An anticancer peptide classification method, system and storage medium based on a graph convolution network

Abstract

The first aspect of the present application provides an anticancer peptide classification method based on a graph convolution network, comprising: S1, obtaining an anticancer peptide training data set and a test data set; S2, constructing an anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encoding to obtain graph structure data; S3, constructing a graph convolution collapse pooling and residual network model, inputting the graph structure data into the model for training; and S4, applying the trained model to classify anticancer peptides. Wherein the graph convolution collapse pooling and residual network model comprises a stacked graph convolution network module, a graph collapse pooling module and a residual network module. The above scheme regards the anticancer peptide data as a kind of data similar to the graph structure, uses the graph convolution neural network to process the classification problem of the anticancer peptide, effectively distinguishes the anticancer peptide and the non-anticancer peptide, and avoids the problems of low recognition accuracy, insufficient generalization ability, lack of large-scale evaluation of features and prediction model of the existing model.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a method, system, and storage medium for classifying anticancer peptides based on graph convolutional networks. Background Technology

[0002] In the heterogeneous information network of drug synthesis, polypeptides with protein activity are multifunctional compounds derived from proteins and can be used as drugs to treat related diseases. A polypeptide is a biologically active protein composed of 10-50 amino acids. Its molecular structure is complex, consisting of molecular polymers composed of several or dozens of amino acids linked by peptide bonds. From the rotation, conformational changes, to conformational rearrangements of individual amino acid chains, protein conformation endows it with unique biological functions. In recent years, research on biologically active polypeptides has been increasingly applied in the field of drug design. Some polypeptides have been shown to possess broad biological properties, including anti-cancer peptides (ACPs), neuropeptides (NPs), antifungal peptides (AFPs), anti-angiogenic peptides (AAPs), and antibacterial peptides (ABPs). Due to their high specificity, selectivity, efficacy, and low toxicity, polypeptides are relatively safer than traditional drugs.

[0003] Anticancer peptides are polypeptides with anticancer activity that can effectively disrupt tumor cell membrane structures by perturbing gene expression or the immune system, thereby inhibiting cancer cell proliferation and migration. Traditional treatment methods for cancer have many shortcomings, such as drug resistance, low efficacy, and significant toxic side effects. Therefore, the application of anticancer peptides in the design of targeted anticancer drugs has received increasing attention.

[0004] In existing technologies, most methods for predicting anticancer peptides treat anticancer peptide data as ordered sequence data, using the primary sequence features of the anticancer peptides and combining them with other features as input to a classifier to build a prediction model. These methods ignore the contextual sequence features of anticancer peptides and the connections between amino acid nodes. Summary of the Invention

[0005] To address the above problems, the first aspect of this invention proposes a method for classifying anticancer peptides based on graph convolutional networks, comprising the following steps:

[0006] S1, Obtain the training and test datasets for the anticancer peptides;

[0007] S2, construct the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encode it to obtain graph structure data;

[0008] S3, construct a graph convolutional collapse pooling and residual network model, and input the graph structure data into the model for training;

[0009] as well as,

[0010] S4. Apply the trained model to classify anticancer peptides.

[0011] The aforementioned scheme proposes a graph convolutional neural network method based on collapse pooling and residual network to predict anticancer peptides (GCNCPR-ACPs). This scheme uses graph theory to address the structural features of anticancer peptides and the connections between amino acid nodes. Specifically, it treats anticancer peptide data as a graph-like structure, with amino acid nodes as vertices and the connections between amino acids as edges. The entire anticancer peptide sequence can be viewed as a spatial topological graph composed of amino acid sequences. From this perspective, graph theory is more suitable for anticancer peptide structural data. Therefore, this invention proposes a method using graph convolutional neural networks to process the graph data structure network of anticancer peptides.

[0012] Furthermore, the attribute data of this graph-structured network includes anticancer peptide chain features, graph labels, and adjacency matrix A. n×n The anticancer peptide chain characteristics include amino acid node type, amino acid node distance, and the physicochemical meaning of the amino acid nodes.

[0013] Furthermore, one-hot encoding is used to represent amino acid node types, Node2vec encoding is used to represent amino acid node distances, and encoding based on physicochemical characteristics is used to represent the physicochemical characteristics of the amino acid nodes themselves.

[0014] Furthermore, the graph convolutional collapse pooling and residual network model includes a stacked graph convolutional network module, a graph collapse pooling module, and a residual network module, where X is graph structure data, and the anticancer peptide chain feature F(X) = L(G(X)||D(X)||R(X)), where || denotes the concatenation operation on the feature vectors. Inputting the graph structure data X into the graph convolutional collapse pooling and residual network model (GCNCPR), after passing through the stacked graph convolutional network module, the overall structural feature G(X) of the anticancer peptide can be obtained; after passing through the graph collapse pooling module, the graph collapse structural feature D(X) of the anticancer peptide can be obtained; and after passing through the residual network module, the anticancer peptide feature R(X) can be obtained. The residual network module is used to prevent the gradient vanishing problem. Connecting the three learned anticancer peptide features together forms the final anticancer peptide chain feature F(X).

[0015] Furthermore, the formula for a single layer in the stacked graph convolutional network module is expressed as follows:

[0016]

[0017] Among them, H (l) ∈R N×M N is the number of nodes in the graph, and each node is represented by an M-dimensional eigenvector. H (l) H is the input feature of the l-th layer. (l+1) For the output features, A represents the adjacency matrix. This represents an adjacency matrix with self-loops. It is an adjacency matrix The degree is specifically represented as the sum of the in-degree and out-degree of all nodes in the graph that forms the adjacency matrix. W (l) Let be the trainable parameters of the l-th layer, and σ be the relevant activation function.

[0018] Furthermore, the calculation formula for the graph collapse operator S in the graph collapse pooling module is as follows:

[0019] S (l) =softmax(GCN) l，pool (A (l) H (l) ))

[0020] The formula for calculating the feature Z of a node is as follows:

[0021] Z (l) =GCN l，embed (A (l) H (l) )

[0022] Among them, GCN l,embed and GCN l,poolThese are two independent graph convolutional network layers, A (l) H is the adjacency matrix of amino acid nodes. (l) These are characteristics of amino acid nodes.

[0023] The formulas for calculating the node features and adjacency matrix of layer l+1 in the graph collapse pooling module are as follows:

[0024] Z (l+1) =S (l)T Z (l)

[0025] A (l+1) =S (l)T A (l) S (l)

[0026] in, Where, n (l) This represents the number of nodes in the l-th layer.

[0027] Furthermore, the loss function ζ is expressed as

[0028]

[0029] Where i represents the sample, N represents the number of samples, and p i y represents the probability of predicting a positive sample. i This represents the true label of the sample.

[0030] Secondly, this invention proposes an anticancer peptide classification system based on graph convolutional networks, comprising:

[0031] The data collection unit is configured to acquire training and test datasets for the anticancer peptides.

[0032] The graph structure data encoding unit is configured to construct and encode the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices, thereby obtaining graph structure data.

[0033] The model training unit is configured to construct graph convolutional collapse pooling and residual network models, and inputs the graph structure data into the model for training.

[0034] as well as,

[0035] A classification unit, configured to classify anticancer peptides using the trained model.

[0036] Thirdly, a computer-readable storage medium for classifying anticancer peptides based on graph convolutional networks, having stored thereon one or more computer programs that, when executed by a computer processor, implement any of the methods described above.

[0037] This invention proposes a method for classifying anticancer peptides based on graph convolutional collapse pooling and residual network models (GCNCPR-ACPs). Graph convolutional neural networks can effectively aggregate information from neighboring nodes and edges in graph data structures, automatically learning useful features from the data and representing network nodes and graph structure data with features. Existing anticancer peptide prediction models almost all use amino acid first-order sequence features, but most models do not use a comprehensive feature space and do not consider combining anticancer peptide structure and sequence data features. The GCNCPR-ACPs model is built based on the idea of ​​graph convolutional neural networks, considering the aggregation of information between nodes in the anticancer peptide sequence structure and extracting useful features from the anticancer peptide graph data structure, effectively distinguishing between anticancer peptides and non-anticancer peptides, and avoiding the problems of low recognition accuracy, insufficient generalization ability, and lack of large-scale evaluation of the features and prediction models used in existing models. Attached Figure Description

[0038] The accompanying drawings are provided to further understand this application. For ease of description, only the parts relevant to the invention are shown in the drawings.

[0039] Figure 1 This is a flowchart of an anticancer peptide classification method based on graph convolutional networks in one embodiment;

[0040] Figure 2 This is a schematic diagram of the framework of an anticancer peptide classification method based on graph convolutional networks in one embodiment;

[0041] Figure 3 This is a schematic diagram of the encoding process based on the Node2vec embedding method in one embodiment;

[0042] Figure 4 The diagram below is a schematic of the Convolutional Collapse Pooling and Residual Network (GCNCPR) framework in one embodiment.

[0043] Figure 5 This is an example diagram illustrating the graph matrix allocation sampling calculation process during graph collapse pooling in one embodiment;

[0044] Figure 6 The diagram below is a schematic diagram of the collapse calculation process in one embodiment;

[0045] Figure 7 This is a schematic diagram of the residual network module structure in one embodiment;

[0046] Figure 8 This is a schematic diagram of the structure of an anticancer peptide classification system based on a graph convolutional network in one embodiment;

[0047] Figure 9 This is a schematic diagram of a computer system structure for an electronic device used to implement the embodiments of this application in one embodiment. Detailed Implementation

[0048] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0049] Figure 1 This is a flowchart of an anticancer peptide classification method based on graph convolutional networks in one embodiment, which includes:

[0050] S1, Obtain the training and test datasets for the anticancer peptides;

[0051] S2, construct the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encode it to obtain graph structure data;

[0052] S3, construct a graph convolutional collapse pooling and residual network model, and input graph structure data into the model for training;

[0053] as well as,

[0054] S4. Use the trained model to classify anticancer peptides.

[0055] Figure 2 This is a schematic diagram of the framework of an anticancer peptide classification method based on graph convolutional networks in one embodiment, including:

[0056] The first step is to collect the amino acid sequences of anticancer peptides from the protein dataset to form the training and testing datasets for anticancer peptides.

[0057] The second step involves constructing and encoding the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices. The graph data attributes include amino acid node features, amino acid node labels, adjacency matrix, anticancer peptide chain labels, and anticancer peptide chain attributes.

[0058] The third step is to construct an anticancer peptide classification model based on graph convolutional collapse pooling and residual networks. This model mainly includes three modules: stacked graph convolution module, graph collapse pooling module, and residual network module.

[0059] The fourth step is to train and optimize the model to achieve the best classification results.

[0060] The following examples describe the construction and encoding methods for graph structure data.

[0061] An anticancer peptide chain is composed of several amino acids. All peptide chains can be constructed into a graph network, where amino acids are considered nodes (V) and connections between amino acids are considered edges (E). The anticancer peptide chains composed of amino acid sequences are then used to construct an anticancer peptide graph network (G). Based on the properties of graph networks, the anticancer peptide graph network (G) possesses graph attribute data, including amino acid node features, amino acid node labels, anticancer peptide chain features, graph labels, and an adjacency matrix A. n×n .

[0062] Adjacency matrix A n×n ∈[0,1] indicates whether there is an edge connecting the two amino acids, 1 indicates that there is an edge between the two amino acid nodes and they are connected to each other, 0 indicates that there is no connection between the two amino acid nodes, and n indicates the number of amino acid nodes.

[0063] Image labels are tags for anticancer peptide chains composed of several amino acid nodes. The value of an image label is 1 or 0, where 1 indicates that the anticancer peptide chain is an anticancer peptide (positive sample), and 0 indicates that it is a non-anticancer peptide (negative sample).

[0064] Anticancer peptide chain characteristics refer to the features formed by the connection of several amino acid nodes, which are used to describe the features of anticancer peptide chain diagrams.

[0065] Amino acid node labels are represented by 20 English letters (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y), which divide amino acid nodes into 20 categories and represent them with numbers 1-20.

[0066] In a preferred embodiment, the amino acid node features are represented by four encoding methods: One-hot encoding, Node2vec encoding, and encoding of the physicochemical properties of the two amino acids themselves.

[0067] One-hot encoding is an efficient binary encoding method where only one bit is valid at any given time, with the other bits set to 0. Since each amino acid sequence consists of 20 letters, one-hot encoding gives each amino acid node a 20-dimensional feature, allowing each amino acid node to be distinguished. For example, amino acid node A can be encoded as 100000000000000000000 (20 bits), and amino acid node C can be encoded as 010000000000000000000 (20 bits), thus representing each amino acid node as a one-dimensional 20-dimensional feature vector using one-hot encoding.

[0068] The Node2vec embedding method primarily uses random walks to sample vertex sequences. Common random walk sampling methods include depth-first search (DFS) and breadth-first search (BFS). Since one-hot encoding only considers the individual differences between two amino acid nodes and cannot represent the distance between nodes, the Node2vec embedding method is introduced to encode amino acid nodes. By encoding amino acid nodes using the Node2vec embedding method, neighbor information of the nodes is obtained, and the combination of vertex and adjacent sequence information is modeled to obtain the representation of the amino acid node. Figure 3 This is a schematic diagram of the encoding process using the Node2vec embedding method in one embodiment. In this diagram, amino acid T is the starting node, amino acid V is the current node of the random walk, and amino acid node V is searched using a DFS or BFS strategy to finally obtain a 2-dimensional feature vector of amino acid node.

[0069] A coding method based on the physicochemical properties of amino acids is introduced to represent the physicochemical meaning of amino acid nodes. This embodiment employs two coding methods based on the physicochemical properties of amino acids.

[0070] Table 1 divides 20 amino acids into 10 groups based on their physicochemical properties, with each group corresponding to a specific physicochemical attribute. In this embodiment, an encoding method based on the physicochemical properties of amino acids is used. According to Table 1, amino acid nodes are divided and encoded using 10-bit binary codes to form a 10-dimensional amino acid feature vector.

[0071] Table 1. Standard Amino Acid Node Division Method Based on Ten Physicochemical Properties

[0072]

[0073] Table 2 categorizes amino acids based on seven physicochemical properties: polarity, standard van der Waals volume, hydrophobicity, secondary structure, solvent accessibility, charge, and polarizability. Furthermore, each physicochemical property is grouped into three groups based on the CTD (composition, transition, and distribution) attributes of the amino acid node. In another encoding method based on the physicochemical characteristics of the amino acid itself, this embodiment uses embedded feature encoding of the amino acid nodes according to Table 2. This encoding method still uses binary encoding, with three groups for each physicochemical property, for a total of seven physicochemical properties. If the amino acid belongs to the corresponding group, the corresponding position is set to 1; otherwise, it is set to 0. Therefore, each amino acid node can be encoded into a 21 (7x3) dimensional feature vector.

[0074] Table 2. Amino acid node segmentation method based on seven physicochemical properties.

[0075]

[0076] In summary, in this embodiment, an amino acid node is encoded using four different encoding methods, resulting in a 51-dimensional binary code, as shown in Table 3.

[0077] Table 3. Encoding of four different amino acids

[0078]

[0079]

[0080] The following examples describe graphical convolutional collapse pooling and residual network models.

[0081] Figure 4 This is a schematic diagram of the Graph Convolutional Collapse Pooling and Residual Network (GCNCPR) framework in one embodiment. The model mainly includes three modules: a stacked graph convolutional network module, a graph collapse pooling module, and a residual network module. In this embodiment, the amino acid nodes on the anticancer peptide chain are encoded to obtain the graph structure data X of the anticancer peptide chain. X is input into the GCNCPR model. After passing through the stacked graph convolutional network module, the overall structural feature G(X) of the anticancer peptide is obtained; after passing through the graph collapse pooling module, the graph collapse structural feature D(X) of the anticancer peptide is obtained; and after passing through the residual network module, the anticancer peptide feature R(X) is obtained. Finally, the three learned anticancer peptide features are concatenated to form the final anticancer peptide chain feature F(X), as shown in the following formula:

[0082] F(X) = L(G(X)||D(X)IIR(X))

[0083] Here, || represents the concatenation operation on the feature vectors, L represents a fully linear concatenation, and the final output is the classification result of the anticancer peptides.

[0084] In one specific embodiment, multi-layer graph convolutional networks are stacked together to form a stacked graph convolutional network, which is used to learn the representation of all amino acid nodes and structural features of the anticancer peptide graph in the anticancer peptide graph network. A single-layer graph convolutional network is a variant of a traditional convolutional neural network, and its formula is as follows:

[0085]

[0086] Among them, H (l) ∈R N×M N is the number of nodes in the graph, and each node is represented by an M-dimensional feature vector. For the problem of predicting anticancer peptides, H... (l) The attributes are amino acid nodes, and M = 51; the input features of the l-th layer are H. (l) The output feature is H (l +1) A is an adjacency matrix. This represents an adjacency matrix with self-loops. Where I is the identity matrix; It is an adjacency matrix The degree, W (l) σ represents the trainable parameters of the l-th layer; σ is the relevant activation function, in this embodiment, σ(·) = ReLU(·). The above formula can be extended to multi-layer stacked graph convolutional networks. For example, using a three-layer graph convolutional network to extract anticancer peptide graph structural features, the complete form of the above formula becomes:

[0087]

[0088] In one specific embodiment, the anticancer peptide feature H containing amino acid nodes in the l-th layer is... (l) and adjacency matrix A (l) The graph data input to the graph collapse pooling module can collapse several amino acid nodes on a peptide chain into a large node. Through layers of graph collapse pooling, all amino acid nodes on the entire anticancer peptide chain are eventually collapsed into a super-large node. The characteristics of this super-large node are the characteristics of anticancer peptide graph collapse.

[0089] Graph collapse pooling is a method that combines graph collapse with graph convolutional neural networks for graph representation learning. Graph collapse pooling represents graph nodes hierarchically, not obtaining the final representation of the graph structure all at once, but rather compressing information layer by layer to obtain the final result. The graph collapse pooling process includes two important matrices: (1) the matrix allocator S∈R N×K Also known as the graph collapse operator, its value represents the probability that a node is assigned to a certain cluster, N represents the number of nodes, and all nodes are assigned to K clusters; (2) Sampling operator C can upsample or downsample the graph, and can also slice the adjacency matrix A in both directions.

[0090] Figure 5 This diagram illustrates the graph matrix allocation and sampling calculation process during graph collapse pooling. Assume nodes of the same color are assigned to the same cluster: white nodes setA = {v1, v2, v3}, and gray nodes setB = {v4, v5, v6, v7}. To obtain the characteristics of the two supernodes (white and gray), an adjacency matrix can be used. and matrix allocator Here, S divides the graph into two clusters. Let A... between =S T AS, A between This represents the connection strength between two supernodes after the collapse. The value A on the diagonal... between [i,i] represents the connection strength within the supernode of the cluster, A between [i,i] represents the connection strength between supernodes in the cluster. If A is set...between [i,i], then the adjacency matrix A between This transforms the adjacency matrix into one where only the diagonal nodes have values ​​and the rest are 0. To obtain the connection structure of the lower-level supernodes, the graph collapse operator S can be calculated using the activation function (softmax), as follows:

[0091] S (l) =softmax(GCN) l，pool (A (l) H (l) ))

[0092] The closer the values ​​of the elements in the graph collapse operator S are, the greater the probability that the nodes will be assigned to the same cluster. The node features Z can also be calculated using graph convolution.

[0093] Z (l) =GCN l，embed (A (l) H (l) )

[0094] Among them, GCN l,embed and GCN l,pool These are two independent graph convolutional network layers, and their inputs are the same: the feature H of the glutamic acid nodes. (l) and adjacency matrix A (l) However, their parameters are different. Therefore, their learning objectives are also different.

[0095] Figure 6 This is a schematic diagram of the graph collapse calculation process in one embodiment. Based on the calculation formulas for the graph collapse operator S and node features Z, the graph structure data composed of anticancer peptide chains is collapsed, yielding the supernode features and adjacency matrix of the (l+1)th layer:

[0096] Z (l+1) =S (l)T Z (l)

[0097] A (l+1) =S (l)T A (l) S (l)

[0098] in, Where, n (l) This represents the number of nodes in the l-th layer. The cluster assignment matrix S of the last layer is filled entirely with "1", which will collapse all the nodes into a single supernode, resulting in the collapsed global representation of the anticancer peptide graph.

[0099] In one specific embodiment, a ResNet residual network module is introduced. Increasing the width and depth of the network can significantly improve its performance, but simply increasing the network depth indefinitely may lead to saturation or even a decrease in model accuracy, resulting in degradation. As the number of network layers increases, the derivative variables may become too large or too small during backpropagation optimization of the model parameters, leading to gradient vanishing or gradient explosion. To address this issue, the ResNet residual network is introduced.

[0100] A residual network consists of several residual modules. Each residual module has two possible structures. Figure 7 This is a schematic diagram of the residual network module structure. Deep model nonlinear mapping structures typically have multiple nonlinear mapping layers F(x). Adding more layers to a deep model can lead to higher training errors. Therefore, by adding an identity mapping x or h(x), information from higher layers is directly fed to lower layers. Features of higher-level mappings are preserved through simple connections such as horizontal or vertical concatenation for feature function computation. When the feature dimension x of the input residual block matches the output feature dimension of the nonlinear mapping structure F(x), it can be directly used... Figure 7 (A) shows the residual module for feature addition. When the feature dimension x of the input residual block is inconsistent with the output feature dimension of the nonlinear mapping structure F(x), then it is necessary to use... Figure 7 The residual module shown in (B) first performs a feature identity mapping h(x), then linearly maps the input feature x, and finally performs an additive function feature operation with the output feature of the nonlinear mapping structure F(x). The specific calculation process is as follows:

[0101] y = F(x) + x

[0102] y = F(x) + h(x)

[0103] Where x is the input feature, y is the output function, F(x) is the deep nonlinear mapping learning module, and h(x) is the identity mapping learning module.

[0104] In a preferred embodiment, the loss function ζ for graph convolutional collapse pooling and the residual network model is expressed as:

[0105]

[0106] Where i represents the sample, N represents the number of samples, and p i y represents the probability of predicting a positive sample. iThis represents the true label of the sample. In classification problems, the cross-entropy loss function is often used in conjunction with the activation function (softmax). The activation function (softmax) processes the output, ensuring that the sum of the predicted values ​​for binary or multi-class classification equals 1, and utilizes the cross-entropy loss function to optimize the model's training process. Since the anti-cancer peptide prediction model is a binary classification problem, the above formula is used to represent the loss function. The smaller the loss function ζ, the better the model's prediction performance.

[0107] The following examples describe the experimental results of the proposed graph convolutional network-based anticancer peptide classification method. This experiment used Python 3.8 for modeling, and the hardware environment was as follows: dual-core Intel Core i5 processor, 2.3GHz; 8GB RAM 2133MHz LPDDR3; Intel Iris Plus Graphics 640 1536MB; and macOS Catalina version 10.15.7.

[0108] The anticancer peptide dataset used in this embodiment is from the literature. Positive samples are anticancer peptides that have been physically verified, while negative samples are selected from 2878 negative samples from the public anticancer peptide database CancerPPD. To avoid homology bias, all sample sequences were filtered using the CD-HIT program to remove redundant sequences with similarity higher than 80%, ensuring that the sequences in the training and test sets are distinct. Ultimately, 332 anticancer peptides (positive samples) and 2878 non-anticancer peptides (negative samples) were obtained to evaluate the performance of the prediction model. To balance the training model parameters, the training set ACPs500 consists of 250 anticancer peptides and 250 non-anticancer peptides. The test set ACPs164 consists of 82 anticancer peptide sequences and 82 non-anticancer peptide sequences. This paper also uses two other datasets: neuropeptides (NPs) and antifungal peptides (AFPs). Detailed information about the datasets is shown in the table below.

[0109] Table 4 Anticancer peptide dataset

[0110]

[0111]

[0112] This embodiment uses the following evaluation metrics to evaluate the model: Sensitivity (SE, also known as Recall), Specificity (SP), Precision, Accuracy (ACC), Matthews Correlation Coefficient (MCC), F1 (the concordance between precision and recall), TPR (the proportion of correctly predicted peptides among all true anticancer peptides, i.e., recall), FPR (the proportion of correctly predicted peptides among all non-anticancer peptides), Correct Index, and AUC (area under the ROC curve). The specific calculation method is shown in the following formula.

[0113]

[0114] Among them, TP (true positive) represents the number of true anticancer peptides that were correctly predicted, FP (false positive) represents the number of true anticancer peptides that were incorrectly predicted, TN (true negative) represents the number of non-anticancer peptides that were correctly predicted, and FN (false negative) represents the number of non-anticancer peptides that were incorrectly predicted.

[0115] For any specific classifier, it's impossible to simultaneously improve all the metrics mentioned above. Similarly, for anticancer peptide prediction, 100% accuracy is unattainable; a certain degree of false positives must be tolerated. Suppose that in 1000 predictions, 5 proteins are predicted as anticancer peptides, with 1 protein actually possessing an anticancer peptide, while the other 4 are false positives. In this case, the accuracy drops from 999 / 1000 = 99.9% to 996 / 1000 = 99.6%, but the recall increases from 0 / 1 = 0% to 1 / 1 = 100%. Although a few anticancer peptides were falsely reported, the truly effective anticancer peptides were not missed and can be further validated in subsequent biological experiments. Therefore, it's generally required that, given a certain level of accuracy, the classifier's recall be as high as possible.

[0116] This embodiment selects the following 7 models for comparison: iACP, ACPred-FL, PEPred-Suite, ACPred-Fuse, AntiCP_ACC, AntiCP_DC, and Hajishirifi.

[0117] The cross-validation with tenfold discount is shown in the table below:

[0118] Table 5 Comparison of 10-fold cross-validation results for different models

[0119]

[0120]

[0121] As can be seen from the table above, the GCNCPR-ACPs model proposed in this application significantly outperforms the other seven prediction models in five metrics (SE, SP, ACC, MCC, and AUC), indicating that it can identify anticancer peptides more accurately than other existing prediction models.

[0122] In one specific embodiment, the GCNCPR-ACPs model is adjusted. The learning rate is set to [0.1, 0.01, 0.001]. It can be seen that the model performance is optimal when the learning rate is 0.01.

[0123] Figure 8 This is a schematic diagram of the structure of an anticancer peptide classification system 800 based on a graph convolutional network in one embodiment, including:

[0124] Data collection unit 801 is configured to acquire training and test datasets for anticancer peptides;

[0125] Graph structure data encoding unit 802 is configured to construct an anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encode it to obtain graph structure data;

[0126] The model training unit 803 is configured to construct a graph convolutional collapse pooling and residual network model, and inputs the graph structure data into the model for training.

[0127] as well as,

[0128] Classification unit 804 is configured to classify anticancer peptides using the trained model.

[0129] Figure 9 A schematic diagram of the structure of a computer system 900 suitable for implementing an electronic device according to embodiments of the present application is shown. Figure 9 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0130] like Figure 9 As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 902 or programs loaded from storage section 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the system 900. The CPU 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0131] The following components are connected to I / O interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a liquid crystal display (LCD) and speakers, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card and a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to I / O interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 910 as needed so that computer programs read from it can be installed into storage section 908 as needed.

[0132] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 909, and / or installed from removable medium 811. When the computer program is executed by central processing unit (CPU) 901, it performs the functions defined in the methods of this application.

[0133] This application proposes an anticancer peptide classification method (GCNCPR-ACPs) based on graph convolutional collapse pooling and residual network models. Ten-fold cross-validation results show that, compared with the baseline model, the GCNCPR-ACPs model significantly outperforms existing anticancer peptide classification models on five metrics (SE, SP, ACC, MCC, and AUC), representing a significant improvement.

[0134] Although the contents of this application have been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that any changes in form and detail made to this application without departing from the spirit and scope of this application as defined by the appended claims and without inventive effort are within the scope of protection of this application.

Claims

1. A method for classifying anticancer peptides based on graph convolutional networks, characterized in that, Includes the following steps: S1, Obtain the training and test datasets for the anticancer peptides; S2, construct the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encode it to obtain graph structure data; S3, construct a graph convolutional collapse pooling and residual network model, and input the graph structure data into the model for training; as well as, S4, Apply the trained model to classify anticancer peptides; The attribute data of the graph structure network described in S2 includes anticancer peptide chain features, graph labels, and adjacency matrix. The anticancer peptide chain characteristics include amino acid node type, amino acid node distance, and the physicochemical meaning of the amino acid nodes; The encoding method described in S2 includes using one-hot encoding to represent amino acid node type, using Node2vec encoding to represent amino acid node distance, and using encoding based on physicochemical characteristics to represent the physicochemical characteristics of the amino acid node itself. The graph convolutional collapse pooling and residual network model described in S3 includes a stacked graph convolutional network module G(X), a graph collapse pooling module D(X), and a residual network module R(X), where X represents graph structure data, and the anticancer peptide chain features... , where || denotes the concatenation operation on the feature vectors; The formula for a single layer in the stacked graph convolutional network module is as follows: ； in, N is the number of nodes in the graph, and each node is represented by an M-dimensional feature vector. For the first The input features of the layer For output features, Represents the adjacency matrix. This represents an adjacency matrix with self-loops. , It is an adjacency matrix The degree, , For the first l The trainable parameters of the layer, For the relevant activation function; The calculation formula for the graph collapse operator S in the graph collapse pooling module is as follows: ； The formula for calculating the feature Z of a node is as follows: ； in, and These are two independent graph convolutional network layers. Let be the adjacency matrix of the amino acid nodes. These are characteristics of amino acid nodes; The graph collapse pooling module is the first The formulas for calculating the node characteristics and adjacency matrix of a layer are as follows: ； in, ,in, Indicates the first The number of nodes in a layer.

2. The method for classifying anticancer peptides based on graph convolutional networks according to claim 1, characterized in that, loss function Represented as ； in, Let N represent the sample, and N represent the sample size. This represents the probability that the sample is predicted to be positive. This represents the true label of the sample.

3. A cancer peptide classification system based on graph convolutional networks, characterized in that, include: The data collection unit acquires the training and test datasets for the anticancer peptides. The graph structure data encoding unit constructs the anticancer peptide sequence into a graph structure network with amino acid nodes as vertices and encodes it to obtain graph structure data; The model training unit constructs a graph convolutional collapse pooling and residual network model, and inputs the graph structure data into the model for training. as well as, The classification unit applies the trained model to classify anticancer peptides. The attribute data of the graph structure network includes anticancer peptide chain features, graph labels, and adjacency matrix. The anticancer peptide chain characteristics include amino acid node type, amino acid node distance, and the physicochemical meaning of the amino acid nodes; The graph structure network is encoded in the following ways: one-hot encoding is used to represent the amino acid node type, Node2vec encoding is used to represent the distance between amino acid nodes, and encoding based on physicochemical characteristics is used to represent the physicochemical characteristics of the amino acid nodes themselves. The graph convolutional collapse pooling and residual network model includes a stacked graph convolutional network module G(X), a graph collapse pooling module D(X), and a residual network module R(X), where X represents graph structure data, and the anticancer peptide chain features... , where || denotes the concatenation operation on the feature vectors; The formula for a single layer in the stacked graph convolutional network module is as follows: ； in, N is the number of nodes in the graph, and each node is represented by an M-dimensional feature vector. For the first The input features of the layer For output features, Represents the adjacency matrix. This represents an adjacency matrix with self-loops. , It is an adjacency matrix The degree, , For the first l The trainable parameters of the layer, For the relevant activation function; The calculation formula for the graph collapse operator S in the graph collapse pooling module is as follows: ； The formula for calculating the feature Z of a node is as follows: ； in, and These are two independent graph convolutional network layers. Let be the adjacency matrix of the amino acid nodes. These are characteristics of amino acid nodes; The graph collapse pooling module is the first The formulas for calculating the node characteristics and adjacency matrix of a layer are as follows: ； in, ,in, Indicates the first The number of nodes in a layer.

4. A computer-readable storage medium for classifying anticancer peptides based on graph convolutional networks, wherein one or more computer programs are stored thereon, characterized in that, When the one or more computer programs are executed by a computer processor, they perform the method according to any one of claims 1 to 2.

Images