conversion, which significantly improves performance indicators. In particular, MVMR-BPF, supported by rich peptide molecular-level features, shows superior experimental results.

[0198] Application of protein language model: Compared with SCN-MLTPP, although the latter integrates 7 different peptide molecular feature representations and has achieved certain performance improvements, its feature extraction process is too dependent on professional experimental tools, which increases the complexity of the experiment. The MVMR-BPF model only extracts the functional and structural features of amino acid sequences through the RDKit library and the ESM-2 protein language model, which not only simplifies the experimental process, but also shows higher performance and stability in the identification of therapeutic functions of bioactive peptides.

[0199] Strategies to enhance feature extraction capabilities: Compared with models such as SCN-MLTPP and Pre-Train, MVMR-BPF performs better in feature extraction. Although SCN-MLTPP obtains feature representation through professional tools, it does not use a deep learning model to further learn these features, but only uses a capsule network for feature fusion, which limits its performance. In contrast, although the Pre-Train model uses natural language processing technology to process peptide sequences and has achieved certain innovative results, it ignores the complex molecular level information of peptides, affecting its prediction accuracy. MVMR-BPF uses the ESM-2 model instead of RoBert for transfer learning, and uses different feature extraction modules in parallel in subsequent tasks, which significantly improves the prediction effect.

[0200] Specifically, we used the same five-fold cross-validation method as MLBP, MPMABP, SCN-MLTPP and Pre-Train, and conducted an exhaustive performance comparison. According to the data in the table, the MVMR-BPF model demonstrated excellent performance in predicting the properties of bioactive peptides, achieving the highest precision, coverage, accuracy, absolute correctness, and the lowest absolute error rate.

[0201] These results reveal that traditional single-label prediction methods—which rely mainly on machine learning and statistical methods to process a single data dimension—are generally ineffective when dealing with peptide molecules with complex properties. Compared with traditional methods, the use of multi-label deep learning methods can significantly improve the results. The MVMR-BPF model significantly improves the performance in the task of predicting bioactive peptides by introducing multi-view features, enhancing feature extraction capabilities, and using a large language model for preliminary feature extraction. MVMR-BPF can fully capture the characteristics of peptide molecules, providing a more effective and comprehensive solution, bringing breakthrough progress to the prediction of the therapeutic functions of bioactive peptides.

[0202]

[0203]

[0204] The above-mentioned embodiments only express the specific implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the protection scope of the present application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the technical solution concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. A method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning, characterized in that: Constructing the MVMR-BPF model, model feature extraction includes the following steps: Construct the graph structure of peptide molecules; Peptide amino acid sequence feature extraction: Use multi-scale dilated convolutional CNN and bidirectional LSTM to extract peptide amino acid sequence information; Peptide molecular graph feature extraction: Graph convolutional neural network and convolutional CNN methods are used to process molecular graph information from different perspectives of node features and topological information respectively; Residue feature extraction: Use the ESM-2 model to extract residue features from the sequence; Molecular fingerprint feature extraction: CNN is used with TLU and FRU layers to capture local features, and the Mamba model is used to capture long-distance dependencies to generate feature vectors; Fusion prediction: Combining the peptide amino acid sequence, molecular fingerprint information, and peptide molecular graph information, using multiple views to extract different modal features, and fusing information to complete multifunctional peptide prediction.

2. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning according to claim 1, characterized in that: The polypeptide amino acid sequence feature extraction specifically comprises the following steps: Send the peptide amino acid sequence vector represented by integer encoding to the sequence feature extraction module; The integer feature vector is converted into a dense continuous feature vector x with a dimension of 100 through the embedding layer, and the feature vector x passes through parallel dilated convolutional layers; for each dilated convolutional layer l, x (l) is the feature vector of the peptide amino acid sequence in the first layer, extracted through dilated convolution learning: in, is the convolution kernel weight of layer l, b (l) is the bias term, r l is the dilation rate of the layer, k is the size of the convolution kernel, max(0,z) represents the ReLU activation function, and for each layer l, the dilation rate r l They are 2, 4, and 8 respectively; For each dilated convolutional layer output x (l) , apply the maximum pooling operation to reduce the sequence dimension and retain the features: Among them, p (l) [i] is the result of the pooling operation, M is the size of the pooling window, and s is the step size; After each convolutional layer l, use Regularization is used to reduce overfitting, expressed as: L total =L+λ∑ l ||W (l) || 2 , Among them, ||W (l) || 2 is the sum of the squares of the convolution kernel weights of the lth layer, and λ is the regularization coefficient, which controls the strength of regularization; The vector after feature fusion is f, and the feature vector result after applying the Dropout layer is f′: f′=f⊙d, Where ⊙ represents element-wise multiplication, d is a random vector whose elements are independent and obey probability p of 1 for retention and probability 1-p of 0 for discarding; The feature f′ processed by the dilated convolution layer and Dropout is input to the bidirectional LSTM layer. The bidirectional LSTM can capture the forward and backward context dependencies of the sequence: h = BiLSTM(f′), Among them, h is the output of the bidirectional LSTM layer, combining the previous and next information of the sequence; Convert h to a one-dimensional vector h by flattening it flat , and compressed into a 128-dimensional feature vector X1 through a fully connected layer: X1=ReLU(W fc h flat +b fc ), Among them, W fc is the weight of the fully connected layer, b fc is the 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 multi-modal representation learning according to claim 1, characterized in that: The residue feature extraction specifically comprises the following steps: For the input polypeptide amino acid sequence X=(x1,x2,…,x n ), where x i Represents the i-th amino acid in the sequence. The calculation process of the self-attention layer is as follows: q i =W Q x i ,k i =W K x i ,v i =W V x i , Among them, W Q ,W K ,W V is the weight matrix, d k is the dimension of the key vector, used to scale the dot product result to avoid too large values; Train the ESM-2 model on large-scale protein sequence data; By performing average pooling on the features output by the specific layer of the ESM-2 model, an overall feature vector Y with a dimension of 1280 is obtained. Y is used as a high-dimensional feature representation of the amino acid sequence of each polypeptide molecule, which contains the structural and functional information of the polypeptide molecule. In the MVMR-BPF model, two fully connected layers are used to perform further feature extraction and dimensionality reduction, and a dropout layer is used to enhance the generalization ability of the model, ultimately obtaining a 64-dimensional compressed feature vector Y1; the structural and functional characteristics of the peptide amino acid sequence are further learned.

4. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning according to claim 1, characterized in that: The molecular fingerprint feature extraction specifically comprises the following steps: Use the RDKit library to simultaneously generate multiple molecular fingerprints from amino acid sequences; The two different feature fingerprints f1 and f2 are combined into a comprehensive feature vector f by point-by-point addition: f=f1+f2, The fused molecular fingerprint feature vector f is further extracted through a deep convolutional network, where each convolutional layer l is followed by an FRN layer and a TLU activation layer; the FRN layer is at the output f of the convolutional layer l. (l) To operate on: Among them, μ and σ 2 They are f (l) The mean and variance of , ∈ is a small constant to avoid division by zero, and τ is a learnable parameter; After the FRN layer, the TLU activation layer enhances y (l) The nonlinear characteristics of the vector are expressed as vector z (l) : z (l) =max(y (l) ,α), Among them, α is the activation threshold, which is a learnable parameter used to control the nonlinear threshold of the activation layer; The extracted features z (l) Enter the Mamba model for in-depth processing; the Mamba model includes multiple customized MambaBlocks. The core operation inside each MambaBlock can be described by the discretized version of the following continuous system: h′(t)=A·h(t)+B·x(t), y(t)=C·h(t), Among them, A, B and C are the state transfer matrix, input matrix and output matrix adjusted by discretization, respectively. The specific discretization steps are: A=exp(ΔA),B=(ΔA) -1 (exp(ΔA)-I)·ΔB, Where Δ is the time scale parameter; MambaBlock uses this method to capture long-range dependencies in amino acid sequences; Finally, through the feature extraction and context analysis layers, a 64-dimensional feature vector M1 is extracted.

5. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning according to claim 1, characterized in that: The molecular graph topology information processing in the polypeptide molecular graph feature extraction comprises the following steps: The peptide amino acid sequence was constructed into a molecular graph using the Rdkit library; network topology statistical indicators were calculated through the molecular graph; The network indicators were organized into a 100 × 7 feature matrix, where each row represents the network topological property of an amino acid, and the topological feature vector N was obtained; Set up a model consisting of a three-layer two-dimensional convolutional neural network to extract and utilize network topology information from the topological feature vector N; The topological feature vector N of the lth convolutional layer in the model (l) , perform convolution operation to get Y (l) : Y (l) =ReLU(W (l) *N (l) +b (l) ), Among them, W (l) and b (l) They are the convolution kernel and bias term of the lth layer, * indicates the convolution operation, and ReLU is the activation function; Get Y returned by the lth convolutional layer (l) Vector, and then through the batch normalization layer to get Z (l) , in, and is the mean and variance of the batch data, Y (l) and β (l) are learnable scaling and offset parameters, ∈ is a small constant to prevent division by zero; After batch normalization, global average pooling is used to reduce the feature dimension and obtain P (l) vector: Among them, N H and N W are the height and width of the feature map respectively; and in order to prevent the learned P (l) The vector is overfitted, and a Dropout layer is added to the module, with the dropout rate set to 0.3: N1 (l) =P (l) ⊙D (l) , D (l) is a random binary vector, ⊙ represents element-wise multiplication; Finally, the network topology information feature N1 is learned and extracted from the molecular graph topology information, and is used for further prediction of the properties of polypeptide molecules.

6. A method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning according to claim 1 or 5, characterized in that: The molecular graph node feature processing in the polypeptide molecular graph feature extraction comprises the following steps: The molecular graph is represented in the form of a graph, marked as G(V,E), where V is a set of nodes, representing atoms in the amino acid sequence, and E is a set of edges, representing chemical bonds between atoms. Each node contains a multidimensional feature vector that comprehensively reflects the properties of the atom. The upper limit of the number of nodes is set to 100. Each node is described by a feature vector containing different chemical property codes, and finally the node feature matrix H and adjacency matrix A of the molecular graph are obtained. Use graph convolutional neural network to extract molecular graph node features, set up a network architecture containing two layers of graph convolution, and perform graph convolution operations on the node feature matrix: Among them, H (l) is the node feature matrix of the lth layer, is the adjacency matrix A with self-loops added and the identity matrix I, yes The degree matrix, whose elements W (l) is the learnable weight matrix of the lth layer; to prevent overfitting, Dropout is applied after each layer of graph convolution: H (l+1) =Dropout(H (l+1) ,0.5), Then, the feature vector H output by the graph convolutional network is pooled through the global average pooling layer. (L) From the last layer L, we can reduce the dimension and get the global feature representation: Where N is the total number of nodes in the graph, It is the feature vector of node i in the last layer. G1 represents the structural and functional characteristics of the amino acid sequence and is suitable for further prediction of the properties of polypeptide molecules.

7. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation 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 five different input features in the feature dimension, including molecular sequence feature X1, molecular fingerprint feature M1, network topology information N1, graph representation feature G1, and molecular structure and function feature Y1 extracted from ESM-2, to obtain the feature fusion vector C; The Dropout layer was introduced and the dropout rate was set to 0.

3. The Dense layer with sigmoid activation function was used to predict the target biological activity, and L2 regularization was added.

8. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation 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 model training process uses binary cross entropy as the loss function, and the accuracy indicator is used to monitor the training progress:

9. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation learning according to claim 5, characterized in that: The network topology statistical indicators include the average neighbor degree, eccentricity, clustering coefficient, node degree, degree centrality, betweenness centrality and closeness centrality of nodes.

10. The method for predicting the function of bioactive peptides based on multi-view multi-modal representation 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 aromatic.

Images